{"schemaVersion":1,"datasetVersion":"271051c653d9a602afe4c96c4da83afb5a783095a416a984a5b76109be990aaf","retrievedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","documentation":"https://epoch.ai/data/ai-models-documentation","license":"CC BY 4.0","attribution":"Epoch AI, Data on AI models","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","rawRecords":3628,"excludedUndatedOrFuture":12,"limitations":"Curated, non-exhaustive model history across AI domains. Publication dates are not API release dates. Parameters and training compute are source estimates with differing architectures and confidence; neither is a performance score. No historical API prices are inferred. Publication-year counts reflect dataset coverage, not all model releases.","models":[{"id":"epoch:52e63de0e3149ec13d58","name":"DeepSeek V4.1 Flash 1017","organization":"DeepSeek","date":"2026-10-17","domains":["Language","Multimodal","Vision"],"country":"China","reference":"https://api-docs.deepseek.com/news/news261017","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:54f6d07c706a2682c1a9","name":"Muse Spark Open","organization":"Meta AI","date":"2026-10-15","domains":["Language","Multimodal","Vision","Audio","Video"],"country":"United States of America","reference":"https://research.meta.ai/blog/muse-spark-open\nhttps://dev.meta.ai/docs/muse-spark-open","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:e646dfa9c6d90f58cc48","name":"Gemini 4 Argon","organization":"Google DeepMind","date":"2026-09-30","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/","confidence":"Confident","accessibility":"Limited access"},{"id":"epoch:4ad54a8238e5ebdd7fdc","name":"GPT-6.1 Sol","organization":"OpenAI","date":"2026-09-29","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://openai.com/index/introducing-gpt-6-1-sol/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:ebc7dfc0b020889622c5","name":"Claude Sonnet 5.5","organization":"Anthropic","date":"2026-09-28","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://www.anthropic.com/claude-sonnet-5-5","confidence":"Confident","accessibility":"API access"},{"id":"epoch:219627283116c2674f5f","name":"Claude Opus 5.5","organization":"Anthropic","date":"2026-09-22","domains":["Language"],"country":"United States of America","reference":"https://www.anthropic.com/claude-opus-5-5","confidence":"Confident","accessibility":"API access"},{"id":"epoch:5094e3b4274c00e73dfc","name":"GPT-6 Sol","organization":"OpenAI","date":"2026-09-22","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://openai.com/index/introducing-gpt-6-sol-and-luna/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:cd84d55f30eb7adb8c4b","name":"GPT-6 Luna","organization":"OpenAI","date":"2026-09-22","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://openai.com/index/introducing-gpt-6-sol-and-luna/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:412993d893d566bd5555","name":"Grok 4.7","organization":"xAI","date":"2026-09-21","domains":["Language"],"country":"United States of America","reference":"https://x.ai/news/grok-4-7","confidence":"Likely","accessibility":"API access"},{"id":"epoch:62eb096966556ab919cf","name":"Step 5 Preview","organization":"StepFun","date":"2026-09-18","domains":["Language","Vision","Multimodal"],"country":"China","reference":"https://platform.stepfun.ai/docs/en/guides/models/step-5-preview","confidence":"Likely","accessibility":"API access"},{"id":"epoch:9f8a4ce481b1194dc259","name":"Periodic Neon","organization":"Periodic Labs","date":"2026-09-15","domains":["Language","Materials science"],"country":"United States of America","reference":"https://periodic.com/news/nature-is-our-learning-environment","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:1f135d59b7429310efa1","name":"SWE-2","organization":"Cognition","date":"2026-09-10","domains":["Language"],"country":"United States of America","reference":"https://cognition.com/blog/swe-2","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:6801998507d114a75aab","name":"DeepSeek V4.1 Flash","organization":"DeepSeek","date":"2026-09-09","domains":["Language","Vision","Multimodal"],"country":"China","reference":"https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/DeepSeek_V41_Tech_Report.pdf \n\nhttps://www.deepseek.com/en/news/deepseek-v4-1-flash/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d23a075b521d8411ed65","name":"ChatGPT Images 2.5","organization":"OpenAI","date":"2026-09-08","domains":["Image generation","Vision"],"country":"United States of America","reference":"https://openai.com/index/introducing-chatgpt-images-2-5/","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:09e2758b5ffead7645c9","name":"GPT-Image-2.5 Flare","organization":"OpenAI","date":"2026-09-08","domains":["Image generation"],"country":"United States of America","reference":"https://openai.com/index/introducing-chatgpt-images-2-5/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:50197a39459fbc20e299","name":"GPT-Image-2.5 Sunburst","organization":"OpenAI","date":"2026-09-08","domains":["Image generation"],"country":"United States of America","reference":"https://openai.com/index/introducing-chatgpt-images-2-5/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:fd50ac4f7952beb4cca1","name":"GPT-6 Astra","organization":"OpenAI","date":"2026-09-03","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"","confidence":"Likely","accessibility":"API access"},{"id":"epoch:a3313fd215f4bdda2418","name":"OpenEvidence Darwin","organization":"OpenEvidence","date":"2026-09-03","domains":["Language","Medicine"],"country":"","reference":"https://www.openevidence.com/blog/model-family","confidence":"Confident","accessibility":"API access"},{"id":"epoch:9d652746bd53ce329e91","name":"OpenEvidence Osler","organization":"OpenEvidence","date":"2026-09-03","domains":["Language","Medicine"],"country":"","reference":"https://www.openevidence.com/blog/model-family","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:810b368e0c23fcc743b0","name":"OpenEvidence Sackett","organization":"OpenEvidence","date":"2026-09-03","domains":["Language","Medicine"],"country":"","reference":"https://www.openevidence.com/blog/model-family","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:f08e5eb621ac3766caf7","name":"OpenEvidence Snow","organization":"OpenEvidence","date":"2026-09-03","domains":["Language","Medicine"],"country":"","reference":"https://www.openevidence.com/blog/model-family","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:d71c85e4f4277bfe69b7","name":"WeatherNext 3","organization":"Google DeepMind,Google Research","date":"2026-09-03","domains":["Earth science"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2609.03582 \n\nhttps://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3/ ","confidence":"Speculative","accessibility":"Hosted access (no API)"},{"id":"epoch:e1332a1d30dd7b676321","name":"Gemini 3.8 Flash","organization":"Google DeepMind","date":"2026-09-02","domains":["Language","Multimodal","Image generation","Video","Audio"],"country":"United States of America","reference":"https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash","confidence":"Confident","accessibility":"API access"},{"id":"epoch:fa69a19df00288f9f963","name":"Muse Spark 1.3","organization":"Meta AI","date":"2026-09-02","domains":["Language"],"country":"United States of America","reference":"https://research.meta.ai/blog/introducing-muse-spark-1-3","confidence":"Confident","accessibility":"API access"},{"id":"epoch:e5bd048f43a27fb67342","name":"Claude Fable 5.1","organization":"Anthropic","date":"2026-09-01","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://www.anthropic.com/claude-fable-and-mythos-5-1","confidence":"Confident","accessibility":"API access"},{"id":"epoch:ee24010e3280faef3362","name":"Tencent Hy4 preview","organization":"Tencent","date":"2026-08-28","domains":["Language"],"country":"China","reference":"https://hy.tencent.com/hy4-preview","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:250ad376b18772157d87","name":"GLM-5.3-Flash","organization":"Z.ai (Zhipu AI)","date":"2026-08-20","domains":["Language","Vision","Multimodal"],"country":"China","reference":"https://z.ai/blog/glm-5.3-flash","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:29658eae7c50e0876ba0","name":"GEN-1.5","organization":"Generalist","date":"2026-08-19","domains":["Robotics","Language","Multimodal"],"country":"United States of America","reference":"https://generalistai.com/blog/gen-1.5","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4b060905d75d9b191b0c","name":"GLM-5.3","organization":"Z.ai (Zhipu AI)","date":"2026-08-14","domains":["Language"],"country":"China","reference":"https://z.ai/blog/glm-5.3","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:d3f7eb8492cda2e0a810","name":"Qwen 3.8 27B","organization":"Alibaba","date":"2026-08-14","domains":["Language","Multimodal","Vision","Video"],"country":"China","reference":"https://x.com/Alibaba_Qwen/status/2088280182356611304\nhttps://huggingface.co/Qwen/Qwen3.8-27B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:592c7fb4d02bc23fb19d","name":"Gemini 3.7 Flash","organization":"Google DeepMind","date":"2026-08-13","domains":["Language","Audio","Video","Multimodal"],"country":"United States of America","reference":"https://x.com/GoogleAIStudio/status/2087953811931320341\n\nhttps://ai.google.dev/gemini-api/docs/models/gemini-3.7-flash ","confidence":"Confident","accessibility":"API access"},{"id":"epoch:4c6d008b71aa4c1863e6","name":"Grok 4.6","organization":"xAI","date":"2026-08-12","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://x.ai/news/grok-4-6","confidence":"Confident","accessibility":"API access"},{"id":"epoch:0c0d72dd0aee1fee9b21","name":"GPT-5.6 Cyber","organization":"OpenAI","date":"2026-08-11","domains":["Cybersecurity","Language"],"country":"United States of America","reference":"https://openai.com/index/expanding-daybreak-as-the-cyber-defense-window-narrows/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:a917df48e391b8e99a88","name":"Nemotron 3.5 Lightning","organization":"Nvidia","date":"2026-08-11","domains":["Language"],"country":"United States of America","reference":"https://developer.nvidia.com/blog/nvidia-nemotron-3-5-lightning-delivers-fast-accurate-specialized-task-execution-for-long-running-agents/","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5bead17b8e1031c76386","name":"Muse Glimmer","organization":"Meta AI","date":"2026-08-10","domains":["Language","Multimodal"],"country":"United States of America","reference":"https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ad7146f299aa000780c0","name":"Qwen3.8-2.4T-A95B","organization":"Alibaba","date":"2026-08-08","domains":["Language"],"country":"China","reference":"https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:9a12d7383424f34c0c36","name":"Motif-3","organization":"Motif Technologies","date":"2026-08-07","domains":["Language"],"country":"South Korea","reference":"https://huggingface.co/Motif-Technologies/Motif-3\nhttps://huggingface.co/Motif-Technologies/Motif-3/blob/main/motif_3_technical_report.pdf\nhttps://github.com/MotifTechnologies/motif3-training-example","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c6017b54bdbf60906f48","name":"Muse Spark 1.2","organization":"Meta AI","date":"2026-08-05","domains":["Language"],"country":"United States of America","reference":"https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2","confidence":"Confident","accessibility":"API access"},{"id":"epoch:296b0466aba910dd5c79","name":"DeepSeek V4 Flash 0731","organization":"DeepSeek","date":"2026-07-31","domains":["Language"],"country":"China","reference":"https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:dc8643a1b4bcc32a8150","name":"K-EXAONE 2.0","organization":"LG AI Research","date":"2026-07-31","domains":["Language"],"country":"South Korea","reference":"https://huggingface.co/LGAI-EXAONE/K-EXAONE-2.0-750B-A37B\nhttps://github.com/LG-AI-EXAONE/K-EXAONE-2.0\nhttps://huggingface.co/LGAI-EXAONE/K-EXAONE-2.0-750B-A37B/blob/main/assets/K-EXAONE-2.0-Technical-Report.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2d546354af8efa621e76","name":"Inkling-Small","organization":"Thinking Machines","date":"2026-07-30","domains":["Language","Multimodal","Audio","Video","Vision"],"country":"United States of America","reference":"https://thinkingmachines.ai/news/introducing-inkling/\n\nhttps://thinkingmachines.ai/model-card/inkling-small/ ","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1f6a3182cc81870d94b5","name":"Gemini Robotics ER 2","organization":"Google DeepMind","date":"2026-07-30","domains":["Robotics"],"country":"United States of America","reference":"https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-robotics-er-2/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:142a168f226ad91fd302","name":"GPT Transcribe","organization":"OpenAI","date":"2026-07-29","domains":["Speech"],"country":"United States of America","reference":"https://developers.openai.com/api/docs/models/gpt-transcribe","confidence":"Confident","accessibility":"API access"},{"id":"epoch:0aa76a4346b47b190d65","name":"GPT Live Transcribe","organization":"OpenAI","date":"2026-07-29","domains":["Speech"],"country":"United States of America","reference":"https://developers.openai.com/api/docs/models/gpt-live-transcribe","confidence":"Confident","accessibility":"API access"},{"id":"epoch:b13077cb88dbdc200700","name":"A.X K2","organization":"SK Telecom","date":"2026-07-29","domains":["Language"],"country":"South Korea","reference":"https://huggingface.co/skt/A.X-K2\nhttps://github.com/SKT-AI/A.X-K2/blob/main/A_X_K2_Tech_Report.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:67dfe29930d4dfd8a302","name":"Qwen3.7 Flash","organization":"Alibaba","date":"2026-07-27","domains":["Language"],"country":"China","reference":"https://www.qwencloud.com/models/qwen3.7-flash","confidence":"Confident","accessibility":"API access"},{"id":"epoch:2aba10437975a48d75e7","name":"Claude Opus 5","organization":"Anthropic","date":"2026-07-24","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://www.anthropic.com/news/claude-opus-5","confidence":"Confident","accessibility":"API access"},{"id":"epoch:947c78321bbe97b38114","name":"Gemini 3.6 Flash","organization":"Google DeepMind","date":"2026-07-21","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:506165ea1adcbb213b57","name":"Gemini 3.5 Flash-Lite","organization":"Google DeepMind","date":"2026-07-21","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:3232f6fce64b337b8ba0","name":"Gemini 3.5 Flash Cyber","organization":"Google DeepMind","date":"2026-07-21","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:23c06baf333574e99133","name":"Qwen 3.8 Max","organization":"Alibaba","date":"2026-07-19","domains":["Language","Vision","Video"],"country":"China","reference":"https://qwen.ai/blog?id=qwen3.8","confidence":"Confident","accessibility":"API access"},{"id":"epoch:e344e2c60ab823cd5cc8","name":"Kimi K3","organization":"Moonshot","date":"2026-07-16","domains":["Language","Multimodal","Vision"],"country":"China","reference":"https://huggingface.co/moonshotai/Kimi-K3","confidence":"Speculative","accessibility":"Open weights (non-commercial)"},{"id":"epoch:cf223bf5394b4dfe83d0","name":"Inkling","organization":"Thinking Machines","date":"2026-07-15","domains":["Language","Multimodal","Audio","Video","Vision"],"country":"United States of America","reference":"https://thinkingmachines.ai/news/introducing-inkling/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c86d5f4d18466ef38cd4","name":"GPT-5.6 Sol","organization":"OpenAI","date":"2026-07-09","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://openai.com/index/gpt-5-6/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:6e6c8be351c12498984a","name":"GPT-5.6 Terra","organization":"OpenAI","date":"2026-07-09","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://openai.com/index/gpt-5-6/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:b79585eddf431cd89591","name":"GPT-5.6 Luna","organization":"OpenAI","date":"2026-07-09","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://openai.com/index/gpt-5-6/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:988980601eb0aaf4c8a3","name":"Muse Spark 1.1","organization":"Meta AI","date":"2026-07-09","domains":["Language"],"country":"United States of America","reference":"https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/\nhttps://ai.meta.com/static-resource/muse-spark-1-1-evaluation-report","confidence":"Confident","accessibility":"API access"},{"id":"epoch:bedf505291d712b2edb4","name":"Grok 4.5","organization":"xAI","date":"2026-07-08","domains":["Language"],"country":"United States of America","reference":"https://x.ai/news/grok-4-5","confidence":"Likely","accessibility":"API access"},{"id":"epoch:7b68f0d1d535b98c8f62","name":"SWE-1.7","organization":"Cognition","date":"2026-07-08","domains":["Language"],"country":"United States of America","reference":"https://cognition.com/blog/swe-1-7","confidence":"Confident","accessibility":""},{"id":"epoch:6e78e3b84c49e03daf87","name":"Muse Image","organization":"Meta AI","date":"2026-07-07","domains":["Image generation"],"country":"United States of America","reference":"https://about.fb.com/news/2026/07/introducing-muse-image-meta-ai/","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:f25e5a7daeb15d3b9225","name":"GPT-Realtime-2.1-Mini","organization":"OpenAI","date":"2026-07-06","domains":[],"country":"United States of America","reference":"","confidence":"Confident","accessibility":"API access"},{"id":"epoch:974f4354fa22ffad802e","name":"Tencent Hy3","organization":"Tencent","date":"2026-07-06","domains":["Language"],"country":"China","reference":"https://huggingface.co/tencent/Hy3","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0bff8c9f9e68612f11d3","name":"Claude Sonnet 5","organization":"Anthropic","date":"2026-06-30","domains":["Language","Multimodal"],"country":"United States of America","reference":"https://www.anthropic.com/news/claude-sonnet-5","confidence":"Confident","accessibility":"API access"},{"id":"epoch:d151aead00e5b4b05a4c","name":"LongCat-2.0","organization":"Meituan Inc","date":"2026-06-29","domains":["Language"],"country":"China","reference":"https://longcat.chat/blog/longcat-2.0/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fbaa821000be951f4440","name":"Solar Open2 250B","organization":"Upstage","date":"2026-06-28","domains":["Language"],"country":"South Korea","reference":"https://huggingface.co/upstage/Solar-Open2-250B","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:44f38ef22b569e94288b","name":"Doubao Seed 2.1 Turbo","organization":"ByteDance","date":"2026-06-23","domains":["Language","Multimodal","Vision","Video"],"country":"China","reference":"https://seed.bytedance.com/en/blog/seed2-1-officially-released-advancing-ai-productivity\n\nhttps://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf ","confidence":"Confident","accessibility":"API access"},{"id":"epoch:b6d8e5d619d9004aca1e","name":"Doubao Seed 2.1 Pro","organization":"ByteDance","date":"2026-06-23","domains":["Language","Multimodal","Vision","Video"],"country":"China","reference":"https://seed.bytedance.com/en/blog/seed2-1-officially-released-advancing-ai-productivity\n\nhttps://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/seed2.1/Seed2_1_Model_Card.pdf ","confidence":"Confident","accessibility":"API access"},{"id":"epoch:00035b648ca396941276","name":"GLM-5.2","organization":"Z.ai (Zhipu AI)","date":"2026-06-16","domains":["Language"],"country":"China","reference":"https://z.ai/blog/glm-5.2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:31068fc70d9d3f02c986","name":"Kimi K2.7 Code","organization":"Moonshot","date":"2026-06-12","domains":["Language","Multimodal","Vision"],"country":"China","reference":"https://platform.kimi.ai/docs/guide/kimi-k2-7-code-quickstart","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c2f0c099ee697b271f4c","name":"DiffusionGemma 26B-A4B","organization":"DeepMind","date":"2026-06-10","domains":["Language","Vision","Multimodal"],"country":"United Kingdom","reference":"https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/ \n\nhttps://huggingface.co/google/diffusiongemma-26B-A4B-it ","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d943f79e4a7161f8eda7","name":"Claude Mythos 5","organization":"Anthropic","date":"2026-06-09","domains":["Language","Multimodal"],"country":"United States of America","reference":"https://www.anthropic.com/news/claude-fable-5-mythos-5","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:9ecc2f028b26d03540da","name":"Claude Fable 5","organization":"Anthropic","date":"2026-06-09","domains":["Language","Multimodal"],"country":"United States of America","reference":"https://www.anthropic.com/news/claude-fable-5-mythos-5","confidence":"Confident","accessibility":"API access"},{"id":"epoch:a2dc63a706d89d0dbf5e","name":"North Mini Code","organization":"Cohere","date":"2026-06-09","domains":["Language"],"country":"Canada","reference":"https://cohere.com/blog/north-mini-code","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:65b6f4dca987e7c3880c","name":"Nemotron 3 Ultra","organization":"Nvidia","date":"2026-06-04","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2512.20856\nhttps://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Ultra-Technical-Report.pdf ","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ecc1e95d828e6c43f840","name":"Gemma 4 12B","organization":"Google DeepMind","date":"2026-06-03","domains":["Language","Multimodal","Audio"],"country":"United States of America","reference":"https://ai.google.dev/gemma/docs/core/model_card_4","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6da4667f6dfd7321effd","name":"Qwen3.7-Plus","organization":"Alibaba","date":"2026-06-03","domains":["Language","Vision","Multimodal"],"country":"China","reference":"https://qwen.ai/blog?id=qwen3.7-plus\n\nhttps://www.qwencloud.com/models/qwen3.7-plus","confidence":"Confident","accessibility":"API access"},{"id":"epoch:229c620dd807bd5eb9b7","name":"MAI-Thinking-1","organization":"Microsoft","date":"2026-06-02","domains":["Language"],"country":"United States of America","reference":"https://microsoft.ai/wp-content/uploads/2026/06/main_20260602_2.pdf","confidence":"Confident","accessibility":"API access"},{"id":"epoch:09b275b8da85ff529202","name":"MAI-Image-2.5","organization":"Microsoft","date":"2026-06-02","domains":["Image generation"],"country":"United States of America","reference":"https://microsoft.ai/news/introducing-mai-image-2-5/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:11215f66d077c43f533a","name":"MAI-Code-1-Flash","organization":"Microsoft","date":"2026-06-02","domains":["Language"],"country":"United States of America","reference":"https://microsoft.ai/news/introducingmai-code-1-flash/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:5cdd3bd86a742e20f2ce","name":"MiniMax-M3","organization":"MiniMax","date":"2026-06-01","domains":["Language","Multimodal","Vision","Video"],"country":"China","reference":"https://www.minimax.io/blog/minimax-m3","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:b59a6c306a3e8cd879b2","name":"Step 3.7 Flash","organization":"StepFun","date":"2026-05-29","domains":["Language","Multimodal","Vision"],"country":"China","reference":"https://huggingface.co/stepfun-ai/Step-3.7-Flash","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0bf10c189d81a3f0019d","name":"Grok Build 0.1","organization":"xAI","date":"2026-05-29","domains":["Language"],"country":"United States of America","reference":"https://x.ai/news/grok-build-0-1","confidence":"Confident","accessibility":"API access"},{"id":"epoch:b6bc991458f6917d30ae","name":"Claude Opus 4.8","organization":"Anthropic","date":"2026-05-28","domains":["Language"],"country":"United States of America","reference":"https://www.anthropic.com/news/claude-opus-4-8","confidence":"Confident","accessibility":"API access"},{"id":"epoch:74c75a0ddc8db9cc09e3","name":"Command A+","organization":"Cohere,Cohere Labs (formerly Cohere for AI)","date":"2026-05-20","domains":["Language","Multimodal","Vision"],"country":"Canada,Canada","reference":"https://docs.cohere.com/changelog/command-a-plus-05-2026","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a756ad75acf515907eab","name":"Gemini 3.5 Flash","organization":"Google DeepMind","date":"2026-05-19","domains":["Language","Multimodal","Audio","Video","Vision"],"country":"United States of America","reference":"https://deepmind.google/models/model-cards/gemini-3-5-flash/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:879504912cfa7719b749","name":"Qwen3.7-Max","organization":"Alibaba","date":"2026-05-19","domains":["Language"],"country":"China","reference":"https://qwen.ai/blog?id=qwen3.7","confidence":"Confident","accessibility":"API access"},{"id":"epoch:ab3acbbdcab183211d9e","name":"Composer 2.5","organization":"Cursor","date":"2026-05-18","domains":["Language"],"country":"United States of America","reference":"https://cursor.com/blog/composer-2-5","confidence":"Likely","accessibility":"API access"},{"id":"epoch:1755cacd27b8a18f0ec5","name":"Ring-2.6-1T","organization":"Ant Group","date":"2026-05-14","domains":["Language"],"country":"China","reference":"https://huggingface.co/inclusionAI/Ring-2.6-1T","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f3f55d7c9260508249d6","name":"TML-Interaction-Small","organization":"Thinking Machines","date":"2026-05-11","domains":["Language","Multimodal","Audio","Video"],"country":"United States of America","reference":"https://thinkingmachines.ai/blog/interaction-models/","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:ae76e8bae6f942ab5a7b","name":"ERNIE 5.1","organization":"Baidu","date":"2026-05-09","domains":["Language"],"country":"China","reference":"https://ernie.baidu.com/blog/posts/ernie-5.1-0508-release/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:c8fddc6a9fb5ab75f71c","name":"GPT-5.5 Cyber","organization":"OpenAI","date":"2026-05-07","domains":["Cybersecurity"],"country":"United States of America","reference":"https://openai.com/index/gpt-5-5-with-trusted-access-for-cyber/","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f472db2f4ed139a499ed","name":"GPT-5.5 Instant","organization":"OpenAI","date":"2026-05-05","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://openai.com/index/gpt-5-5-instant/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:7ffd46da978d65483d1b","name":"Mistral Medium 3.5","organization":"Mistral AI","date":"2026-04-29","domains":["Language","Multimodal"],"country":"France","reference":"https://huggingface.co/mistralai/Mistral-Medium-3.5-128B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:375c245aa976d2a1f6a0","name":"Granite 4.1 30B","organization":"IBM","date":"2026-04-29","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/blog/ibm-granite/granite-4-1","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:cbf60930da30f0c2edd5","name":"Granite 4.1 8B","organization":"IBM","date":"2026-04-29","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/blog/ibm-granite/granite-4-1","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d13f4ddd46a63f0ef33d","name":"DeepSeek-V4-Pro","organization":"DeepSeek","date":"2026-04-24","domains":["Language"],"country":"China","reference":"https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:610d3c8957bbd565ab8a","name":"DeepSeek-V4-Flash","organization":"DeepSeek","date":"2026-04-24","domains":["Language"],"country":"China","reference":"https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:983c416750f4c2021793","name":"GPT-5.5 Pro","organization":"OpenAI","date":"2026-04-23","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://openai.com/index/introducing-gpt-5-5/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:34032a5f5f61c2c6a0b7","name":"GPT-5.5","organization":"OpenAI","date":"2026-04-23","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://openai.com/index/introducing-gpt-5-5/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:485bbf8e96149f181b0f","name":"Grok Voice Think Fast 1.0","organization":"xAI","date":"2026-04-23","domains":["Audio"],"country":"United States of America","reference":"https://x.ai/news/grok-voice-think-fast-1","confidence":"Likely","accessibility":"API access"},{"id":"epoch:b2ce4aa48bedb1dea6c1","name":"MiMo-V2.5-Pro","organization":"Xiaomi Corp","date":"2026-04-23","domains":["Language"],"country":"China","reference":"https://mimo.xiaomi.com/mimo-v2-5-pro/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1c5940505868ae1d5446","name":"Tencent Hy3 preview","organization":"Tencent","date":"2026-04-23","domains":["Language"],"country":"China","reference":"https://hy.tencent.com/hy3-preview","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:2673a798543ca20759d3","name":"Qwen3.6 27B","organization":"Alibaba","date":"2026-04-22","domains":["Language","Vision","Multimodal"],"country":"China","reference":"https://qwen.ai/blog?id=qwen3.6-27b \n\nhttps://huggingface.co/Qwen/Qwen3.6-27B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ec6c5eec8b20fee10fc4","name":"GPT Image 2","organization":"OpenAI","date":"2026-04-21","domains":["Image generation"],"country":"United States of America","reference":"https://openai.com/index/introducing-chatgpt-images-2-0/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:ffa72da0a4d9afce4efe","name":"Ling-2.6-flash","organization":"Ant Group","date":"2026-04-21","domains":["Language"],"country":"China","reference":"https://huggingface.co/inclusionAI/Ling-2.6-flash","confidence":"Confident","accessibility":""},{"id":"epoch:61087fa55644676c60f4","name":"Qwen 3.6-27B","organization":"","date":"2026-04-21","domains":["Language"],"country":"","reference":"https://qwen.ai/blog?id=qwen3.6-27b","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:478f875317e225cf29ed","name":"Kimi K2.6","organization":"Moonshot","date":"2026-04-20","domains":["Language"],"country":"China","reference":"https://www.kimi.com/ai-models/kimi-k2-6","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5a6cd58d35b4e91d8e22","name":"Grok 4.3 Beta","organization":"xAI","date":"2026-04-17","domains":["Language"],"country":"United States of America","reference":"https://grok.com/release-notes#:~:text=Apr%2017%2C%202026-,Grok%204.3%20Beta,-Grok%204.3%20is","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:9a4cc8abbf9dffa06bb8","name":"Claude Opus 4.7","organization":"Anthropic","date":"2026-04-16","domains":["Language"],"country":"United States of America","reference":"https://www.anthropic.com/news/claude-opus-4-7","confidence":"Likely","accessibility":"API access"},{"id":"epoch:16b92b1531480635224a","name":"\u03c00.7 (pi-0.7)","organization":"Physical Intelligence","date":"2026-04-16","domains":["Robotics","Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2604.15483","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2a9cc911ed7c1b02a812","name":"Gemini Flash 3.1 TTS","organization":"Google DeepMind","date":"2026-04-15","domains":["Audio"],"country":"United States of America","reference":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-flash-tts/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:85670df7f16a377233af","name":"Gemini Robotics-ER 1.6","organization":"Google","date":"2026-04-14","domains":["Robotics","Multimodal","Vision","Language","Video","Audio"],"country":"United States of America","reference":"https://deepmind.google/blog/gemini-robotics-er-1-6/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:e47b63422784a5872c3f","name":"GPT-5.4-Cyber","organization":"OpenAI","date":"2026-04-14","domains":["Cybersecurity","Language"],"country":"United States of America","reference":"https://openai.com/index/scaling-trusted-access-for-cyber-defense/","confidence":"Unknown","accessibility":"Limited access"},{"id":"epoch:58f722d5d09e787cd11c","name":"EXAONE 4.5","organization":"LG AI Research","date":"2026-04-09","domains":["Language","Multimodal"],"country":"South Korea","reference":"Eunbi Choi, Kibong Choi, Sehyun Chun, Seokhee Hong, Junwon Hwang, Hyojin Jeon, Ahra Jo, Hyunjik Jo, Yeonsik Jo, Joonkee Kim, Seonghwan Kim, Soyeon Kim, Sunkyoung Kim, Yireun Kim, Yongil Kim, Changhun Lee, Haeju Lee, Jinsik Lee, Kyungmin Lee, Sangha Park, Kwangrok Ryoo, Minju Seo, Sejong Yang, Heuiyeen Yeen, Hwan Chang, Stanley Jungkyu Choi, Yejin Choi, Kyubeen Han, Joonwon Jang, Kijeong Jeon, Geunyeong Jeong, Gerrard Jeongwon Jo, Jiyeon Jung, Daeseong Kim, Dohoon Kim, Dohyun Kim, Hyunseo Kim, Minu Kim, Myoungshin Kim, Youchul Kim, Byungoh Ko, Christopher Lee, Edward Hwayoung Lee, Honglak Lee, Jiyoung Lee, Sangeun Lee, Seungwon Lim, Woohyung Lim, Jueun Mun, Jaewoo Park, Jimin Park, Jinho Park, Yongmin Park, Wooseok Seo, Yongwoo Song, Sihyuk Yi, Kyungjae Yoo, Sangyeon Yoon","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:10cd362ccfd87ebe03be","name":"Muse Spark","organization":"Meta AI","date":"2026-04-08","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://ai.meta.com/blog/introducing-muse-spark-msl/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:5b69fde1cce50388aea6","name":"GLM-5.1","organization":"Z.ai (Zhipu AI)","date":"2026-04-07","domains":["Language"],"country":"China","reference":"https://z.ai/blog/glm-5.1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:af0a09a44ead649637c1","name":"SWE-1.6","organization":"Cognition","date":"2026-04-07","domains":["Language"],"country":"United States of America","reference":"https://cognition.com/blog/swe-1-6","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:ca731361ae3767207ba1","name":"Claude Mythos Preview","organization":"Anthropic","date":"2026-04-07","domains":["Language","Multimodal"],"country":"United States of America","reference":"https://www-cdn.anthropic.com/7624816413e9b4d2e3ba620c5a5e091b98b190a5/Claude%20Mythos%20Preview%20System%20Card.pdf\nhttps://www.anthropic.com/glasswing","confidence":"Unknown","accessibility":"Limited access"},{"id":"epoch:e72cc9df044c662c4f2c","name":"MolmoAct 2","organization":"Allen Institute for AI,University of Washington,National University of Singapore,University of Pennsylvania,Johns Hopkins University,Amazon,Cortex AI","date":"2026-04-05","domains":["Vision","Language","Robotics"],"country":"United States of America,United States of America,Singapore,United States of America,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2605.02881","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:eae9ecca8b217527600a","name":"Gemma 4 31B IT","organization":"Google DeepMind","date":"2026-04-02","domains":["Language"],"country":"United States of America","reference":"","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:106734caf2c98b05d6aa","name":"Gemma 4 26B A4B","organization":"Google DeepMind","date":"2026-04-02","domains":["Language"],"country":"United States of America","reference":"https://ai.google.dev/gemma/docs/core/model_card_4","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c7cecc8eec2e23b74878","name":"Qwen 3.6 Plus","organization":"Alibaba","date":"2026-04-01","domains":["Language","Multimodal"],"country":"China","reference":"https://qwen.ai/blog?id=qwen3.6","confidence":"Confident","accessibility":"API access"},{"id":"epoch:3391e6d972b7a1f383d3","name":"Qwen3.5-Omni-Plus","organization":"Alibaba","date":"2026-03-29","domains":["Language","Multimodal","Vision","Audio"],"country":"China","reference":"https://arxiv.org/abs/2604.15804","confidence":"Likely","accessibility":"API access"},{"id":"epoch:2cf472d38063d6c35bd5","name":"Qwen3.5-Omni-Flash","organization":"Alibaba","date":"2026-03-29","domains":["Language","Multimodal","Vision","Audio"],"country":"China","reference":"https://arxiv.org/abs/2604.15804","confidence":"Likely","accessibility":"API access"},{"id":"epoch:88838f9aff9cf31d5e18","name":"Composer 2","organization":"Cursor","date":"2026-03-19","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2603.24477","confidence":"Confident","accessibility":"API access"},{"id":"epoch:ddd3b4509a777bf741dc","name":"MiniMax-M2.7","organization":"MiniMax","date":"2026-03-18","domains":["Language"],"country":"China","reference":"https://www.minimax.io/news/minimax-m27-en","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:84a3eecbf78681c14f14","name":"MiMo-V2-Pro","organization":"Xiaomi Corp","date":"2026-03-18","domains":["Language"],"country":"China","reference":"https://mimo.xiaomi.com/mimo-v2-pro","confidence":"Confident","accessibility":"API access"},{"id":"epoch:b6581c60fb667b456329","name":"MiMo-V2-Omni","organization":"Xiaomi Corp","date":"2026-03-18","domains":["Vision","Language","Audio","Video"],"country":"China","reference":"https://mimo.xiaomi.com/mimo-v2-omni","confidence":"Confident","accessibility":"API access"},{"id":"epoch:f8bb895ca42da9ac3563","name":"GPT-5.4 Mini","organization":"OpenAI","date":"2026-03-17","domains":["Language"],"country":"United States of America","reference":"https://openai.com/index/introducing-gpt-5-4-mini-and-nano/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:170e15209a07ea1a6104","name":"GPT-5.4 Nano","organization":"OpenAI","date":"2026-03-17","domains":["Language"],"country":"United States of America","reference":"https://openai.com/index/introducing-gpt-5-4-mini-and-nano/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:e908826740ea38fefd7d","name":"Mistral Small 4","organization":"Mistral AI","date":"2026-03-16","domains":["Language","Multimodal","Vision"],"country":"France","reference":"https://mistral.ai/news/mistral-small-4/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7942801966ec19f644d4","name":"Nemotron 3 Super","organization":"Nvidia","date":"2026-03-11","domains":["Language"],"country":"United States of America","reference":"https://developer.nvidia.com/blog/introducing-nemotron-3-super-an-open-hybrid-mamba-transformer-moe-for-agentic-reasoning/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0c7b277fbb155aef222c","name":"GPT-5.4","organization":"OpenAI","date":"2026-03-05","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://openai.com/index/introducing-gpt-5-4/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:c85cdaf9ab2a1b256c9b","name":"GPT-5.4 Pro","organization":"OpenAI","date":"2026-03-05","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://openai.com/index/introducing-gpt-5-4/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:bf19e9f9e5eca32a3b2c","name":"Gemini 3.1 Flash-Lite","organization":"Google","date":"2026-03-03","domains":["Language","Multimodal"],"country":"United States of America","reference":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-flash-lite/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:ec14626b199b4aadd3d1","name":"Gemini 3.0 Flash-lite","organization":"Google DeepMind","date":"2026-03-03","domains":["Language"],"country":"United States of America","reference":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-flash-lite/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:9969413dcd4e64f5c5d6","name":"SWE 1.6","organization":"Cognition","date":"2026-03-01","domains":["Language"],"country":"United States of America","reference":"https://cognition.ai/blog/swe-1-6-preview","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:ac00fefd227a2a020881","name":"Qwen 3.5 Flash (hosted 35B-A3B)","organization":"Alibaba","date":"2026-02-25","domains":[],"country":"China","reference":"","confidence":"Likely","accessibility":""},{"id":"epoch:c75e9f16e73d0b767fb4","name":"Qwen3.5-122B-A10B","organization":"Alibaba","date":"2026-02-24","domains":["Language"],"country":"China","reference":"https://github.com/QwenLM/Qwen3.5?tab=readme-ov-file ","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:e8a6abf56830297b492f","name":"Qwen3.5-27B","organization":"Alibaba","date":"2026-02-24","domains":["Language"],"country":"China","reference":"https://github.com/QwenLM/Qwen3.5?tab=readme-ov-file ","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:2c870bab654dcaa1f35c","name":"Qwen3.5-9B","organization":"Alibaba","date":"2026-02-24","domains":["Language"],"country":"China","reference":"https://github.com/QwenLM/Qwen3.5?tab=readme-ov-file ","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:ba764ad003e7d977e884","name":"Qwen3.5-4B","organization":"Alibaba","date":"2026-02-24","domains":["Language"],"country":"China","reference":"https://github.com/QwenLM/Qwen3.5?tab=readme-ov-file ","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:f904ace180d6e7cbd6e3","name":"Qwen3.5-2B","organization":"Alibaba","date":"2026-02-24","domains":["Language"],"country":"China","reference":"https://github.com/QwenLM/Qwen3.5?tab=readme-ov-file ","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:e2f5c8e46181f95fe2bb","name":"Qwen3.5-0.8B","organization":"Alibaba","date":"2026-02-24","domains":["Language"],"country":"China","reference":"https://github.com/QwenLM/Qwen3.5?tab=readme-ov-file ","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:65a29bb3983edd3a8f17","name":"Qwen3.5-35B-A3B","organization":"Alibaba","date":"2026-02-24","domains":["Language"],"country":"China","reference":"https://huggingface.co/Qwen/Qwen3.5-35B-A3B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:26f5729c8b3d89e1ddb3","name":"Gemini 3.1 Pro","organization":"Google DeepMind","date":"2026-02-19","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:4a4949c193d182667168","name":"Arcee Trinity Large","organization":"Prime Intellect,Arcee AI","date":"2026-02-19","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2602.17004","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:7c5edc45e41a3a1b319c","name":"Claude Sonnet 4.6","organization":"Anthropic","date":"2026-02-17","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://www.anthropic.com/news/claude-sonnet-4-6","confidence":"Likely","accessibility":"API access"},{"id":"epoch:1f8df1f8d34ae8e608ac","name":"Grok 4.20","organization":"xAI","date":"2026-02-17","domains":["Language"],"country":"United States of America","reference":"https://docs.x.ai/developers/models","confidence":"Likely","accessibility":"API access"},{"id":"epoch:0393e6c54fad1cd3650a","name":"DreamZero","organization":"Nvidia","date":"2026-02-17","domains":["Robotics"],"country":"United States of America","reference":"https://arxiv.org/abs/2602.15922","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:49a96ab1300e2c0c5f71","name":"Qwen 3.5 Plus (hosted 397B-A17B)","organization":"Alibaba","date":"2026-02-16","domains":[],"country":"China","reference":"","confidence":"Likely","accessibility":""},{"id":"epoch:dad8d348286df9176cd5","name":"Qwen3.5 397B-A17B","organization":"Alibaba","date":"2026-02-15","domains":["Language","Vision"],"country":"China","reference":"https://qwen.ai/blog?id=qwen3.5","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6d6a7d85b7337be2fdb9","name":"GLM-5","organization":"Z.ai (Zhipu AI)","date":"2026-02-12","domains":["Language"],"country":"China","reference":"https://arxiv.org/pdf/2602.15763 \n\nhttps://z.ai/blog/glm-5 ","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:22a595cce118707f7898","name":"MiniMax-M2.5","organization":"MiniMax","date":"2026-02-12","domains":["Language"],"country":"China","reference":"https://www.minimax.io/news/minimax-m25","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:28ea57b78e0efebec0ed","name":"Seedance 2.0","organization":"ByteDance","date":"2026-02-12","domains":["Image generation","Video","Audio"],"country":"China","reference":"https://seed.bytedance.com/en/seedance2_0","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:2b98eaffc8619c35474a","name":"GPT-5.3-Codex-Spark","organization":"OpenAI","date":"2026-02-12","domains":["Language"],"country":"United States of America","reference":"https://openai.com/index/introducing-gpt-5-3-codex-spark/","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:f54ba1b962bbfa8ab900","name":"Composer 1.5","organization":"Cursor","date":"2026-02-09","domains":["Language"],"country":"United States of America","reference":"https://cursor.com/blog/composer-1-5","confidence":"Likely","accessibility":"API access"},{"id":"epoch:1dd690ba4fd47724df0c","name":"Claude Opus 4.6","organization":"Anthropic","date":"2026-02-05","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://www.anthropic.com/news/claude-opus-4-6","confidence":"Likely","accessibility":"API access"},{"id":"epoch:a0b7ba5ab9aeed707c10","name":"GPT-5.3 Codex","organization":"OpenAI","date":"2026-02-05","domains":["Language"],"country":"United States of America","reference":"https://openai.com/index/introducing-gpt-5-3-codex/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:8c163add5edbe1bac26b","name":"Step 3.5 Flash","organization":"StepFun","date":"2026-02-02","domains":["Language"],"country":"China","reference":"https://static.stepfun.com/blog/step-3.5-flash/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0d782a9fd8f1847707cb","name":"Kimi K2.5","organization":"Moonshot","date":"2026-02-02","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2602.02276","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1c90fdfd5611642e8868","name":"Qwen3-Coder-Next","organization":"Alibaba","date":"2026-02-02","domains":["Language"],"country":"China","reference":"https://qwen.ai/blog?id=qwen3-coder-next","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:abb90e35858f05c3300d","name":"Helix 02","organization":"Figure AI","date":"2026-01-27","domains":["Robotics"],"country":"United States of America","reference":"https://www.figure.ai/news/helix-02","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:8752ece372ac9dccdebc","name":"Solar Pro 3","organization":"Upstage","date":"2026-01-27","domains":["Language"],"country":"South Korea","reference":"https://www.upstage.ai/blog/en/solar-pro-3-0127","confidence":"Confident","accessibility":"API access"},{"id":"epoch:63844df01a2de92b5da9","name":"Qwen3-Max-Thinking","organization":"Alibaba","date":"2026-01-25","domains":["Language"],"country":"China","reference":"https://qwen.ai/blog?id=qwen3-max-thinking","confidence":"Likely","accessibility":"API access"},{"id":"epoch:79735fc52bf59d0d6c15","name":"GLM 4.7 Flash","organization":"Z.ai (Zhipu AI)","date":"2026-01-19","domains":["Language"],"country":"China","reference":"https://huggingface.co/zai-org/GLM-4.7-Flash","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:59c94aaee2a0bb196cf7","name":"Solar Open 100B","organization":"Upstage","date":"2025-12-31","domains":["Language"],"country":"South Korea","reference":"https://huggingface.co/upstage/Solar-Open-100B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:e522d8577d61c7974eac","name":"K-EXAONE","organization":"LG AI Research","date":"2025-12-31","domains":["Language"],"country":"South Korea","reference":"https://huggingface.co/LGAI-EXAONE","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:19250504a73e36311588","name":"VAETKI","organization":"NC AI","date":"2025-12-30","domains":["Language"],"country":"South Korea","reference":"https://huggingface.co/NC-AI-consortium-VAETKI/VAETKI","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5adb9034eed4fe41582b","name":"A.X K1","organization":"SK Telecom","date":"2025-12-30","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/abs/2601.09200 \n\nhttps://huggingface.co/skt/A.X-K1 ","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:502a6c60bc0e7b37f6e7","name":"HyperCLOVA X SEED 32B Think","organization":"NAVER","date":"2025-12-29","domains":["Multimodal","Language","Vision"],"country":"South Korea","reference":"https://huggingface.co/naver-hyperclovax/HyperCLOVAX-SEED-Think-32B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:af25685532c535bb6df5","name":"MiniMax-M2.1","organization":"MiniMax","date":"2025-12-23","domains":["Language"],"country":"China","reference":"https://huggingface.co/MiniMaxAI/MiniMax-M2.1","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:86b359dbdb31d629f765","name":"GLM-4.7","organization":"Z.ai (Zhipu AI)","date":"2025-12-22","domains":["Language"],"country":"China","reference":"https://z.ai/blog/glm-4.7","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:53bef73c9a37fe9e88f7","name":"GPT-5.2 Codex","organization":"OpenAI","date":"2025-12-18","domains":["Language"],"country":"United States of America","reference":"https://openai.com/index/introducing-gpt-5-2-codex/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:e87baec699793a13f69c","name":"Doubao Seed 1.8","organization":"ByteDance","date":"2025-12-18","domains":["Language","Vision","Multimodal","Video"],"country":"China","reference":"https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/research/Seed-1.8-Modelcard.pdf\n\nhttps://seed.bytedance.com/en/seed1_8 ","confidence":"Confident","accessibility":"API access"},{"id":"epoch:d6106d598fc73ef2a152","name":"Gemini 3 Flash","organization":"Google DeepMind","date":"2025-12-17","domains":[],"country":"United States of America","reference":"","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:90562e2b3e6e1882347c","name":"MiMo-V2-Flash","organization":"Xiaomi Corp","date":"2025-12-16","domains":["Language"],"country":"China","reference":"https://github.com/XiaomiMiMo/MiMo-V2-Flash/blob/main/paper.pdf","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0cb37aa18a94add5217a","name":"Nemotron 3-Nano-30B-A3B","organization":"Nvidia","date":"2025-12-15","domains":["Language"],"country":"United States of America","reference":"https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Nano-Technical-Report.pdf","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:3f4d25858c08a69e4b8b","name":"Nemotron-Cascade 14B","organization":"Nvidia","date":"2025-12-15","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2512.13607","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e2dbc91aa07d949c2f60","name":"GPT-5.2","organization":"OpenAI","date":"2025-12-11","domains":[],"country":"United States of America","reference":"","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:680df17d868527a1a84c","name":"GPT-5.2 Pro","organization":"OpenAI","date":"2025-12-11","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://openai.com/index/introducing-gpt-5-2/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:7f50014373ac0f4a90b2","name":"Nomos 1","organization":"Nous Research","date":"2025-12-11","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/NousResearch/nomos-1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:041bd113b20556507f1c","name":"Devstral 2 (123B)","organization":"Mistral AI","date":"2025-12-09","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/devstral-2-vibe-cli","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:3f5b59d4408e24610fbe","name":"Devstral 2 (24B)","organization":"Mistral AI","date":"2025-12-09","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/devstral-2-vibe-cli","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:7a20dacb8e19b70d2d10","name":"Devstral Small 2","organization":"Mistral AI","date":"2025-12-09","domains":["Language","Vision","Multimodal"],"country":"France","reference":"https://mistral.ai/news/devstral-2-vibe-cli/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8ca06a15e0834dfc9aae","name":"SIMA 2","organization":"Google DeepMind","date":"2025-12-04","domains":["3D modeling"],"country":"United States of America","reference":"https://arxiv.org/abs/2512.04797v1","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:f2e5b036dd35ebd0bf9c","name":"Nova 2","organization":"Amazon Web Services (AWS)","date":"2025-12-02","domains":["Language","Vision","Multimodal"],"country":"","reference":"https://aws.amazon.com/about-aws/whats-new/2025/12/nova-2-foundation-models-amazon-bedrock/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:136e4fae5548c41ea399","name":"Nova 2 Pro (Preview)","organization":"Amazon","date":"2025-12-02","domains":["Language"],"country":"United States of America","reference":"https://aws.amazon.com/about-aws/whats-new/2025/12/nova-2-foundation-models-amazon-bedrock/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4a79fa47584a43d31fb5","name":"Ministral 3 14B","organization":"Mistral AI","date":"2025-12-02","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/mistral-3","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e1ded62ea1482b528879","name":"Ministral 3 8B","organization":"Mistral AI","date":"2025-12-02","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/mistral-3","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ba1dcf990da763c6edd4","name":"Ministral 3 3B","organization":"Mistral AI","date":"2025-12-02","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/mistral-3","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f047eb7d38b35f0bf091","name":"Mistral Large 3","organization":"Mistral AI","date":"2025-12-02","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/mistral-3/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:075ce2e1b9b9831ee481","name":"Nova 2.0 Pro Preview","organization":"Amazon","date":"2025-12-02","domains":["Language","Multimodal","Audio","Video"],"country":"United States of America","reference":"https://cdn.amazon.science/c5/3d/84514a224666b5be6de4b43ef4aa/nova-2-0-technical-report2.pdf","confidence":"Confident","accessibility":"API access"},{"id":"epoch:2f3ee310859dc8e22678","name":"DeepSeek-V3.2","organization":"DeepSeek","date":"2025-12-01","domains":[],"country":"China","reference":"https://api-docs.deepseek.com/news/news251201","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:cae88c0faa7cd4816c5c","name":"Runway Gen-4.5","organization":"Runway","date":"2025-12-01","domains":["Video"],"country":"United States of America","reference":"https://runwayml.com/research/introducing-runway-gen-4.5","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:79de87c2bbcc3ccf5061","name":"DeepSeekMath-V2","organization":"DeepSeek","date":"2025-11-27","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2511.22570","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:91761965cca4e08edfc0","name":"INTELLECT-3","organization":"Prime Intellect","date":"2025-11-26","domains":["Language"],"country":"United States of America","reference":"https://storage.googleapis.com/intellect-3-paper/INTELLECT_3_Technical_Report.pdf","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:33000eaebd8ff22e003c","name":"FLUX.2","organization":"Black Forest Labs","date":"2025-11-25","domains":["Image generation"],"country":"Germany","reference":"https://bfl.ai/blog/flux-2, https://github.com/black-forest-labs/flux2","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:aeffeace2acb935c08dd","name":"Claude Opus 4.5","organization":"Anthropic","date":"2025-11-24","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://www.anthropic.com/news/claude-opus-4-5","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:8d89b09d4179957d70ba","name":"Olmo 3","organization":"Allen Institute for AI","date":"2025-11-20","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2512.13961","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b48927fbdd66ecc8c392","name":"Gemini 3 Pro Image (Nano Banana Pro)","organization":"Google DeepMind","date":"2025-11-20","domains":["Image generation"],"country":"United States of America","reference":"https://blog.google/innovation-and-ai/products/nano-banana-pro/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:b43df9a83f83be6b5c4f","name":"SAM 3","organization":"","date":"2025-11-20","domains":["Vision","Video"],"country":"","reference":"https://arxiv.org/abs/2511.16719, https://github.com/facebookresearch/sam3","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:5691e80707a7c847df44","name":"Olmo 3 32B Instruct","organization":"Allen Institute for AI,University of Washington,Carnegie Mellon University (CMU),Stanford University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Universit\u00e9 de Montr\u00e9al,Princeton University,Massachusetts Institute of Technology (MIT),University of Maryland","date":"2025-11-20","domains":["Language"],"country":"United States of America,United States of America,United States of America,United States of America,Canada,Canada,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2512.13961, https://allenai.org/blog/olmo3, https://huggingface.co/allenai/Olmo-3-1125-32B","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1f1c7757591daa34a5d1","name":"Olmo 3 7B Think","organization":"Allen Institute for AI,University of Washington,Carnegie Mellon University (CMU),Stanford University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Universit\u00e9 de Montr\u00e9al,Princeton University,Massachusetts Institute of Technology (MIT),University of Maryland","date":"2025-11-20","domains":["Language"],"country":"United States of America,United States of America,United States of America,United States of America,Canada,Canada,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2512.13961, https://allenai.org/blog/olmo3, https://huggingface.co/allenai/Olmo-3-1125-32B","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9421f210cc0f386fb009","name":"Olmo 3.1 32B Think","organization":"Allen Institute for AI,University of Washington,Carnegie Mellon University (CMU),Stanford University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Universit\u00e9 de Montr\u00e9al,Princeton University,Massachusetts Institute of Technology (MIT),University of Maryland","date":"2025-11-20","domains":["Language"],"country":"United States of America,United States of America,United States of America,United States of America,Canada,Canada,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2512.13961, https://allenai.org/blog/olmo3, https://huggingface.co/allenai/Olmo-3-1125-32B","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d9a1b26c323d94191c63","name":"Grok 4.1 Fast","organization":"xAI","date":"2025-11-19","domains":["Language"],"country":"United States of America","reference":"https://x.ai/news/grok-4-1-fast","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:43ba891022df010eb05a","name":"GPT-5.1-Codex-Max","organization":"OpenAI","date":"2025-11-19","domains":["Language"],"country":"United States of America","reference":"https://openai.com/index/gpt-5-1-codex-max/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:a544fad233f2d0df42e0","name":"Cogito v2.1","organization":"Deep Cogito","date":"2025-11-19","domains":["Language"],"country":"","reference":"https://huggingface.co/deepcogito/cogito-671b-v2.1\n\nhttps://github.com/DeepCogito/cogito-v2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:081938d269e62da1c11f","name":"Gemini 3 Pro","organization":"Google DeepMind","date":"2025-11-18","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://blog.google/products/gemini/gemini-3/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:f2515e9138058061fa8a","name":"Gemini 3 Deep Think","organization":"Google DeepMind","date":"2025-11-18","domains":["Language"],"country":"United States of America","reference":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-deep-think/","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:adade8dd9f254a5245f5","name":"Grok 4.1","organization":"xAI","date":"2025-11-17","domains":["Language"],"country":"United States of America","reference":"https://x.ai/news/grok-4-1","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:4e7f8efdbb96194e95c4","name":"\u03c00.6 (pi-0.6)","organization":"Physical Intelligence","date":"2025-11-17","domains":["Robotics","Vision"],"country":"United States of America","reference":"https://www.pi.website/download/pistar06.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0195d2bfb8f382f1166a","name":"P1-235B-A22B","organization":"Shanghai AI Lab","date":"2025-11-17","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2511.13612, https://github.com/PRIME-RL/P1","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c539cb1df1697a5969fc","name":"Depth Anything 3","organization":"","date":"2025-11-14","domains":["Vision","3D modeling"],"country":"","reference":"https://arxiv.org/abs/2511.10647, https://github.com/ByteDance-Seed/Depth-Anything-3, https://depth-anything-3.github.io/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c39202cfc8581c76d1f6","name":"GPT-5.1","organization":"OpenAI","date":"2025-11-13","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://openai.com/index/gpt-5-1/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:0226653d4e22c3a669fd","name":"GPT-5.1 Instant","organization":"OpenAI","date":"2025-11-13","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://openai.com/index/gpt-5-1/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:64a3e162462a03f32f1a","name":"GPT-5.1-Codex-mini","organization":"OpenAI","date":"2025-11-12","domains":["Language","Multimodal"],"country":"United States of America","reference":"","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:d62c14442fdf50f558c5","name":"GPT-5.1-Codex","organization":"OpenAI","date":"2025-11-12","domains":["Language"],"country":"United States of America","reference":"","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:349511a4c2c27853a237","name":"Meta's Generative Ads Model (GEM)","organization":"Meta AI","date":"2025-11-10","domains":["Recommendation"],"country":"United States of America","reference":"https://engineering.fb.com/2025/11/10/ml-applications/metas-generative-ads-model-gem-the-central-brain-accelerating-ads-recommendation-ai-innovation/","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:9b593af386e3e85f66a5","name":"Kaiju Large","organization":"Character.ai","date":"2025-11-07","domains":["Language"],"country":"United States of America","reference":"https://blog.character.ai/inside-kaiju-building-conversational-models-at-scale/ ","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:9ac88e88662c3708b719","name":"Kaiju Medium","organization":"Character.ai","date":"2025-11-07","domains":["Language"],"country":"United States of America","reference":"https://blog.character.ai/inside-kaiju-building-conversational-models-at-scale/ ","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:f684f7af5403a2dcb89b","name":"Kaiju Small","organization":"Character.ai","date":"2025-11-07","domains":["Language"],"country":"United States of America","reference":"https://blog.character.ai/inside-kaiju-building-conversational-models-at-scale/ ","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:7b8547d6a3308103b345","name":"Kimi K2 Thinking","organization":"Moonshot","date":"2025-11-06","domains":["Language"],"country":"China","reference":"https://moonshotai.github.io/Kimi-K2/thinking","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:a67654651def79c59d77","name":"Gen-0","organization":"Generalist","date":"2025-11-04","domains":["Robotics"],"country":"United States of America","reference":"https://generalistai.com/blog/nov-04-2025-GEN-0","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:6d1a3d66aa9202513ef4","name":"Emu3.5","organization":"Beijing Academy of Artificial Intelligence / BAAI","date":"2025-10-30","domains":["Video","Multimodal","Image generation","Vision","Language","Speech"],"country":"China","reference":"https://emu.world/Emu35_tech_report.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b4af9537942bc841be6d","name":"Kimi Linear","organization":"Moonshot","date":"2025-10-30","domains":["Language"],"country":"China","reference":"https://github.com/MoonshotAI/Kimi-Linear/blob/master/tech_report.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:aa8ac453c448706f20bf","name":"Composer","organization":"Cursor","date":"2025-10-29","domains":["Language"],"country":"United States of America","reference":"https://cursor.com/blog/composer","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:9bf964cbb93a2d750e4e","name":"SWE-1.5","organization":"Cognition","date":"2025-10-29","domains":["Language"],"country":"United States of America","reference":"https://cognition.ai/blog/swe-1-5","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:dac367ca21d03903903f","name":"Tongyi DeepResearch","organization":"Alibaba","date":"2025-10-28","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2510.24701","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b2a8afa9d867574e99de","name":"MiniMax-M2","organization":"MiniMax","date":"2025-10-27","domains":["Language"],"country":"China","reference":"https://www.minimax.io/news/minimax-m2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ff2eb268e9efa9d5bf47","name":"LoongRL 7B","organization":"Microsoft Research Asia,Shanghai Jiao Tong University,Carnegie Mellon University (CMU)","date":"2025-10-27","domains":["Language"],"country":"China,China,United States of America","reference":"https://arxiv.org/abs/2510.19363","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:386109e71cdc80a56570","name":"LoongRL 14B","organization":"Microsoft Research Asia,Shanghai Jiao Tong University,Carnegie Mellon University (CMU)","date":"2025-10-27","domains":["Language"],"country":"China,China,United States of America","reference":"https://arxiv.org/abs/2510.19363","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b38f679ebcf125bee813","name":"Lapa LLM","organization":"Ukrainian Catholic University,Igor Sikorsky Kyiv Polytechnic Institute,AGH University of Krakow,Lviv Polytechnic","date":"2025-10-25","domains":["Language","Multimodal"],"country":"Ukraine,Ukraine,Poland,Ukraine","reference":"https://huggingface.co/spaces/lapa-llm/lapa","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:670df5d18df4ed2e2720","name":"Ring-mini-linear-2.0","organization":"Ant Group","date":"2025-10-23","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2510.19338","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:59b9c7faeded38b756dc","name":"Ring-flash-linear-2.0","organization":"Ant Group","date":"2025-10-23","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2510.19338","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1c52f8d948795ae26e03","name":"Deepseek OCR","organization":"DeepSeek","date":"2025-10-21","domains":["Vision","Language"],"country":"China","reference":"https://arxiv.org/abs/2510.18234v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e438575ad2dbda08c43d","name":"BAPO 32B","organization":"Fudan University,Shanghai Qiji Zhifeng","date":"2025-10-21","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2510.18927","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b23fc5e1ee932d41ee4b","name":"Odyssey 102B","organization":"Anthrogen","date":"2025-10-18","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/10.1101/2025.10.15.682677v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:bf94d0a6aae0d037b431","name":"Odyssey 12B","organization":"Anthrogen","date":"2025-10-18","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/10.1101/2025.10.15.682677v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c63852c76ab3f11c3a9b","name":"Odyssey 1.2B","organization":"Anthrogen","date":"2025-10-18","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/10.1101/2025.10.15.682677v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:35d210cdd2bb0f2201ef","name":"Claude Haiku 4.5","organization":"Anthropic","date":"2025-10-15","domains":["Language"],"country":"United States of America","reference":"https://www.anthropic.com/news/claude-haiku-4-5","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:66cc7031561bf62d77a6","name":"Veo 3.1","organization":"Google DeepMind","date":"2025-10-15","domains":["Video","Vision"],"country":"United States of America","reference":"https://blog.google/technology/ai/veo-updates-flow/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:9eeb546ac898582099b0","name":"Llama 4 Scout + ScaleRL","organization":"Meta AI","date":"2025-10-15","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2510.13786","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:63a50bcfe78f00ebe5f7","name":"MAI-Image-1","organization":"Microsoft","date":"2025-10-13","domains":["Image generation"],"country":"United States of America","reference":"https://microsoft.ai/news/introducing-mai-image-1-debuting-in-the-top-10-on-lmarena/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:30b8fca504eff135530a","name":"C2S-Scale","organization":"Google Research,Yale University,Google DeepMind,Brown University,University of Southern California","date":"2025-10-11","domains":["Language","Biology"],"country":"United States of America,United States of America,United States of America,United States of America,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2025.04.14.648850v2","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b292c509592906e8200d","name":"Ring-1T","organization":"Ant Group","date":"2025-10-10","domains":["Language"],"country":"China","reference":"https://huggingface.co/inclusionAI/Ring-1T","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5d5e21c9354e4536eec6","name":"Ling-1T","organization":"Ant Group","date":"2025-10-10","domains":["Language"],"country":"China","reference":"https://huggingface.co/inclusionAI/Ling-1T\nhttps://arxiv.org/pdf/2510.22115","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b54798bd81b445638b7a","name":"Grok Imagine","organization":"xAI","date":"2025-10-08","domains":["Image generation","Video"],"country":"United States of America","reference":"https://grokimagine.ai/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:260bae65d688a6176682","name":"GPT-5 Pro","organization":"OpenAI","date":"2025-10-07","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://cdn.openai.com/gpt-5-system-card.pdf","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:143513dd04c135807c79","name":"Gemini 2.5 Computer Use","organization":"Google","date":"2025-10-07","domains":["Language","Vision","Video","Multimodal","Speech"],"country":"United States of America","reference":"https://blog.google/technology/google-deepmind/gemini-computer-use-model/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:de022b5763435a23e563","name":"Tiny Recursive Model (TRM-Att)","organization":"Samsung SAIT AI Lab","date":"2025-10-06","domains":["Language","Vision","Multimodal"],"country":"South Korea","reference":"https://arxiv.org/abs/2510.04871","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:43d2037f33c42f14373f","name":"Granite-4.0-H-Tiny","organization":"IBM","date":"2025-10-02","domains":["Language"],"country":"United States of America","reference":"https://www.ibm.com/new/announcements/ibm-granite-4-0-hyper-efficient-high-performance-hybrid-models#:~:text=We're%20launching%20Granite%204,costs%20compared%20to%20conventional%20LLMs.","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e0e790b0e05a3b7d061a","name":"Granite-4.0-H-Micro","organization":"IBM","date":"2025-10-02","domains":["Language"],"country":"United States of America","reference":"https://www.ibm.com/new/announcements/ibm-granite-4-0-hyper-efficient-high-performance-hybrid-models#:~:text=We're%20launching%20Granite%204,costs%20compared%20to%20conventional%20LLMs.","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:dcb5f85853791d6212e8","name":"Granite-4.0-H-Small","organization":"IBM","date":"2025-10-02","domains":["Language"],"country":"United States of America","reference":"https://www.ibm.com/new/announcements/ibm-granite-4-0-hyper-efficient-high-performance-hybrid-models#:~:text=We're%20launching%20Granite%204,costs%20compared%20to%20conventional%20LLMs.","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2cba2f895fd56811febf","name":"Octave 2","organization":"Hume","date":"2025-10-01","domains":["Speech"],"country":"United States of America","reference":"https://www.hume.ai/blog/octave-2-launch","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:cd749e2bf9f8a421ebdf","name":"EVI 4 mini","organization":"Hume","date":"2025-10-01","domains":["Speech"],"country":"United States of America","reference":"https://www.hume.ai/blog/octave-2-launch","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:e757efa786032276c02b","name":"GLM-4.6","organization":"Z.ai (Zhipu AI),Tsinghua University","date":"2025-09-30","domains":["Language"],"country":"China,China","reference":"https://z.ai/blog/glm-4.6","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b0ed01c2bd0eb9902a3d","name":"Sora 2.0","organization":"OpenAI","date":"2025-09-30","domains":["Video"],"country":"United States of America","reference":"https://openai.com/index/sora-2/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:62c7cc06f36f17afd291","name":"Kandinsky 5.0 Video Lite","organization":"Sber","date":"2025-09-30","domains":["Video"],"country":"Russia","reference":"https://habr.com/ru/companies/sberbank/articles/951800/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:786a499e26a6ec501b21","name":"Claude Sonnet 4.5","organization":"Anthropic","date":"2025-09-29","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://www.anthropic.com/news/claude-sonnet-4-5","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:4b1ef722a052b453449c","name":"NVIDIA Isaac GR00T N1.6","organization":"Nvidia","date":"2025-09-29","domains":["Robotics","Vision","Language"],"country":"United States of America","reference":"https://nvidianews.nvidia.com/news/nvidia-accelerates-robotics-research-and-development-with-new-open-models-and-simulation-libraries","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:ce2b4f27cbe29c1552f2","name":"Cosmos-Transfer2.5-2B","organization":"Nvidia","date":"2025-09-29","domains":["Video","Image generation"],"country":"United States of America","reference":"https://research.nvidia.com/labs/dir/cosmos-transfer2.5/\n\nhttps://research.nvidia.com/publication/2025-09_world-simulation-video-foundation-models-physical-ai","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:4258e3c4faa2a2c18348","name":"Cosmos-Predict2.5-14B","organization":"Nvidia","date":"2025-09-29","domains":["Video"],"country":"United States of America","reference":"https://research.nvidia.com/publication/2025-09_world-simulation-video-foundation-models-physical-ai","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:93d8290871eea0671e7a","name":"Cosmos-Predict2.5 2B","organization":"Nvidia","date":"2025-09-29","domains":["Video"],"country":"United States of America","reference":"https://research.nvidia.com/publication/2025-09_world-simulation-video-foundation-models-physical-ai","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:cfb2217fc520033ecfd6","name":"DeepSeek-V3.2-Exp","organization":"DeepSeek","date":"2025-09-29","domains":["Language"],"country":"China","reference":"https://github.com/deepseek-ai/DeepSeek-V3.2-Exp/blob/main/DeepSeek_V3_2.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e4ca585901f1cab801fa","name":"MinerU2.5","organization":"Shanghai AI Lab,Peking University,Shanghai Jiao Tong University","date":"2025-09-29","domains":["Language","Vision","Multimodal"],"country":"China,China,China","reference":"https://arxiv.org/abs/2509.22186","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:52b7d773267038403ab0","name":"Wan 2.5","organization":"Alibaba","date":"2025-09-29","domains":["Video","Vision"],"country":"China","reference":"https://www.wan-ai.co/wan-2-5","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:710139501acae252c92d","name":"Seedream 4.0","organization":"ByteDance","date":"2025-09-28","domains":["Image generation"],"country":"China","reference":"https://arxiv.org/abs/2509.20427","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:79b79978a2cf896b97d4","name":"Kling 2.5 Turbo","organization":"Kuaishou Technology","date":"2025-09-26","domains":["Video","Vision"],"country":"China","reference":"https://ir.kuaishou.com/news-releases/news-release-details/kling-ai-launches-25-turbo-video-model-industry-leading","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:c53ec584707c48e99fb9","name":"Suno v5","organization":"Suno","date":"2025-09-25","domains":["Audio"],"country":"United States of America","reference":"https://help.suno.com/en/articles/8105153","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:4e6a3905b49b6607e241","name":"Gemini Robotics 1.5 ","organization":"Google DeepMind","date":"2025-09-25","domains":["Robotics","Vision","Language"],"country":"United States of America","reference":"https://deepmind.google/discover/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/\n\nhttps://storage.googleapis.com/deepmind-media/gemini-robotics/Gemini-Robotics-1-5-Tech-Report.pdf","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:ab69a7cafc8b7a3701c1","name":"Gemini Robotics-ER 1.5","organization":"Google DeepMind","date":"2025-09-25","domains":["Vision","Language","Speech"],"country":"United States of America","reference":"https://deepmind.google/discover/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/\n\nhttps://storage.googleapis.com/deepmind-media/gemini-robotics/Gemini-Robotics-1-5-Tech-Report.pdf","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:5b9dbd29b67a420d9cc5","name":"GigaEmbeddings","organization":"Sber,Moscow Institute of Physics and Technology","date":"2025-09-25","domains":["Language"],"country":"Russia,Russia","reference":"https://aclanthology.org/2025.bsnlp-1.3/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:63b3122d44f845c08b46","name":"SimpleFold","organization":"Apple","date":"2025-09-23","domains":["Biology"],"country":"United States of America","reference":"https://arxiv.org/abs/2509.18480v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e9c33511272b902137d2","name":"DeepSeek-V3.1-Terminus","organization":"DeepSeek","date":"2025-09-22","domains":["Language"],"country":"China","reference":"https://api-docs.deepseek.com/news/news250922","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9599cb54300ad3418fa9","name":"Qwen3-Omni-Flash","organization":"Alibaba","date":"2025-09-22","domains":["Multimodal","Language","Vision","Speech","Video"],"country":"China","reference":"https://arxiv.org/abs/2509.17765","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:167dd39626e060acda7c","name":"Qwen3-Omni-30B-A3B","organization":"Alibaba","date":"2025-09-22","domains":["Multimodal","Language","Vision","Speech","Video"],"country":"China","reference":"https://arxiv.org/abs/2509.17765","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:71e4adc0ae9e49433860","name":"Grok 4 Fast","organization":"xAI","date":"2025-09-19","domains":["Language"],"country":"United States of America","reference":"https://x.ai/news/grok-4-fast","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:50f4973a4fef719479ff","name":"Magistral Small 1.2","organization":"Mistral AI","date":"2025-09-18","domains":["Language","Vision","Multimodal"],"country":"France","reference":"https://huggingface.co/mistralai/Magistral-Small-2509","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d16965d11c862866a0f0","name":"Magistral Medium 1.2","organization":"Mistral AI","date":"2025-09-18","domains":["Language"],"country":"France","reference":"","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:a6b56a032c09298adfc7","name":"Granite-Docling","organization":"IBM","date":"2025-09-17","domains":["Vision","Multimodal","Language"],"country":"United States of America","reference":"https://www.ibm.com/new/announcements/granite-docling-end-to-end-document-conversion","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f1047f616e51b6111bfa","name":"AgentFounder-30B","organization":"Alibaba","date":"2025-09-16","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2509.13310","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7f0d2e897c938d7bcc5b","name":"Fabric 1.0","organization":"Veed","date":"2025-09-15","domains":["Video"],"country":"United Kingdom","reference":"https://www.veed.io/ai/fabric-1-0\n\nhttps://www.veed.io/blog/fabric-use-cases","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:12563e301b32727a96e6","name":"GPT-5-Codex","organization":"OpenAI","date":"2025-09-15","domains":["Language"],"country":"United States of America","reference":"https://openai.com/index/introducing-upgrades-to-codex/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:f7b24f31855253b648ff","name":"Qwen3-Next-80B-A3B","organization":"Alibaba","date":"2025-09-10","domains":["Language"],"country":"China","reference":"https://qwen.ai/blog?id=4074cca80393150c248e508aa62983f9cb7d27cd&from=research.latest-advancements-list","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:194d954d9fd3e5c5fea5","name":"Lucid Origin","organization":"Leonardo AI","date":"2025-09-10","domains":["Video"],"country":"Australia","reference":"https://intercom.help/leonardo-ai/en/articles/11890238-lucid-origin","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:bba229caf04b81395135","name":"Ling-mini-base-2.0-20T","organization":"Ant Group","date":"2025-09-10","domains":["Language"],"country":"China","reference":"https://huggingface.co/inclusionAI/Ling-mini-base-2.0-20T","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:21de0ab39830c91166d0","name":"Ling-flash-base-2.0-20T","organization":"Ant Group","date":"2025-09-10","domains":["Language"],"country":"China","reference":"https://huggingface.co/inclusionAI/Ling-flash-base-2.0","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d69d34f4b2915384ac4a","name":"K2 Think","organization":"Mohamed bin Zayed University of Artificial Intelligence (MBZUAI),G42","date":"2025-09-09","domains":["Language"],"country":"United Arab Emirates,United Arab Emirates","reference":"https://arxiv.org/abs/2509.07604","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:94b6981dc3ade2181c86","name":"Signal Processing Transformer","organization":"Softbank","date":"2025-09-09","domains":["Other"],"country":"Japan","reference":"https://arxiv.org/abs/2508.17960","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:b7a78e7c8b88087cc748","name":"Qwen3-Max","organization":"Alibaba","date":"2025-09-05","domains":["Language"],"country":"China","reference":"https://modelstudio.console.alibabacloud.com/?tab=doc#/doc/?type=model&url=2840914_2&modelId=qwen3-max-preview\n\nhttps://qwen.ai/blog?id=87dc93fc8a590dc718c77e1f6e84c07b474f6c5a&from=home.latest-research-list","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:cbd9223c0b4c9a5086cc","name":"EmbeddingGemma","organization":"Google DeepMind","date":"2025-09-05","domains":["Language"],"country":"United States of America","reference":"https://ollama.com/library/embeddinggemma","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:78219257607cac8ecf18","name":"Chatterbox Multilingual","organization":"Resemble AI","date":"2025-09-04","domains":["Speech"],"country":"United States of America","reference":"https://www.resemble.ai/introducing-chatterbox-multilingual-open-source-tts-for-23-languages/\n\nhttps://github.com/resemble-ai/chatterbox","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ed22cb6c71d9038e986e","name":"Apertus 8B","organization":"ETH Zurich,Ecole Polytechnique F\u00b4ed\u00b4erale de Lausanne (EPFL),Swiss National Supercomputing Centre (CSCS),Swisscom","date":"2025-09-02","domains":["Language"],"country":"Switzerland,Switzerland,Switzerland,Switzerland","reference":"https://ethz.ch/en/news-and-events/eth-news/news/2025/09/press-release-apertus-a-fully-open-transparent-multilingual-language-model.html\n\nhttps://github.com/swiss-ai/apertus-tech-report/blob/main/Apertus_Tech_Report.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b79ea51b073432776b2e","name":"Apertus 70B","organization":"ETH Zurich,Ecole Polytechnique F\u00b4ed\u00b4erale de Lausanne (EPFL),Swiss National Supercomputing Centre (CSCS),Swisscom","date":"2025-09-02","domains":["Language"],"country":"Switzerland,Switzerland,Switzerland,Switzerland","reference":"https://ethz.ch/en/news-and-events/eth-news/news/2025/09/press-release-apertus-a-fully-open-transparent-multilingual-language-model.html\n\nhttps://github.com/swiss-ai/apertus-tech-report/blob/main/Apertus_Tech_Report.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3ac59eaa9c1b367eb1e3","name":"LongCat-Flash","organization":"Meituan Inc","date":"2025-09-01","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2509.01322","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9472f7b460fc19cd577f","name":"MultiverSeg","organization":"Massachusetts Institute of Technology (MIT),Databricks","date":"2025-08-31","domains":["Vision","Biology","Medicine"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2412.15058","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e37646eebb9a53eddb48","name":"gpt-realtime","organization":"OpenAI","date":"2025-08-28","domains":["Speech","Vision","Language"],"country":"United States of America","reference":"https://openai.com/index/introducing-gpt-realtime/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:e3e39b1d5b318bb7f569","name":"MAI-Voice-1","organization":"Microsoft","date":"2025-08-28","domains":["Speech"],"country":"United States of America","reference":"https://microsoft.ai/news/two-new-in-house-models/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:6bc546912200a880516d","name":"Grok Code Fast 1","organization":"xAI","date":"2025-08-28","domains":["Language"],"country":"United States of America","reference":"https://data.x.ai/2025-08-26-grok-code-fast-1-model-card.pdf","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:a548a7536fd64102902f","name":"Wan 2.2 14B S2V","organization":"Alibaba","date":"2025-08-26","domains":["Video","Speech"],"country":"China","reference":"http://humanaigc.github.io/wan-s2v-webpage/content/wan-s2v.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d5949bf3ab89efc379e5","name":"Gemini 2.5 Flash Image (Nano Banana)","organization":"Google","date":"2025-08-26","domains":["Image generation"],"country":"United States of America","reference":"https://gemini.google/overview/image-generation/\n\nhttps://aistudio.google.com/models/gemini-2-5-flash-image","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:493e162017dc2ce3ed48","name":"YandexGPT 5.1 Pro","organization":"Yandex","date":"2025-08-25","domains":["Language"],"country":"Russia","reference":"https://ya.ru/ai/gpt","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:a9c9ca9a2c572228497d","name":"DeepSeek-V3.1","organization":"DeepSeek","date":"2025-08-21","domains":["Language"],"country":"China","reference":"https://api-docs.deepseek.com/news/news250821\n\nhttps://huggingface.co/deepseek-ai/DeepSeek-V3.1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c9010661ce113486201d","name":"Seed-OSS-36B-Base","organization":"ByteDance","date":"2025-08-21","domains":["Language"],"country":"China","reference":"https://seed.bytedance.com/en/blog/seed-oss-open-source-models-release","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:91dae98968ba4f30dde3","name":"Cohere Command A Reasoning","organization":"Cohere","date":"2025-08-21","domains":["Language"],"country":"Canada","reference":"https://cohere.com/blog/command-a-reasoning","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:9618fcf4c166f022b886","name":"Teuken 7B","organization":"OpenGPT-X,Fraunhofer Institute for Algorithms and Scientific Computing,Forschungszentrum Julich,Technische Universit\u00e4t Dresden","date":"2025-08-21","domains":["Language"],"country":"Germany,Germany,Germany,Germany","reference":"https://arxiv.org/abs/2410.03730","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:172cb9c19ab266274bb9","name":"Surya","organization":"NASA,University of Alabama,IBM Research","date":"2025-08-20","domains":["Earth science"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2508.14112","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a3a156d63aaaec19950f","name":"Gemma-SEA-LION-v4-27B-IT","organization":"AI Singapore","date":"2025-08-20","domains":["Multimodal","Language","Vision"],"country":"Singapore","reference":"https://sea-lion.ai/our-models/#sealionv4","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:66d8ce96cfbb58df4488","name":"FlowER","organization":"Massachusetts Institute of Technology (MIT)","date":"2025-08-20","domains":["Materials science"],"country":"United States of America","reference":"https://www.nature.com/articles/s41586-025-09426-9","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:631f430a021e19d56bed","name":"MolmoAct-7B-D","organization":"Allen Institute for AI,University of Washington","date":"2025-08-19","domains":["Robotics"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2508.07917","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a81159aaa0b7e0bf8bb2","name":"NVIDIA-Nemotron-Nano-9B-v2","organization":"Nvidia","date":"2025-08-18","domains":["Language"],"country":"United States of America","reference":"https://build.nvidia.com/nvidia/nvidia-nemotron-nano-9b-v2/modelcard","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:dd834757df0271fb6f82","name":"NVIDIA-Nemotron-Nano-12B-v2","organization":"Nvidia","date":"2025-08-18","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:2abfa80cd8afc6d4cb90","name":"Qwen Image Edit","organization":"Alibaba","date":"2025-08-18","domains":["Image generation"],"country":"China","reference":"https://qwen.ai/blog?id=a6f483777144685d33cd3d2af95136fcbeb57652&from=research.research-list","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:33b9484926673363a0c2","name":"Ovis2.5 9B","organization":"Alibaba","date":"2025-08-15","domains":["Multimodal","Language","Vision","Video"],"country":"China","reference":"https://arxiv.org/abs/2508.11737","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c621b08818663c7530ea","name":"Ovis2.5 2B","organization":"Alibaba","date":"2025-08-15","domains":["Multimodal","Language","Vision","Video"],"country":"China","reference":"https://arxiv.org/abs/2508.11737","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:dbc71839149acaebe641","name":"GLM-4.5V","organization":"Z.ai (Zhipu AI),Tsinghua University","date":"2025-08-15","domains":["Multimodal","Vision","Language","Video"],"country":"China,China","reference":"https://arxiv.org/abs/2507.01006","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3ab31778ebbad6d7ee6c","name":"GLM-4.1V-Thinking","organization":"Z.ai (Zhipu AI),Tsinghua University","date":"2025-08-15","domains":["Multimodal","Vision","Language"],"country":"China,China","reference":"https://arxiv.org/abs/2507.01006","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b109e51226f6f7e9777e","name":"imagen 4 fast","organization":"Google","date":"2025-08-15","domains":["Image generation"],"country":"United States of America","reference":"https://developers.googleblog.com/en/announcing-imagen-4-fast-and-imagen-4-family-generally-available-in-the-gemini-api/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:50eb697237ff8f90ccb9","name":"Canary 1B v2","organization":"Nvidia","date":"2025-08-14","domains":["Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2509.14128","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:53fc556c240d35af99db","name":"Parakeet-tdt-0.6b-v3","organization":"Nvidia","date":"2025-08-14","domains":["Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2509.14128","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6761af6bf1cef2ba3649","name":"Gemma 3 270M","organization":"Google DeepMind","date":"2025-08-14","domains":["Language"],"country":"United States of America","reference":"https://developers.googleblog.com/en/introducing-gemma-3-270m/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:1f50c97347e4922c9265","name":"Marey Realism v1.5","organization":"Moonvalley","date":"2025-08-14","domains":["Video"],"country":"Canada","reference":"https://blog.fal.ai/moonvalleys-marey-realism-v1-5-is-now-on-fal-the-most-advanced-generative-filmmaking-yet/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:092a1011beca69a8240d","name":"DINOv3","organization":"Meta AI","date":"2025-08-13","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2508.10104, https://ai.meta.com/dinov3/, https://github.com/facebookresearch/dinov3","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:eef1f1b73dabe9fd6469","name":"Mistral Medium 3.1","organization":"Mistral AI","date":"2025-08-12","domains":["Language"],"country":"France","reference":"https://x.com/MistralAI/status/1955316715417382979?lang=en","confidence":"Confident","accessibility":"API access"},{"id":"epoch:9dc79002375c3984662a","name":"GPT-5","organization":"OpenAI","date":"2025-08-07","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://cdn.openai.com/gpt-5-system-card.pdf\nhttps://openai.com/index/introducing-gpt-5/ ","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:ff39d774f49000397d5f","name":"GPT-5 mini","organization":"OpenAI","date":"2025-08-07","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://cdn.openai.com/gpt-5-system-card.pdf","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:8356ebab574ae173ecb5","name":"GPT-5 nano","organization":"OpenAI","date":"2025-08-07","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://cdn.openai.com/gpt-5-system-card.pdf","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:e730864299f81cd6f6bf","name":"Ideogram Character","organization":"Ideogram","date":"2025-08-07","domains":["Image generation"],"country":"Canada","reference":"https://ideogram.ai/features/character","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:255c68b31d531327bc38","name":"Claude Opus 4.1","organization":"Anthropic","date":"2025-08-05","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://www.anthropic.com/news/claude-opus-4-1","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:cb6481e6c1c0164f34fd","name":"gpt-oss-120b","organization":"OpenAI","date":"2025-08-05","domains":["Language"],"country":"United States of America","reference":"https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:73edcf2c54f4d6411f96","name":"gpt-oss-20b","organization":"OpenAI","date":"2025-08-05","domains":["Language"],"country":"United States of America","reference":"https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7627f119f99acd1610d5","name":"Genie 3","organization":"Google DeepMind","date":"2025-08-05","domains":["Games","Video","3D modeling"],"country":"United States of America","reference":"https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:77d80db11953a2a70586","name":"Qwen Image","organization":"Alibaba","date":"2025-08-04","domains":["Image generation","Vision"],"country":"China","reference":"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-Image/Qwen_Image.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e84540b76fe6639b6c78","name":"Hierarchical Reasoning Model (HPM)","organization":"Sapient Intelligence","date":"2025-08-04","domains":["Language","Vision","Multimodal"],"country":"Singapore","reference":"https://arxiv.org/abs/2506.21734","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:663a17fd68a310cfe8be","name":"Seed Diffusion Preview","organization":"ByteDance,Tsinghua University,SIA Lab Tsinghua AIR  ByteDance Seed","date":"2025-08-04","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/abs/2508.02193","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:e728095ecd4e8d89d146","name":"rStar-Math (Qwen2.5-Math-7B base)","organization":"Microsoft Research Asia,Peking University,Tsinghua University","date":"2025-08-01","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/pdf/2501.04519","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:fe7cd60c377e4f9ac553","name":"rStar-Math (Qwen2-Math-7B base)","organization":"Microsoft Research Asia,Peking University,Tsinghua University","date":"2025-08-01","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/abs/2501.04519","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:5b5aea282185703720a2","name":"MindLink-72B","organization":"Kunlun Inc.","date":"2025-08-01","domains":["Language"],"country":"China","reference":"https://github.com/SkyworkAI/MindLink/blob/main/mindlink.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0d4fe2d01b38b168c9ea","name":"MindLink-32B","organization":"Kunlun Inc.","date":"2025-08-01","domains":["Language"],"country":"China","reference":"https://github.com/SkyworkAI/MindLink/blob/main/mindlink.pdf","confidence":"Confident","accessibility":""},{"id":"epoch:fef896691b180f761519","name":"Gemini 2.5 Deep Think","organization":"Google,Google DeepMind","date":"2025-08-01","domains":["Language","Multimodal","Vision","Video","Audio","Mathematics"],"country":"United States of America,United States of America","reference":"https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-5-Deep-Think-Model-Card.pdf\n\nhttps://storage.googleapis.com/deepmind-media/gemini/gemini_v2_5_report.pdf","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:3f86ff76b79cdfe3a0c6","name":"Tri-21B","organization":"Trillion Labs","date":"2025-08-01","domains":["Language"],"country":"South Korea","reference":"https://huggingface.co/trillionlabs/Tri-21B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:ce799f4da2fafb57176e","name":"Veo 3 Fast","organization":"Google DeepMind","date":"2025-07-31","domains":["Video","Vision"],"country":"United States of America","reference":"https://developers.googleblog.com/en/veo-3-fast-image-to-video-capabilities-now-available-gemini-api/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:b9caee3821b5bca9c7f8","name":"Command A Vision","organization":"Cohere","date":"2025-07-31","domains":["Vision","Multimodal","Language"],"country":"Canada","reference":"https://cohere.com/blog/command-a-vision","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:7f71fed75d8e1a533f78","name":"AlphaEarth Foundations (AEF)","organization":"Google DeepMind,Google","date":"2025-07-30","domains":["Earth science"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2507.22291","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:3ac3df690ca7b84d945c","name":"Llama Nemotron Super v1.5","organization":"Nvidia","date":"2025-07-29","domains":["Language"],"country":"United States of America","reference":"https://developer.nvidia.com/blog/build-more-accurate-and-efficient-ai-agents-with-the-new-nvidia-llama-nemotron-super-v1-5/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:9397530c23042a7a1e58","name":"Qwen3-30B-A3B (Jul 2025)","organization":"Alibaba","date":"2025-07-29","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen3/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3f5556a01570b7c2101a","name":"GLM-4.5","organization":"Z.ai (Zhipu AI),Tsinghua University","date":"2025-07-28","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2508.06471","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5d19d4a0aea61ebe1e05","name":"Wan 2.2 14B T2V","organization":"Alibaba","date":"2025-07-28","domains":["Video"],"country":"China","reference":"https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:17e9da64ac3152545d37","name":"Wan 2.2 14B I2V","organization":"Alibaba","date":"2025-07-28","domains":["Video","Vision"],"country":"China","reference":"https://huggingface.co/Wan-AI/Wan2.2-I2V-A14B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:33a3c8e573f59cae5573","name":"GLM-4.5-Air","organization":"Z.ai (Zhipu AI),Tsinghua University","date":"2025-07-28","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2508.06471","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:86cfc26ae60788dd6fe1","name":"Agentar-Fin-R1 32B","organization":"Ant Group","date":"2025-07-27","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2507.16802","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:897db880ed75e877411d","name":"Agentar-Fin-R1 8B","organization":"Ant Group","date":"2025-07-27","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2507.16802","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:aa4abb71e847f49243d0","name":"Qwen3-235B-A22B-Thinking (Jul 2025)","organization":"Alibaba","date":"2025-07-25","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen3/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0a302eaaff7a5c0c4ea0","name":"Qwen3-235B-A22B (Jul 2025)","organization":"Alibaba","date":"2025-07-25","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen3/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fe82a5d3b9350c3ead35","name":"Magistral Small 1.1","organization":"Mistral AI","date":"2025-07-24","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/magistral","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:cbe87866827d723c1b5f","name":"Seed Prover","organization":"ByteDance","date":"2025-07-23","domains":["Language"],"country":"China","reference":"https://seed.bytedance.com/en/blog/bytedance-seed-prover-achieves-silver-medal-score-in-imo-2025","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:dc1fe5a0dae754e8866b","name":"Aeneas","organization":"Google DeepMind,University of Nottingham,University of Warwick,Athens University of Economics and Business,Google,University of Oxford","date":"2025-07-23","domains":["Vision","Multimodal","Language"],"country":"United States of America,United Kingdom,United Kingdom,Greece,United States of America,United Kingdom","reference":"https://www.nature.com/articles/s41586-025-09292-5","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:0f68009569bcdd5d651a","name":"Qwen3-Coder-480B-A35B","organization":"Alibaba","date":"2025-07-22","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen3-coder/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:18682002f407858f58c3","name":"GR-3","organization":"ByteDance","date":"2025-07-21","domains":["Robotics","Multimodal","Vision","Language"],"country":"China","reference":"https://arxiv.org/abs/2507.15493","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:6ea6cac0e4c2f8d301f6","name":"T-Pro 2.0","organization":"T-Bank","date":"2025-07-18","domains":["Language"],"country":"Russia","reference":"https://www.tbank.ru/about/news/18072025-t-technologies-group-has-released-model-with-hybrid-reasoning-mode-t-pro-2-0/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:dac30ea251e8b5f20b0f","name":"ChatGPT agent","organization":"OpenAI","date":"2025-07-17","domains":["Language"],"country":"United States of America","reference":"https://cdn.openai.com/pdf/6bcccca6-3b64-43cb-a66e-4647073142d7/chatgpt_agent_system_card_launch.pdf","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:e06547b1f4a6268525e1","name":"OpenReasoning-Nemotron-32B","organization":"Nvidia","date":"2025-07-16","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/nvidia/OpenReasoning-Nemotron-32B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:845698026b414428bec9","name":"EXAONE 4.0 (32B)","organization":"LG AI Research","date":"2025-07-15","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/abs/2507.11407","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:7dc656257a4e67e18f08","name":"EXAONE 4.0 (1.2B)","organization":"LG AI Research","date":"2025-07-15","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/abs/2507.11407","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:20d69cce4a48c14d7db9","name":"Voxtral Small","organization":"Mistral AI","date":"2025-07-15","domains":["Speech"],"country":"France","reference":"https://arxiv.org/abs/2507.13264","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:774d1a4ce51b256a28b6","name":"Voxtral Mini","organization":"Mistral AI","date":"2025-07-15","domains":["Speech","Language"],"country":"France","reference":"https://arxiv.org/abs/2507.13264","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:822b9ef347acb31056a0","name":"Gemini Embedding","organization":"Google DeepMind","date":"2025-07-14","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2503.07891","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:9d6edd63babf878faa6f","name":"Kimi K2","organization":"Moonshot","date":"2025-07-11","domains":["Language"],"country":"China","reference":"https://moonshotai.github.io/Kimi-K2/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:43045afe17a823763db4","name":"Grok 4 Heavy","organization":"xAI","date":"2025-07-10","domains":["Language"],"country":"United States of America","reference":"https://x.ai/news/grok-4","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:d5f1e9c16d3f096b439b","name":"Grok 4","organization":"xAI","date":"2025-07-09","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://x.ai/news/grok-4","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:8486a659eb702b2565ef","name":"EXAONE Path 2.0","organization":"LG AI Research","date":"2025-07-09","domains":["Vision","Medicine"],"country":"South Korea","reference":"https://arxiv.org/abs/2507.06639","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:efbf341f81c6d81227a1","name":"MedSigLIP","organization":"Google","date":"2025-07-09","domains":["Vision","Medicine"],"country":"United States of America","reference":"https://developers.google.com/health-ai-developer-foundations/medsiglip","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:af8ac44aba8ad40b119a","name":"dots.llm1","organization":"Rednote","date":"2025-07-06","domains":["Language"],"country":"China","reference":"https://www.arxiv.org/abs/2506.05767","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9a1b15246fe0e67a171a","name":"Finix-P1-32B","organization":"Ant Group","date":"2025-07-01","domains":[],"country":"China","reference":"https://rank.opencompass.org.cn/model-detail-v2/Finix-P1-32B%20(Thinking)","confidence":"Confident","accessibility":""},{"id":"epoch:df4abc174600019f86c5","name":"ERNIE-4.5-VL-28B-A3B","organization":"Baidu","date":"2025-06-29","domains":["Multimodal","Language","Vision","Video"],"country":"China","reference":"https://yiyan.baidu.com/blog/publication/ERNIE_Technical_Report.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a947cd51f1dc448192c9","name":"ERNIE-4.5-300B-A47B","organization":"Baidu","date":"2025-06-29","domains":["Language"],"country":"China","reference":"https://yiyan.baidu.com/blog/publication/ERNIE_Technical_Report.pdf","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5c00c2f166e2ed93bdea","name":"ERNIE-4.5-21B-A3B","organization":"Baidu","date":"2025-06-29","domains":["Language"],"country":"China","reference":"https://yiyan.baidu.com/blog/publication/ERNIE_Technical_Report.pdf","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:681ad7e718f8c0ff8449","name":"ERNIE-4.5-0.3B","organization":"Baidu","date":"2025-06-29","domains":["Language"],"country":"China","reference":"https://yiyan.baidu.com/blog/publication/ERNIE_Technical_Report.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9d1b4f2c8932c78bca0a","name":"FLUX.1 Kontext [dev]","organization":"Black Forest Labs","date":"2025-06-26","domains":["Image generation","Vision"],"country":"Germany","reference":"https://bfl.ai/announcements/flux-1-kontext","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:99faa010e98cdd431419","name":"DiLoCoX (Qwen1.5-107B on WT-103)","organization":"China Mobile,Zero Gravity Labs (0g AI)","date":"2025-06-26","domains":["Language"],"country":"China,United States of America","reference":"https://arxiv.org/abs/2506.21263","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f38f1b9bd0f25d57dd0e","name":"BlueOcean LLM 2.0 (\u8424\u77f3\u84dd\u6d77)","organization":"Hangzhou EZVIZ Software Co., Ltd. (Hikvision)","date":"2025-06-26","domains":["Multimodal","Vision","Video","Speech","Language"],"country":"China","reference":"https://web.archive.org/web/20250710111638/https://www.anfangnews.com/2025/0703/12905.shtml","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:72713e68a8cf72338b86","name":"AlphaGenome","organization":"Google DeepMind","date":"2025-06-25","domains":["Biology"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/papers/alphagenome.pdf","confidence":"Likely","accessibility":"API access"},{"id":"epoch:aa4b4aa79eb4fa2b30f5","name":"Kimi-VL","organization":"Moonshot","date":"2025-06-23","domains":["Multimodal","Vision","Language"],"country":"China","reference":"https://arxiv.org/pdf/2504.07491","confidence":"Confident","accessibility":""},{"id":"epoch:d2121124e45562f25719","name":"Minimax Hailuo 02","organization":"MiniMax,Hailuo AI","date":"2025-06-18","domains":["Video"],"country":"China,Singapore","reference":"https://www.minimax.io/news/minimax-hailuo-02","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:ce31fd8580122f99acea","name":"video-SALMONN 2+","organization":"ByteDance","date":"2025-06-18","domains":["Multimodal"],"country":"China","reference":"https://github.com/bytedance/video-SALMONN-2","confidence":"Likely","accessibility":""},{"id":"epoch:39d29de306d7c78a9e2e","name":"Mercury Coder Mini","organization":"Inception Labs","date":"2025-06-17","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2506.17298v1","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:2fb5cc68811ee33df81a","name":"Mercury Coder Small","organization":"Inception Labs","date":"2025-06-17","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2506.17298v1","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:007702d9d1a7fa3bdef6","name":"Kimi Dev 72b","organization":"Moonshot","date":"2025-06-16","domains":["Language"],"country":"China","reference":"https://moonshotai.github.io/Kimi-Dev/\n","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7f1341dbd7dc0c4fa89f","name":"Gemini 2.5 Flash-Lite","organization":"Google DeepMind","date":"2025-06-15","domains":["Language","Vision","Video","Speech","Multimodal"],"country":"United States of America","reference":"https://deepmind.google/models/gemini/flash-lite/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:1aa95aa1b5e3f30532da","name":"Mistral Small 3.2","organization":"Mistral AI","date":"2025-06-15","domains":["Language","Vision","Multimodal"],"country":"France","reference":"https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:bec13afd78de9d51ab87","name":"MiniMax-M1-80k","organization":"MiniMax","date":"2025-06-13","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2506.13585","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:961af46f4d6e6146681f","name":"MiniMax-M1-40k","organization":"MiniMax","date":"2025-06-13","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2506.13585","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1052689ece242983c0ac","name":"FGN","organization":"Google DeepMind","date":"2025-06-12","domains":["Earth science"],"country":"United States of America","reference":"https://arxiv.org/abs/2506.10772v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:eb1fd3dee2d72a4d7394","name":"CollabLLM","organization":"Stanford University,Microsoft,Georgia Institute of Technology","date":"2025-06-12","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2502.00640","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:c3c4e1eb3787069e2138","name":"V-JEPA 2","organization":"Facebook AI Research","date":"2025-06-11","domains":["Vision","Video","Robotics"],"country":"United States of America,France","reference":"https://arxiv.org/abs/2506.09985","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:009d61cd2b11fad55433","name":"Cosmos-Predict2-14B-Text2Image","organization":"Nvidia","date":"2025-06-11","domains":["Image generation"],"country":"United States of America","reference":"https://research.nvidia.com/labs/dir/cosmos-predict2/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:de3a49072721b71b31e8","name":"Cosmos-Predict2-2B-Video2World","organization":"Nvidia","date":"2025-06-11","domains":["Video","Vision","Robotics"],"country":"United States of America","reference":"https://research.nvidia.com/labs/dir/cosmos-predict2/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:bafd6b6a3cc00e109e7f","name":"Cosmos-Predict2-14B-Video2World","organization":"Nvidia","date":"2025-06-11","domains":["Video","Vision","Robotics"],"country":"United States of America","reference":"https://research.nvidia.com/labs/dir/cosmos-predict2/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:41141bda368d8b4505ba","name":"Cosmos-Predict2-2B-Text2Image","organization":"Nvidia","date":"2025-06-11","domains":["Image generation"],"country":"United States of America","reference":"https://huggingface.co/spaces/multimodalart/Cosmos-Predict2-2B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:d3345bb2e14256fdd645","name":"Seed-1.6-Thinking","organization":"ByteDance","date":"2025-06-11","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://seed.bytedance.com/en/blog/introduction-to-techniques-used-in-seed1-6","confidence":"Confident","accessibility":"API access"},{"id":"epoch:c688ce49b96ca1468705","name":"Seed 1.6","organization":"ByteDance","date":"2025-06-11","domains":["Multimodal","Language","Vision","Video"],"country":"China","reference":"https://www.aibase.com/news/18831","confidence":"Confident","accessibility":"API access"},{"id":"epoch:8fee70e035cda9c83aa3","name":"Magistral Medium 1.1","organization":"Mistral AI","date":"2025-06-10","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/magistral","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:44649e5248cfe27b0101","name":"o3-pro","organization":"OpenAI","date":"2025-06-10","domains":[],"country":"United States of America","reference":"","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:3f0f9af29e14ff270f05","name":"Seedance 1.0","organization":"ByteDance","date":"2025-06-10","domains":["Video"],"country":"China","reference":"https://arxiv.org/abs/2506.09113","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:b3e8d7aa52dcb525bcf3","name":"Skywork-R1V3","organization":"Kunlun Inc.","date":"2025-06-10","domains":["Vision"],"country":"China","reference":"https://arxiv.org/abs/2507.06167","confidence":"Confident","accessibility":""},{"id":"epoch:77c75aba54c853f88fa5","name":"Devstral Medium","organization":"Mistral AI,All Hands AI","date":"2025-06-10","domains":["Language"],"country":"France,United States of America","reference":"https://mistral.ai/news/devstral-2507","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:b4d3f23b41eb27b19694","name":"Magistral Small 1.0","organization":"Mistral AI","date":"2025-06-10","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/magistral","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5c7c24d6da0c05d24532","name":"MiniCPM-4-8B","organization":"OpenBMB (Open Lab for Big Model Base)","date":"2025-06-09","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2506.07900","confidence":"Confident","accessibility":""},{"id":"epoch:abc40c281e21cb4575f3","name":"Redwood AI","organization":"1X","date":"2025-06-09","domains":["Robotics"],"country":"United States of America","reference":"https://www.1x.tech/discover/redwood-ai","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:6f44d15bd7fd7983e32d","name":"Boltz-2","organization":"Massachusetts Institute of Technology (MIT),Recursion Pharmaceuticals,ETH Zurich,Valence Labs","date":"2025-06-06","domains":["Biology"],"country":"United States of America,United States of America,Switzerland,Canada","reference":"http://jeremywohlwend.com/assets/boltz2.pdf","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fba773cedcd7e4c17e8e","name":"MiMo-7B-Base","organization":"Xiaomi Corp","date":"2025-06-05","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2505.07608","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ed3c0f8924637842e582","name":"Qwen3 Embedding","organization":"Alibaba","date":"2025-06-05","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2506.05176v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9006bc210701949a5c96","name":"Qwen3 Reranker","organization":"Alibaba","date":"2025-06-05","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2506.05176v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:852d55f589acb2dedae0","name":"ether0","organization":"FutureHouse","date":"2025-06-05","domains":["Materials science","Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2506.17238","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f94cb38c2fd760b9d5fa","name":"Claude Gov","organization":"Anthropic","date":"2025-06-05","domains":["Language"],"country":"United States of America","reference":"https://www.anthropic.com/news/claude-gov-models-for-u-s-national-security-customers","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:39adc3f8823183f3b614","name":"Gemini 2.5 Pro (Jun 2025)","organization":"Google DeepMind","date":"2025-06-05","domains":["Language","Vision","Video","Multimodal","Speech"],"country":"United States of America","reference":"https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/#gemini-2-5-thinking","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:96dace77635bbd995a89","name":"Ink Whisper","organization":"Cartesia","date":"2025-06-04","domains":["Speech"],"country":"United States of America","reference":"https://cartesia.ai/ink\n\nhttps://cartesia.ai/blog/introducing-ink-speech-to-text\n\nhttps://docs.cartesia.ai/2024-11-13/build-with-cartesia/models/stt","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:f594100b20c8d4478f26","name":"MiMo-VL-7B-SFT","organization":"Xiaomi Corp","date":"2025-06-04","domains":["Vision","Multimodal","Language","Video"],"country":"China","reference":"https://arxiv.org/abs/2506.03569","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c0afa5429dbebcd3e6d0","name":"MiMo\u2011VL\u20117B\u2011RL","organization":"Xiaomi Corp","date":"2025-06-04","domains":["Vision","Multimodal","Language","Video"],"country":"China","reference":"https://arxiv.org/abs/2506.03569","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1b4aa336e057f1125e08","name":"Seed-Coder","organization":"ByteDance","date":"2025-06-04","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2506.03524","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9cfb7084afbe1c5cf2ba","name":"Eleven v3","organization":"ElevenLabs","date":"2025-06-03","domains":["Speech"],"country":"United States of America","reference":"https://elevenlabs.io/v3","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:3c4f72476838d9211fa6","name":"OpenAudio-S1","organization":"Fish Audio","date":"2025-06-03","domains":["Speech"],"country":"United States of America","reference":"https://openaudio.com/blogs/s1","confidence":"Likely","accessibility":"API access"},{"id":"epoch:1fa0afe084467171672a","name":"OpenAudio-S1-mini","organization":"Fish Audio","date":"2025-06-03","domains":["Speech"],"country":"United States of America","reference":"https://huggingface.co/fishaudio/openaudio-s1-mini","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:a416fe4088f6fee3f664","name":"Gemini 2.5 Flash Native Audio","organization":"Google DeepMind","date":"2025-06-03","domains":["Speech"],"country":"United States of America","reference":"https://blog.google/technology/google-deepmind/gemini-2-5-native-audio/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:2b8fdba0d7858a106fb3","name":"Kling 2.1","organization":"Kuaishou Technology","date":"2025-05-29","domains":["Video","Vision"],"country":"China","reference":"https://ir.kuaishou.com/news-releases/news-release-details/kling-ai-celebrates-first-anniversary-achieves-annualized","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:601a207c5788ad4e9947","name":"FLUX.1 Kontext [pro]","organization":"Black Forest Labs","date":"2025-05-29","domains":["Image generation","Vision"],"country":"Germany","reference":"https://arxiv.org/abs/2506.15742","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:f388f968f33bbd7d949a","name":"FLUX.1 Kontext [max]","organization":"Black Forest Labs","date":"2025-05-29","domains":["Image generation","Vision"],"country":"Germany","reference":"https://arxiv.org/abs/2506.15742","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:cc157ea5095f67c9fddf","name":"EVI 3","organization":"Hume","date":"2025-05-29","domains":["Speech"],"country":"United States of America","reference":"https://www.hume.ai/blog/introducing-evi-3","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:1906d8cfa1a4554eb3bd","name":"Skywork-OR1-32B","organization":"Kunlun Inc.","date":"2025-05-29","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2505.22312","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:cfcf2426d1dadfc33797","name":"Codestral Embed","organization":"Mistral AI","date":"2025-05-28","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/codestral-embed","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:ff8df82825c27d83321a","name":"Pangu Pro MoE","organization":"Huawei","date":"2025-05-28","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2505.21411","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:728b5b604bddad11da1e","name":"DeepSeek-R1 (May 2025)","organization":"DeepSeek","date":"2025-05-28","domains":["Language"],"country":"China","reference":"https://api-docs.deepseek.com/news/news250120","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f3de53c3a82fe907e2c5","name":"SignGemma","organization":"Google","date":"2025-05-27","domains":["Language"],"country":"United States of America","reference":"https://officechai.com/ai/google-announces-signgemma-a-model-that-translates-sign-language/","confidence":"Unknown","accessibility":""},{"id":"epoch:f7ed9809957fd01a97f5","name":"DMind-1","organization":"","date":"2025-05-24","domains":["Language"],"country":"","reference":"https://huggingface.co/DMindAI/DMind-1","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:681e558490fde1b7e00e","name":"OpenOmni","organization":"Chinese Academy of Sciences,Shenzhen Institute of Advanced Technology,University of Chinese Academy of Sciences,National University of Singapore,University of Science and Technology of China (USTC)","date":"2025-05-24","domains":["Multimodal","Language","Vision","Speech"],"country":"China,China,China,Singapore,China","reference":"https://arxiv.org/abs/2501.04561","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fe7bd3b5f6b3ffe4b4fe","name":"o3 Operator","organization":"OpenAI","date":"2025-05-23","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://cdn.openai.com/pdf/4375e605-f9a6-438d-bcc8-190599c183a6/o3_cua_system_card.pdf\n\nhttps://help.openai.com/en/articles/10561834-operator-release-notes","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:9a1efdadba4173fb8669","name":"DataRater test model (1B)","organization":"Google DeepMind","date":"2025-05-23","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2505.17895","confidence":"Likely","accessibility":""},{"id":"epoch:18e1ffd65c8adfb57a56","name":"Claude Opus 4","organization":"Anthropic","date":"2025-05-22","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://www.anthropic.com/claude/opus\n\nhttps://www.anthropic.com/news/claude-4","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:b7f1a90c78ca72f68b9a","name":"Claude Sonnet 4","organization":"Anthropic","date":"2025-05-22","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://www.anthropic.com/claude/sonnet\n\nhttps://www.anthropic.com/news/claude-4","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:6f1cf7069da696cba849","name":"Reason-ModernColBERT","organization":"LightOn","date":"2025-05-22","domains":["Language"],"country":"France","reference":"https://huggingface.co/lightonai/Reason-ModernColBERT","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:3ce6262b04bbe5b990fd","name":"Veo 3","organization":"Google DeepMind","date":"2025-05-21","domains":["Video","Vision"],"country":"United States of America","reference":"https://deepmind.google/models/veo/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:ec5a0dddc72d99ef3760","name":"Falcon-Arabic","organization":"Technology Innovation Institute","date":"2025-05-21","domains":["Language"],"country":"United Arab Emirates","reference":"https://huggingface.co/blog/tiiuae/falcon-arabic","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:38075516310fff3e4d98","name":"Falcon-H1","organization":"Technology Innovation Institute","date":"2025-05-21","domains":["Language"],"country":"United Arab Emirates","reference":"https://arxiv.org/abs/2507.22448","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e24d551e49ca89fbfdc1","name":"Devstral Small","organization":"Mistral AI,All Hands AI","date":"2025-05-21","domains":["Language"],"country":"France,United States of America","reference":"https://mistral.ai/news/devstral","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:95bda1a3b67b1d843bfd","name":"MedGemma 27B","organization":"Google","date":"2025-05-20","domains":["Medicine","Language","Vision","Multimodal"],"country":"United States of America","reference":"https://arxiv.org/abs/2507.05201","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:979bc9e691dcd5a6371c","name":"Gemma 3n","organization":"Google","date":"2025-05-20","domains":["Language","Multimodal","Speech","Vision"],"country":"United States of America","reference":"https://developers.googleblog.com/en/introducing-gemma-3n/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:d11e89e90d1f571e8fc9","name":"Lyria RealTime","organization":"Google DeepMind","date":"2025-05-20","domains":["Audio"],"country":"United States of America","reference":"https://deepmind.google/models/lyria/realtime/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:089e5c87a3a22d22f9ed","name":"Imagen 4","organization":"Google","date":"2025-05-20","domains":["Image generation"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/Model-Cards/Imagen-4-Model-Card.pdf","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:5c533fa2eb7578aa08d3","name":"voyage-3.5","organization":"Voyage AI","date":"2025-05-20","domains":["Language"],"country":"United States of America","reference":"https://blog.voyageai.com/2025/05/20/voyage-3-5/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:59427ee53a53826e9444","name":"Imagen 4 ultra","organization":"Google","date":"2025-05-20","domains":["Image generation"],"country":"United States of America","reference":"https://cloud.google.com/vertex-ai/generative-ai/docs/models/imagen/4-0-ultra-generate-001","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:1fd6286517d3f71c29c6","name":"Marin 8B","organization":"Marin","date":"2025-05-19","domains":["Language"],"country":"United States of America","reference":"https://marin.readthedocs.io/en/latest/reports/marin-8b-retro/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:17dcb099743ba1b576b5","name":"Cosmos-Reason1 7B","organization":"Nvidia","date":"2025-05-19","domains":["Multimodal","Language","Vision","Robotics","Video"],"country":"United States of America","reference":"https://arxiv.org/abs/2503.15558","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:256efebcad774541deae","name":"Cosmos-Reason1 56B","organization":"Nvidia","date":"2025-05-19","domains":["Multimodal","Language","Vision","Robotics","Video"],"country":"United States of America","reference":"https://arxiv.org/abs/2503.15558","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c4367aedb5d1a565545f","name":"NVIDIA Isaac GR00T N1.5 3B","organization":"Nvidia","date":"2025-05-18","domains":["Robotics","Vision","Language"],"country":"United States of America","reference":"https://nvidianews.nvidia.com/news/nvidia-powers-humanoid-robot-industry-with-cloud-to-robot-computing-platforms-for-physical-ai","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:02dd85d1a14073fff95c","name":"SANA 1.5 4.8B","organization":"Nvidia,Massachusetts Institute of Technology (MIT),Tsinghua University,Playground,Peking University,The University of Hong Kong","date":"2025-05-17","domains":["Image generation"],"country":"United States of America,United States of America,China,United States of America,China,Hong Kong","reference":"https://arxiv.org/abs/2501.18427","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:85c6b729ffed841e017e","name":"codex-1","organization":"OpenAI","date":"2025-05-16","domains":["Language"],"country":"United States of America","reference":"https://openai.com/index/introducing-codex/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:4966a7d546db99f89d27","name":"codex-mini","organization":"OpenAI","date":"2025-05-16","domains":["Language"],"country":"United States of America","reference":"https://openai.com/index/introducing-codex/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:9789c855af7a50a00cd6","name":"Repress","organization":"DeepGenomics","date":"2025-05-16","domains":["Biology"],"country":"Canada","reference":"https://www.biorxiv.org/content/10.1101/2025.05.15.654105v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:449c8aae6727eed596e9","name":"COLLECTIVE-1","organization":"","date":"2025-05-15","domains":["Language"],"country":"","reference":"https://www.vana.org/posts/vana-flower-labs-partnership","confidence":"Unknown","accessibility":""},{"id":"epoch:20298f0dd6ddea59a95d","name":"II-Medical-8B","organization":"Intelligent Internet","date":"2025-05-15","domains":["Language"],"country":"United Kingdom","reference":"https://ii.inc/web/blog/post/ii-medical","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:343f4b0396e37c23e024","name":"AlphaEvolve","organization":"DeepMind","date":"2025-05-14","domains":["Language"],"country":"United Kingdom","reference":"https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/","confidence":"Unknown","accessibility":""},{"id":"epoch:b1c8d53f262bbcd92ca9","name":"LTX-Video-0.9.7 13B distilled","organization":"Lightricks","date":"2025-05-14","domains":["Video"],"country":"Israel","reference":"https://huggingface.co/Lightricks/LTX-Video-0.9.7-distilled","confidence":"Unknown","accessibility":"Open weights (restricted use)"},{"id":"epoch:115c638bc80cd4da3ee6","name":"Pixverse v4.5","organization":"PixVerse AI","date":"2025-05-13","domains":["Video"],"country":"China","reference":"https://app.pixverse.ai/onboard\n\nhttps://www.veed.io/ai-models/video/pixverse-v4.5","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:240f4b71dd3dbba91087","name":"Hunyuan T1-Vision","organization":"Tencent","date":"2025-05-12","domains":["Video"],"country":"China","reference":"https://www.aibase.com/news/17980","confidence":"Unknown","accessibility":""},{"id":"epoch:f8a3ab09709c808429b6","name":"Minimax-Speech-02-HD","organization":"MiniMax","date":"2025-05-12","domains":["Speech"],"country":"China","reference":"https://arxiv.org/abs/2505.07916","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:41632e83bcad2e2d9534","name":"NTele-R1-32B-V1","organization":"ZTE","date":"2025-05-12","domains":["Mathematics","Language"],"country":"China","reference":"https://huggingface.co/ZTE-AIM/NTele-R1-32B-V1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1559549554cee1bb11a9","name":"Seed1.5-VL","organization":"ByteDance","date":"2025-05-11","domains":["Vision","Language","Multimodal","Video"],"country":"China","reference":"https://arxiv.org/abs/2505.07062","confidence":"Confident","accessibility":"API access"},{"id":"epoch:277f2a0ccf74d1cfa392","name":"Earth-2 (cBottle-SR)","organization":"Nvidia","date":"2025-05-10","domains":["Earth science","Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2505.06474v1","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:93444fcfb2befe7bba10","name":"Tianxi-32B","organization":"Lenovo","date":"2025-05-09","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250515004811/https://www.qbitai.com/2025/05/281131.html\nhttps://web.archive.org/web/20250710113516/https://zhidx.com/news/42238.html","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:8cb0d277ccceb2d13728","name":"Tianxi-72B","organization":"Lenovo","date":"2025-05-09","domains":["Language","Multimodal","Speech","Vision"],"country":"China","reference":"https://web.archive.org/web/20250515004811/https://www.qbitai.com/2025/05/281131.html\nhttps://web.archive.org/web/20250710113516/https://zhidx.com/news/42238.html","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:ccce9cc206007f876cb2","name":"Mistral Medium 3","organization":"Mistral AI","date":"2025-05-07","domains":["Multimodal","Language","Vision"],"country":"France","reference":"https://mistral.ai/news/mistral-medium-3","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:86da392f6a33e3a96b38","name":"Pangu Ultra MoE","organization":"Huawei","date":"2025-05-07","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2505.04519","confidence":"Speculative","accessibility":"Hosted access (no API)"},{"id":"epoch:10319c16702a051199cd","name":"Apriel Nemotron 15B","organization":"Nvidia,ServiceNow","date":"2025-05-06","domains":["Language"],"country":"United States of America,United States of America","reference":"https://www.servicenow.com/company/media/press-room/nvidia-enterprise-ai-agents.html","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:85cb23d1efb1115ae339","name":"Kevin-32B","organization":"Cognition,Stanford University","date":"2025-05-06","domains":["Language"],"country":"United States of America,United States of America","reference":"https://cognition.ai/blog/kevin-32b","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:947852db6cbd9f032d86","name":"Gemini 2.5 Pro (May 2025)","organization":"Google DeepMind","date":"2025-05-06","domains":["Language","Vision","Video","Multimodal","Speech"],"country":"United States of America","reference":"https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/#gemini-2-5-thinking","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:2b91950248b498239ea9","name":"Typhoon 2.1 Gemma 4B","organization":"Typhoon / SCB 10X","date":"2025-05-05","domains":["Language"],"country":"Thailand","reference":"https://opentyphoon.ai/blog/en/typhoon2-1-gemma-release","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:d03adb44eb834030afcc","name":"Typhoon 2.1 Gemma 12B","organization":"Typhoon / SCB 10X","date":"2025-05-05","domains":["Language"],"country":"Thailand","reference":"https://opentyphoon.ai/blog/en/typhoon2-1-gemma-release","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:824d1b3fdc72a4a73a68","name":"LTX-Video-0.9.7 13B","organization":"Lightricks","date":"2025-05-05","domains":["Video"],"country":"Israel","reference":"https://huggingface.co/Lightricks/LTX-Video-0.9.7-dev","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:f0c644a5913bb1dad606","name":"LinOSS","organization":"Massachusetts Institute of Technology (MIT)","date":"2025-05-02","domains":[],"country":"United States of America","reference":"https://news.mit.edu/2025/novel-ai-model-inspired-neural-dynamics-from-brain-0502","confidence":"Unknown","accessibility":""},{"id":"epoch:464666b3bb6a5d410f1d","name":"MiniMax-Speech-02-turbo","organization":"MiniMax","date":"2025-05-02","domains":["Speech"],"country":"China","reference":"https://www.minimaxi.com/news/minimax-speech-02","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:0cd069e5dd243aea2ffd","name":"Phi-4-Reasoning","organization":"Microsoft","date":"2025-04-30","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2504.21318","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b8caf7b9643ad7ce28f6","name":"Phi-4-Reasoning-plus","organization":"Microsoft","date":"2025-04-30","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2504.21318","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:319f4b6d9f7c355da3e1","name":"DeepSeek-Prover-V2-671B","organization":"DeepSeek","date":"2025-04-30","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2504.21801","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:5d638eaf00697a6a3794","name":"DeepSeek-Prover-V2-7B","organization":"DeepSeek","date":"2025-04-30","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2504.21801","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:6c59ace5b1f749715745","name":"TranscriptFormer","organization":"Chan Zuckerberg Initiative","date":"2025-04-30","domains":["Biology"],"country":"United States of America","reference":"https://chanzuckerberg.com/blog/transcriptformer-model-overview/","confidence":"Unknown","accessibility":""},{"id":"epoch:8ce14bc9ce6b1978dc96","name":"Amazon Nova Premier","organization":"Amazon","date":"2025-04-30","domains":["Multimodal","Language","Video","Vision"],"country":"United States of America","reference":"https://www.amazon.science/publications/amazon-nova-premier-technical-report-and-model-card","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:2460017028c8b422baa7","name":"GTE-ModernColBERT-v1","organization":"LightOn","date":"2025-04-30","domains":["Language"],"country":"France","reference":"https://huggingface.co/lightonai/GTE-ModernColBERT-v1","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:75f5a3b7b72e77603e1d","name":"Qwen3-235B-A22B","organization":"Alibaba","date":"2025-04-29","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen3/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fba54ae106dfd5a70ce0","name":"Qwen3-30B-A3B","organization":"Alibaba","date":"2025-04-29","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen3/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6e5a075da0f25059fa02","name":"Qwen3-32B","organization":"Alibaba","date":"2025-04-29","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen3/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:da9f3ecd73e073c1c6f6","name":"Qwen3-14B","organization":"Alibaba","date":"2025-04-29","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen3/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9b6b853ef9e9df019d76","name":"Qwen3-8B","organization":"Alibaba","date":"2025-04-29","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen3/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:54885e7ae9e3e59d211c","name":"Qwen3-4B","organization":"Alibaba","date":"2025-04-29","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen3/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3ab49504665065ad260f","name":"Qwen3-1.7B","organization":"Alibaba","date":"2025-04-29","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen3/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:13549f3d8308f054c74e","name":"Qwen3-0.6B","organization":"Alibaba","date":"2025-04-29","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen3/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c6bd651709a25b3c0904","name":"Foundation-sec-8b","organization":"Cisco","date":"2025-04-28","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2504.21039","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:607c12d47ae483b6ab15","name":"Palmyra X5","organization":"Writer","date":"2025-04-28","domains":["Language"],"country":"United States of America","reference":"https://writer.com/engineering/long-context-palmyra-x5/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:c29eaee96291920529da","name":"Atlantes","organization":"Allen Institute for AI","date":"2025-04-26","domains":[],"country":"United States of America","reference":"https://arxiv.org/abs/2504.19036","confidence":"Unknown","accessibility":""},{"id":"epoch:c1219fedb6a19bbd2857","name":"Pleias-RAG-350m","organization":"PleIAs","date":"2025-04-25","domains":["Language"],"country":"France","reference":"https://arxiv.org/abs/2504.18225","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1c55bc13a81788040c58","name":"Pleias-RAG-1B","organization":"PleIAs","date":"2025-04-25","domains":["Language"],"country":"France","reference":"https://arxiv.org/abs/2504.18225","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:06d08f9382a05a3ff00c","name":"HiDream-I1","organization":"HiDream","date":"2025-04-25","domains":["Image generation"],"country":"China","reference":"https://huggingface.co/HiDream-ai/HiDream-I1-Full","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0ac907c8247a5fcdc598","name":"Firefly Image 4","organization":"Adobe","date":"2025-04-24","domains":["Image generation"],"country":"United States of America","reference":"https://blog.adobe.com/en/publish/2025/04/24/adobe-firefly-next-evolution-creative-ai-is-here","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:eec563edd1bbf9070604","name":"Firefly Image 4 Ultra","organization":"Adobe","date":"2025-04-24","domains":["Image generation"],"country":"United States of America","reference":"https://blog.adobe.com/en/publish/2025/04/24/adobe-firefly-next-evolution-creative-ai-is-here","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:e47bc60ffb0c82385160","name":"gpt-image-1","organization":"OpenAI","date":"2025-04-23","domains":["Image generation","Vision"],"country":"United States of America","reference":"https://openai.com/index/image-generation-api/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:cd389f76a6f60940e4f9","name":"MamayLM","organization":"INSAIT,ETH Zurich","date":"2025-04-23","domains":["Language"],"country":"Bulgaria,Switzerland","reference":"https://huggingface.co/INSAIT-Institute/MamayLM-Gemma-2-9B-IT-v0.1","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:78467f8fcf3c244348ce","name":"\u03c00.5 (pi-0.5)","organization":"Physical Intelligence","date":"2025-04-22","domains":["Robotics","Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2504.16054","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:5b837a512c5e32ecaa0d","name":"Eagle 2.5","organization":"Nvidia,Nanjing University,Hong Kong Polytechnic University,Rutgers University","date":"2025-04-21","domains":["Vision","Robotics","Language"],"country":"United States of America,China,Hong Kong,United States of America","reference":"https://arxiv.org/abs/2504.15271\nhttps://nvlabs.github.io/EAGLE/\nhttps://github.com/NVlabs/EAGLE","confidence":"Confident","accessibility":""},{"id":"epoch:1bd1eecb204fcaa6c109","name":"SkyReels-V2","organization":"Kunlun Inc.","date":"2025-04-21","domains":["Video"],"country":"China","reference":"https://arxiv.org/abs/2504.13074","confidence":"Confident","accessibility":""},{"id":"epoch:eb1226825de424adbc37","name":"Llama-Primus-Nemotron-70B","organization":"Trend Micro","date":"2025-04-21","domains":["Language"],"country":"Japan,United States of America","reference":"https://huggingface.co/trend-cybertron/Llama-Primus-Nemotron-70B-Base","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:ac1bed91d2c8f9dffe1f","name":"Trillion-7B","organization":"Trillion Labs","date":"2025-04-21","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/abs/2504.15431","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5ee6e4f7c31ea1cce113","name":"Gemma 3 QAT 4B","organization":"Google DeepMind","date":"2025-04-18","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://developers.googleblog.com/en/gemma-3-quantized-aware-trained-state-of-the-art-ai-to-consumer-gpus/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:0bf72f477c6f945e20fd","name":"Gemma 3 QAT 1B","organization":"Google DeepMind","date":"2025-04-18","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://developers.googleblog.com/en/gemma-3-quantized-aware-trained-state-of-the-art-ai-to-consumer-gpus/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:455cf217660d1aa4207a","name":"Gemma 3 QAT 12B","organization":"Google DeepMind","date":"2025-04-18","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://developers.googleblog.com/en/gemma-3-quantized-aware-trained-state-of-the-art-ai-to-consumer-gpus/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:448fcfd8fa9ae2d677cf","name":"Gemma 3 QAT 27B","organization":"Google DeepMind","date":"2025-04-18","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://developers.googleblog.com/en/gemma-3-quantized-aware-trained-state-of-the-art-ai-to-consumer-gpus/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:334a3e913a9aeba5103e","name":"Gemini 2.5 Flash","organization":"Google DeepMind","date":"2025-04-17","domains":["Language","Multimodal","Vision","Speech","Video"],"country":"United States of America","reference":"https://deepmind.google/models/gemini/flash/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:0fd82c216348da635af4","name":"Demist-2","organization":"Darktrace","date":"2025-04-17","domains":["Language"],"country":"United Kingdom","reference":"https://www.darktrace.com/research/demist-2-darktrace-embedding-model-for-investigation-of-security-threats","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:7e30d6519cbea7a34373","name":"o4-mini","organization":"OpenAI","date":"2025-04-16","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://openai.com/index/introducing-o3-and-o4-mini/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:200af89cc582478b94df","name":"Seedream 3.0","organization":"ByteDance","date":"2025-04-16","domains":["Image generation"],"country":"China","reference":"https://arxiv.org/abs/2504.11346","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:163f2036ebd205f1c5cf","name":"Digest: Cyber AI Analyst","organization":"Darktrace","date":"2025-04-16","domains":["Language","Vision"],"country":"United Kingdom","reference":"https://www.darktrace.com/blog/ai-uncovered-introducing-darktrace-incident-graph-evaluation-for-security-threats-digest","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:283983eedf274ecdac11","name":"MAI-DS-R1","organization":"Microsoft","date":"2025-04-16","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/microsoft/MAI-DS-R1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1103127f04aaee63c2ec","name":"Nova-3","organization":"Deepgram","date":"2025-04-15","domains":["Speech"],"country":"United States of America","reference":"https://deepgram.com/learn/introducing-nova-3-speech-to-text-api","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:b453db63c0fec67a5b0d","name":"Kling 2.0 Video Generation","organization":"Kuaishou Technology","date":"2025-04-15","domains":["Video","Vision"],"country":"China","reference":"https://ir.kuaishou.com/news-releases/news-release-details/kling-ai-advances-20-era-empowering-everyone-tell-great-stories","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:aaa85b0a2ae9cd060562","name":"Kolors 2.0 Image Generation","organization":"Kuaishou Technology","date":"2025-04-15","domains":["Image generation"],"country":"China","reference":"https://ir.kuaishou.com/news-releases/news-release-details/kling-ai-advances-20-era-empowering-everyone-tell-great-stories","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:2907e7680fe044744bb5","name":"TerraMind","organization":"IBM,Forschungszentrum Julich,European Space Agency (ESA),NASA","date":"2025-04-15","domains":["Earth science","Vision"],"country":"United States of America,Germany,Multinational,United States of America","reference":"https://arxiv.org/abs/2504.11171","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b74d8cd4affeb8f64e16","name":"Cohere Embed 4","organization":"Cohere","date":"2025-04-15","domains":["Language","Vision","Multimodal"],"country":"Canada","reference":"https://cohere.com/blog/embed-4","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:ed2f4f44a0e61044f31b","name":"GPT-4.1","organization":"OpenAI","date":"2025-04-14","domains":["Multimodal","Language","Vision","Video"],"country":"United States of America","reference":"https://openai.com/index/gpt-4-1/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:6cc3272b360461f866b9","name":"GPT-4.1 mini","organization":"OpenAI","date":"2025-04-14","domains":["Multimodal","Language","Vision","Video"],"country":"United States of America","reference":"https://openai.com/index/gpt-4-1/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:97a26f4eef8c8d1b0e8d","name":"GPT-4.1 nano","organization":"OpenAI","date":"2025-04-14","domains":["Multimodal","Language","Vision","Video"],"country":"United States of America","reference":"https://openai.com/index/gpt-4-1/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:80556a95ce695a246bd3","name":"Nemotron-H 8B","organization":"Nvidia","date":"2025-04-14","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2504.03624","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:aed06a4e00ea3a359324","name":"Nemotron-H 47B","organization":"Nvidia","date":"2025-04-14","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2504.03624","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:19980ca14e437dfe8634","name":"Nemotron-H 56B","organization":"Nvidia","date":"2025-04-14","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2504.03624","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:d047ba114026ff4e0102","name":"DolphinGemma","organization":"Google DeepMind,Georgia Institute of Technology,Wild Dolphin Project","date":"2025-04-14","domains":["Audio"],"country":"United States of America,United States of America,United States of America","reference":"https://blog.google/technology/ai/dolphingemma/","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:36dba52f1c1f1ea7bed7","name":"GLM-Z1-Rumination-32B-0414","organization":"Tsinghua University","date":"2025-04-14","domains":["Language"],"country":"China","reference":"https://huggingface.co/THUDM/GLM-Z1-Rumination-32B-0414","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:763aeaec15e5360dacf2","name":"GLM-4-9B-0414","organization":"Tsinghua University","date":"2025-04-14","domains":["Language"],"country":"China","reference":"https://huggingface.co/THUDM/GLM-4-9B-0414","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d0949d91e59cadc80e89","name":"360Zhinao3-7B-O1.5","organization":"360 Security Technology","date":"2025-04-14","domains":["Language"],"country":"China","reference":"https://huggingface.co/qihoo360/360Zhinao3-7B-O1.5","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b659715b7f5b4c2f6ebc","name":"GLM-4-32B-0414","organization":"Z.ai (Zhipu AI),Tsinghua University","date":"2025-04-14","domains":["Language"],"country":"China,China","reference":"https://huggingface.co/zai-org/GLM-4-32B-0414","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f398f828e37354fc8626","name":"SenseNova V6","organization":"SenseTime","date":"2025-04-12","domains":["Multimodal","Language","Video","Vision"],"country":"Hong Kong","reference":"https://www.prnewswire.com/apac/news-releases/sensetimes-sensenova-v6-chinas-most-advanced-multimodal-model-with-the-lowest-cost-in-the-industry-302426998.html\nhttps://platform.sensenova.cn/technology/fusion/Pro/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:21cfdb9e3947d0cf21c1","name":"Seaweed-7B","organization":"ByteDance","date":"2025-04-11","domains":["Video"],"country":"China","reference":"https://arxiv.org/abs/2504.08685","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2aa56cf25aea2c4bb003","name":"Pangu Ultra","organization":"Huawei","date":"2025-04-10","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2504.07866","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:1769c24fb34c72a42763","name":"AMIE (Articulate Medical Intelligence Explorer)","organization":"Google DeepMind,Google Research","date":"2025-04-09","domains":["Medicine","Language"],"country":"United States of America,United States of America","reference":"https://www.nature.com/articles/s41586-025-08866-7?linkId=13898052\nhttps://www.nature.com/articles/s41586-025-08869-4?linkId=13898054","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:747850828a53ec755947","name":"SHIFT-SUV","organization":"Luminary Cloud,Nvidia,Honda","date":"2025-04-09","domains":["3D modeling"],"country":"United States of America,United States of America,Japan","reference":"https://www.luminarycloud.com/resources/luminary-cloud-unveils-first-physics-ai-open-source-automotive-foundation-model-for-suv-aerodynamics-in-collaboration-with-honda-and-nvidia/","confidence":"Unknown","accessibility":"Open weights (non-commercial)"},{"id":"epoch:fc300c6383dd59491363","name":"Gen-4 Turbo","organization":"Runway","date":"2025-04-09","domains":["Video","Vision"],"country":"United States of America","reference":"https://help.runwayml.com/hc/en-us/articles/37327109429011-Creating-with-Gen-4-Video","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:03576375080d0029d366","name":"QWQ-Plus","organization":"Alibaba","date":"2025-04-08","domains":["Language"],"country":"China","reference":"https://www.alibabacloud.com/help/en/model-studio/qwq","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:8ff0fb577fe719c90c2c","name":"TxGemma 27B","organization":"Google DeepMind,Google Research","date":"2025-04-08","domains":["Language","Biology"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2504.06196","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:2d2f4328e07b79f01457","name":"TxGemma 9B","organization":"Google DeepMind,Google Research","date":"2025-04-08","domains":["Language","Biology"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2504.06196","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:b679a53aed7d5c619a3f","name":"TxGemma 2B","organization":"Google DeepMind,Google Research","date":"2025-04-08","domains":["Language","Biology"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2504.06196","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:a0742b99fcce6ad60bb7","name":"Cogito 1 series","organization":"","date":"2025-04-08","domains":["Language"],"country":"","reference":"https://techcrunch.com/2025/04/08/deep-cogito-emerges-from-stealth-with-hybrid-ai-reasoning-models/","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3c96fedffe5126f54b92","name":"Amazon Nova Sonic","organization":"Amazon","date":"2025-04-08","domains":["Speech"],"country":"United States of America","reference":"https://aws.amazon.com/blogs/aws/introducing-amazon-nova-sonic-human-like-voice-conversations-for-generative-ai-applications/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:ab11d85204bb1d611c39","name":"T5Gemma (Gemma 9B-9B)","organization":"Google DeepMind","date":"2025-04-08","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2504.06225","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:b3c62fb4a05324a4b3b5","name":"Amazon Nova Reel","organization":"Amazon","date":"2025-04-07","domains":["Video","Vision"],"country":"United States of America","reference":"https://docs.aws.amazon.com/ai/responsible-ai/nova-reel/overview.html\n\nhttps://assets.amazon.science/96/7d/0d3e59514abf8fdcfafcdc574300/nova-tech-report-20250317-0810.pdf","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:3472cefebea381936e3b","name":"Llama 4 Scout","organization":"Meta AI","date":"2025-04-05","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://ai.meta.com/blog/llama-4-multimodal-intelligence/","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:66c02b081abde74e371c","name":"Llama 4 Maverick","organization":"Meta AI","date":"2025-04-05","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://ai.meta.com/blog/llama-4-multimodal-intelligence/","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:22f4f3edc6b38db24d62","name":"Llama 4 Behemoth (preview)","organization":"Meta AI","date":"2025-04-05","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://ai.meta.com/blog/llama-4-multimodal-intelligence/","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:d324e62fe36fce5aea15","name":"Sec-Gemini v1","organization":"Google","date":"2025-04-04","domains":[],"country":"United States of America","reference":"https://security.googleblog.com/2025/04/google-launches-sec-gemini-v1-new.html","confidence":"Unknown","accessibility":""},{"id":"epoch:894558f1a1eb84fbac37","name":"FoundationStereo","organization":"Nvidia","date":"2025-04-04","domains":["3D modeling"],"country":"United States of America","reference":"https://arxiv.org/abs/2501.09898","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:e653417d6feddf98a69a","name":"Midjourney V7","organization":"Midjourney","date":"2025-04-03","domains":["Vision"],"country":"United States of America","reference":"https://docs.midjourney.com/hc/en-us/articles/32199405667853-Version","confidence":"Unknown","accessibility":""},{"id":"epoch:59d65a351c3abe7da509","name":"SkyReels-A2","organization":"Kunlun Inc.","date":"2025-04-03","domains":["Video"],"country":"China","reference":"https://arxiv.org/pdf/2504.02436","confidence":"Confident","accessibility":""},{"id":"epoch:2410f19ba962dcb0323e","name":"OpenThaiGPT 1.6 /  OTG-1.6 (72B)","organization":"Mahidol University,AI Entrepreneurs Association of Thailand","date":"2025-04-02","domains":["Language"],"country":"Thailand,Thailand","reference":"https://arxiv.org/abs/2504.01789","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:be0d696a7d4e169f175e","name":"OpenThaiGPT R1 32b / OTG-R1 (32B)","organization":"Mahidol University,AI Entrepreneurs Association of Thailand","date":"2025-04-02","domains":["Language"],"country":"Thailand,Thailand","reference":"https://arxiv.org/abs/2504.01789","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:07f5b31ba9804e7e0255","name":"AutoGLM Rumination","organization":"Z.ai (Zhipu AI)","date":"2025-04-01","domains":[],"country":"China","reference":"https://www.techinasia.com/news/chinese-ai-firm-launches-model-claims-to-outperform-deepseek","confidence":"Unknown","accessibility":""},{"id":"epoch:8eed2a10dcb6144141fc","name":"Amazon Nova Act","organization":"Amazon","date":"2025-03-31","domains":["Other"],"country":"United States of America","reference":"https://labs.amazon.science/blog/nova-act","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:60f211bab646da52d7a4","name":"Gen-4","organization":"Runway","date":"2025-03-31","domains":["Video","Vision"],"country":"United States of America","reference":"https://runwayml.com/research/introducing-runway-gen-4\n\nhttps://help.runwayml.com/hc/en-us/articles/37327109429011-Creating-with-Gen-4-Video","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:76e17f5f3f18545e2668","name":"Papla P1","organization":"Papla Media","date":"2025-03-31","domains":["Speech"],"country":"Poland","reference":"https://papla.media/blog/p1","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:0d33a7682201582f7b2c","name":"DeepHermes 3 - Mistral 24B","organization":"Nous Research","date":"2025-03-29","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/NousResearch/DeepHermes-3-Mistral-24B-Preview","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:08faf2a0caa346341326","name":"QVQ-Max","organization":"Alibaba","date":"2025-03-28","domains":["Language","Vision","Multimodal"],"country":"China","reference":"https://qwenlm.github.io/blog/qvq-max-preview/","confidence":"Confident","accessibility":""},{"id":"epoch:cd6bd39d362913bf2756","name":"Lingju Lingnao","organization":"Guangzhou Lingju Information Technology Co Ltd.","date":"2025-03-28","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250624092437/https://m.sohu.com/a/881435995_121956424","confidence":"Unknown","accessibility":""},{"id":"epoch:12985b291a8d4610ee8f","name":"Lumina-Image-2.0","organization":"Shanghai AI Lab,University of Sydney,Chinese University of Hong Kong (CUHK),Shanghai Jiao Tong University,Krea AI","date":"2025-03-27","domains":["Image generation"],"country":"China,Australia,Hong Kong,China,United States of America","reference":"https://arxiv.org/abs/2503.21758","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:69e880ad6a95f55ae024","name":"CassetteAI","organization":"CassetteAI","date":"2025-03-27","domains":["Audio"],"country":"United States of America","reference":"https://cassetteai.com/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:7a6d12161b2688e445c0","name":"FASHN v1.5","organization":"FASHN AI","date":"2025-03-27","domains":["Vision","Image generation"],"country":"United Kingdom","reference":"https://fashn.ai/blog/fashn-v1-5-our-next-generation-virtual-try-on-model","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:3763f78bcf848e9c4109","name":"GPT-4o (Mar 2025)","organization":"OpenAI","date":"2025-03-27","domains":["Multimodal","Language","Audio","Speech","Vision"],"country":"United States of America","reference":"https://help.openai.com/en/articles/6825453-chatgpt-release-notes","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:2da27fd03744a3d73a8d","name":"GAIA-2","organization":"Wayve","date":"2025-03-26","domains":["Video","Vision","Multimodal","Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2503.20523v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e31ca1c1c49f269bd01e","name":"Ideogram 3.0","organization":"Ideogram","date":"2025-03-26","domains":["Image generation"],"country":"Canada","reference":"https://about.ideogram.ai/3.0","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:78141c3bdf39ab22eb5a","name":"Qwen2.5-Omni 7B","organization":"Alibaba","date":"2025-03-26","domains":["Multimodal","Language","Vision","Video","Audio","Speech"],"country":"China","reference":"https://arxiv.org/abs/2503.20215","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:222bb9ff86cf61301653","name":"Qwen2.5-Omni 3B","organization":"Alibaba","date":"2025-03-26","domains":["Multimodal","Language","Vision","Video","Audio","Speech"],"country":"China","reference":"https://arxiv.org/abs/2503.20215","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:bf449d8adc1a79db38be","name":"Gemini 2.5 Pro (Mar 2025)","organization":"Google DeepMind","date":"2025-03-25","domains":["Language","Vision","Video","Multimodal","Speech"],"country":"United States of America","reference":"https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/#gemini-2-5-thinking","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:2c7f08707aba75238da8","name":"4o Image Generation","organization":"OpenAI","date":"2025-03-25","domains":["Image generation"],"country":"United States of America","reference":"https://openai.com/index/introducing-4o-image-generation/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:1dedb86e6190ec39fa0b","name":"Stable Video 4D 2.0 (SV4D 2.0)","organization":"Stability AI,Northeastern University","date":"2025-03-25","domains":["Vision","Video","3D modeling"],"country":"United Kingdom,United States of America","reference":"https://arxiv.org/abs/2503.16396","confidence":"Unknown","accessibility":"Open weights (restricted use)"},{"id":"epoch:b1e7c747dc884facf533","name":"3DGUT","organization":"Nvidia,University of Toronto","date":"2025-03-24","domains":["3D modeling"],"country":"United States of America,Canada","reference":"https://arxiv.org/abs/2412.12507","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f8623733a00764fc9962","name":"DeepSeek-V3 (Mar 2025)","organization":"DeepSeek","date":"2025-03-24","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2412.19437","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:76ad829292ae8b545126","name":"Diffusion Renderer","organization":"Nvidia,University of Toronto,Vector Institute,University of Illinois Urbana-Champaign (UIUC)","date":"2025-03-22","domains":["Video"],"country":"United States of America,Canada,Canada,United States of America","reference":"https://arxiv.org/abs/2501.18590","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:c3682d9f25b822f46df6","name":"HiCEGNN","organization":"University of Missouri","date":"2025-03-22","domains":["Biology"],"country":"United States of America","reference":"https://academic.oup.com/nargab/article/7/1/lqaf027/8090331","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b5d6102c6663952edad1","name":"CodeScientist","organization":"Allen Institute for AI","date":"2025-03-20","domains":[],"country":"United States of America","reference":"https://arxiv.org/abs/2503.22708","confidence":"Unknown","accessibility":""},{"id":"epoch:f373b7a1a5de51dc071d","name":"Gezhi (\u683c\u81f4\u5927\u6a21\u578b)","organization":"Troy Information Technology Co., Ltd.","date":"2025-03-19","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250624072945/https://finance.sina.com.cn/stock/relnews/cn/2025-03-19/doc-ineqfcye5539068.shtml","confidence":"Unknown","accessibility":""},{"id":"epoch:b0539e38d5b06a668612","name":"o1-pro","organization":"OpenAI","date":"2025-03-19","domains":["Language","Mathematics","Multimodal"],"country":"United States of America","reference":"https://openai.com/index/introducing-chatgpt-pro/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:2361069e08ab48ef3682","name":"Llama Nemotron Nano 8B","organization":"Nvidia","date":"2025-03-18","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2505.00949","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:19e6eaa5b3abd90f2a43","name":"Llama Nemotron Ultra 253B","organization":"Nvidia","date":"2025-03-18","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2505.00949","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:8f35c80d0e4c1ccafc27","name":"Llama Nemotron Super 49B","organization":"Nvidia","date":"2025-03-18","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2505.00949","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:ac4b7bd68531200d1051","name":"GR00T N1 2B","organization":"Nvidia","date":"2025-03-18","domains":["Robotics","Vision","Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2503.14734v2\nhttps://arxiv.org/abs/2503.14734v1\nhttps://github.com/NVIDIA/Isaac-GR00T\nhttps://nvidianews.nvidia.com/news/nvidia-isaac-gr00t-n1-open-humanoid-robot-foundation-model-simulation-frameworks\nhttps://developer.nvidia.com/blog/accelerate-generalist-humanoid-robot-development-with-nvidia-isaac-gr00t-n1/\nhttps://huggingface.co/nvidia/GR00T-N1-2B\nhttps://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-Sim\n\n","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:962d16db993962670648","name":"Cosmos-Transfer1-7B","organization":"Nvidia","date":"2025-03-18","domains":["Video","Vision","Robotics"],"country":"United States of America","reference":"https://arxiv.org/abs/2503.14492","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:2bd1784d4add94a28a6e","name":"Mistral Small 3.1","organization":"Mistral AI","date":"2025-03-17","domains":["Language","Vision","Multimodal"],"country":"France","reference":"https://mistral.ai/news/mistral-small-3-1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8783759f0a18ce897199","name":"Chirp 3 Speech-to-Text","organization":"Google,Google DeepMind","date":"2025-03-17","domains":["Speech"],"country":"United States of America,United States of America","reference":"https://cloud.google.com/speech-to-text/v2/docs/chirp_3-model","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:3ae0a82acda098bdd659","name":"Chirp 3 HD Text-to-Speech","organization":"Google,Google DeepMind","date":"2025-03-17","domains":["Speech"],"country":"United States of America,United States of America","reference":"https://cloud.google.com/text-to-speech/docs/chirp3-hd","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:bfb836ba2629c5b9e7b2","name":"EXAONE Deep 32B","organization":"LG AI Research","date":"2025-03-16","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/abs/2503.12524","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:b5e13497ad41d7d0f728","name":"EXAONE Deep 7.8B","organization":"LG AI Research","date":"2025-03-16","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/abs/2503.12524","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:803e1b0f9aa959cad783","name":"EXAONE Deep 2.4B","organization":"LG AI Research","date":"2025-03-16","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/abs/2503.12524","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:b6148581dfa781693925","name":"ERNIE-4.5-VL-424B-A47B (\u6587\u5fc3\u5927\u6a21\u578b4.5)","organization":"Baidu","date":"2025-03-16","domains":["Multimodal","Language","Vision","Video"],"country":"China","reference":"https://www.datacamp.com/blog/ernie-4-5-x1\n\nhttps://yiyan.baidu.com/blog/publication/ERNIE_Technical_Report.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:08e34262cf11a5f7cf1e","name":"ERNIE x1 (\u6587\u5fc3\u5927\u6a21\u578bX1)","organization":"Baidu","date":"2025-03-16","domains":["Language","Multimodal"],"country":"China","reference":"https://www.datacamp.com/blog/ernie-4-5-x1","confidence":"Unknown","accessibility":""},{"id":"epoch:130f2bda6b54a607e4d2","name":"EXAONE 3.5-R 2.4B","organization":"LG AI Research","date":"2025-03-14","domains":["Language"],"country":"South Korea","reference":"","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f645ebeedc5d674e6ffc","name":"EXAONE 3.5-R 32B","organization":"LG AI Research","date":"2025-03-14","domains":["Language"],"country":"South Korea","reference":"","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c8fcd4e7f2ed9f266582","name":"EXAONE 3.5-R 7.8B","organization":"LG AI Research","date":"2025-03-14","domains":["Language"],"country":"South Korea","reference":"","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:ae095dee8cad2778800c","name":"Cohere Command A","organization":"Cohere","date":"2025-03-13","domains":["Language"],"country":"Canada","reference":"https://cohere.com/blog/command-a\nhttps://huggingface.co/CohereForAI/c4ai-command-a-03-2025?ref=cohere-ai.ghost.io","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:a9784ff9703668fbc9d9","name":"Meissonic","organization":"National University of Singapore,Skywork AI,Hong Kong University of Science and Technology (HKUST),University of California (UC) Berkeley,Zhejiang University (ZJU)","date":"2025-03-13","domains":["Image generation"],"country":"Singapore,Singapore,Hong Kong,United States of America,China","reference":"https://arxiv.org/abs/2410.08261","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:bb5a76fb8be0be20bcad","name":"OLMo 2 32B","organization":"Allen Institute for AI","date":"2025-03-13","domains":["Language"],"country":"United States of America","reference":"https://allenai.org/blog/olmo2-32B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b1de71d24f1c9e6adcdc","name":"GigaChat 2 MAX","organization":"Sber","date":"2025-03-13","domains":["Language"],"country":"Russia","reference":"https://www.sberbank.ru/ru/sberpress/all/article?newsID=50dce36e-befd-431d-9633-6e19ed7a1a76&blockID=1303&regionID=77&lang=ru&type=NEWS","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:a488ab4363a4eceaac00","name":"GigaChat 2 Lite","organization":"Sber","date":"2025-03-13","domains":["Language"],"country":"Russia","reference":"https://www.sberbank.ru/ru/sberpress/all/article?newsID=50dce36e-befd-431d-9633-6e19ed7a1a76&blockID=1303&regionID=77&lang=ru&type=NEWS","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:eb9b6acf844f042569d5","name":"GigaChat 2 Pro","organization":"Sber","date":"2025-03-13","domains":["Language"],"country":"Russia","reference":"https://www.sberbank.ru/ru/sberpress/all/article?newsID=50dce36e-befd-431d-9633-6e19ed7a1a76&blockID=1303&regionID=77&lang=ru&type=NEWS","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:21e4376fdd6db57cdc44","name":"Gemma 3 27B","organization":"Google DeepMind","date":"2025-03-12","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/gemma/Gemma3Report.pdf","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:e4b9a931658f8c645392","name":"Gemma 3 12B","organization":"Google DeepMind","date":"2025-03-12","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/gemma/Gemma3Report.pdf","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:605b4d866f5ce10cceed","name":"Gemma 3 4B","organization":"Google DeepMind","date":"2025-03-12","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/gemma/Gemma3Report.pdf","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:e13eaa75ecd355be938f","name":"Gemma 3 1B","organization":"Google DeepMind","date":"2025-03-12","domains":["Language"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/gemma/Gemma3Report.pdf","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:35ca0410b361e22e1be2","name":"Gemini Robotics","organization":"Google DeepMind","date":"2025-03-12","domains":["Robotics","Vision","Speech"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/gemini-robotics/gemini_robotics_report.pdf\nhttps://arxiv.org/abs/2503.20020","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:33daf1ce19682d0ff266","name":"Gemini Robotics-ER","organization":"Google DeepMind","date":"2025-03-12","domains":["Robotics","Vision","Speech"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/gemini-robotics/gemini_robotics_report.pdf\nhttps://arxiv.org/abs/2503.20020","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:85357e8fb6780bbe9310","name":"Marey","organization":"Moonvalley","date":"2025-03-12","domains":["Video"],"country":"Canada","reference":"https://venturebeat.com/ai/moonvalleys-marey-is-a-state-of-the-art-ai-video-model-trained-on-fully-licensed-data/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:8e07ab422fdf09c8936a","name":"Hunyuan-TurboS","organization":"Tencent","date":"2025-03-11","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250408105622/https://www.dapingtime.com/article/2171.html\n\nhttps://medium.com/data-science-in-your-pocket/tencent-hunyuan-turbo-s-the-fastest-reasoning-llm-d64a02bed5c8\n\nhttps://arxiv.org/abs/2505.15431","confidence":"Confident","accessibility":"API access"},{"id":"epoch:b3b27ac1cef413b167d2","name":"YuE","organization":"Hong Kong University of Science and Technology (HKUST),Multimodal Art Projection (MAP)","date":"2025-03-11","domains":["Speech"],"country":"Hong Kong,China","reference":"https://arxiv.org/abs/2503.08638","confidence":"Confident","accessibility":""},{"id":"epoch:a64c6f6037746ad2a159","name":"Reka Flash 3","organization":"Reka AI","date":"2025-03-10","domains":["Multimodal","Language","Vision","Video","Speech"],"country":"United States of America","reference":"https://www.reka.ai/news/introducing-reka-flash","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2d0ceb0662393f23305f","name":"Seedream 2.0","organization":"ByteDance","date":"2025-03-10","domains":["Image generation"],"country":"China","reference":"https://arxiv.org/abs/2503.07703","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:1de29593fc079d6f66af","name":"FoxBrain","organization":"Foxconn","date":"2025-03-10","domains":["Language"],"country":"Taiwan","reference":"https://huggingface.co/FoxconnAI/Llama_3.1-FoxBrain-70B\n\nhttps://aimagazine.com/articles/foxconn-creates-the-first-traditional-chinese-llm-foxbrain","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:126b10007cbf9d8028bb","name":"Ling-lite-1.5 (\"Bailing\")","organization":"Ant Group","date":"2025-03-10","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2503.05139","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c8a562d98eaef7b0ad86","name":"Ling-Plus (\"Bailing\")","organization":"Ant Group","date":"2025-03-10","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2503.05139","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4a598d2a709943d0778f","name":"Sonic 2","organization":"Cartesia","date":"2025-03-07","domains":["Speech"],"country":"United States of America","reference":"https://cartesia.ai/sonic\n\nhttps://docs.cartesia.ai/2024-11-13/build-with-cartesia/models/tts","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:933f332efa7ad07d147b","name":"QwQ-32B","organization":"Alibaba","date":"2025-03-06","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwq-32b/","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:04f49ed7e83e31f7fcf4","name":"Mistral OCR","organization":"Mistral AI","date":"2025-03-06","domains":["Multimodal","Vision","Language"],"country":"France","reference":"https://mistral.ai/news/mistral-ocr","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:f1a3d0dd311044e7c190","name":"Jamba 1.6 Mini","organization":"AI21 Labs","date":"2025-03-06","domains":["Language"],"country":"Israel","reference":"https://www.ai21.com/blog/introducing-jamba-1-6/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:ca54e85b3a9ce6544cc0","name":"Jamba 1.6 Large","organization":"AI21 Labs","date":"2025-03-06","domains":["Language"],"country":"Israel","reference":"https://www.ai21.com/blog/introducing-jamba-1-6/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:5d4e9be8f1f3033de699","name":"Character-3","organization":"Hedra AI","date":"2025-03-06","domains":["Video","Audio"],"country":"United States of America","reference":"https://www.linkedin.com/posts/hedra-labs_introducing-hedra-studio-and-character-3-activity-7303470158843379712-Gjxy","confidence":"Unknown","accessibility":"Open weights (non-commercial)"},{"id":"epoch:f0b383e96de82446b514","name":"LTX-Video-0.9.5 2B","organization":"Lightricks","date":"2025-03-05","domains":["Video"],"country":"Israel","reference":"https://huggingface.co/Lightricks/LTX-Video-0.9.5","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:c43180ded75eec4dd24c","name":"Phi-4 Mini","organization":"Microsoft","date":"2025-03-03","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2503.01743","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3f042fd97eccbefa2f5b","name":"Phi-4-Multimodal","organization":"Microsoft","date":"2025-03-03","domains":["Multimodal","Language","Vision","Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2503.01743","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:eff4ef18c2f6bdd58093","name":"Aya Vision 32B","organization":"Cohere","date":"2025-03-03","domains":["Multimodal","Vision","Language"],"country":"Canada","reference":"https://cohere.com/blog/aya-vision","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:afbef752245e446792b2","name":"Hailuo I2V-01-Director","organization":"MiniMax,Hailuo AI","date":"2025-03-03","domains":["Video","Vision"],"country":"China,Singapore","reference":"https://www.minimax.io/news/01-director","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:8de1a70a05dbd39fbe92","name":"Hailuo T2V-01-Director","organization":"MiniMax,Hailuo AI","date":"2025-03-03","domains":["Video"],"country":"China,Singapore","reference":"https://www.minimax.io/news/01-director","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:19d39c50be8694a76ede","name":"Spark-X1","organization":"iFlytek","date":"2025-03-03","domains":["Language","Mathematics"],"country":"China","reference":"https://en.tmtpost.com/news/7477906\nhttps://news.cgtn.com/news/2025-01-15/China-releases-Spark-X1-deep-reasoning-model-that-packs-a-punch-1AbIq8PzzEI/p.html\nhttps://wallstreetcn.com/articles/3740575","confidence":"Confident","accessibility":""},{"id":"epoch:36fa96ae09c7d303d9ff","name":"Difix3D+","organization":"Nvidia,National University of Singapore,University of Toronto,Vector Institute","date":"2025-03-03","domains":["3D modeling"],"country":"United States of America,Singapore,Canada,Canada","reference":"https://arxiv.org/abs/2503.01774","confidence":"Unknown","accessibility":"Open weights (non-commercial)"},{"id":"epoch:34f10c6f7bd6089d8db5","name":"Image-01","organization":"MiniMax","date":"2025-02-28","domains":["Image generation"],"country":"China","reference":"https://www.minimax.io/news/image-01","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:2d31eaa3a97805c153ba","name":"GPT-4.5","organization":"OpenAI","date":"2025-02-27","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://openai.com/index/introducing-gpt-4-5/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:eb28b4766e3e755465d9","name":"Kimi 1.6","organization":"Moonshot","date":"2025-02-27","domains":["Language"],"country":"China","reference":"","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:0bac0f0ced34bac3d752","name":"Mercury","organization":"Inception Labs","date":"2025-02-27","domains":["Language"],"country":"United States of America","reference":"https://machine-learning-made-simple.medium.com/is-the-mercury-llm-the-first-of-a-new-generation-of-llms-b64de1d36029\n\nhttps://www.inceptionlabs.ai/news\n\nhttps://www.inceptionlabs.ai/introducing-mercury-our-general-chat-model\n","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:6e85865f84c405c0bfa8","name":"Pika 2.2","organization":"Pika Labs","date":"2025-02-27","domains":["Video","Vision"],"country":"United States of America","reference":"https://www.linkedin.com/posts/pika-labs_launch-alert-pika-22-is-here-with-pikaframeskey-activity-7300923033542688770-BRfc?","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:b222914fde96d6b18d4b","name":"Granite 3.2 8B","organization":"IBM","date":"2025-02-26","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/ibm-granite/granite-3.2-8b-instruct","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8afca030fc8466a7c8d1","name":"Granite 3.2 2B","organization":"IBM","date":"2025-02-26","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/ibm-granite/granite-3.2-2b-instruct","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f6751f66e18d1f21761c","name":"Granite Guardian 3.2","organization":"IBM","date":"2025-02-26","domains":["Language"],"country":"United States of America","reference":"Unclear which model is being referred to","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a4a420046aa03cf384d0","name":"Euler (Euler\u5927\u6a21\u578b)","organization":"","date":"2025-02-26","domains":["Language"],"country":"","reference":"https://web.archive.org/web/20250624101313/https://finance.sina.com.cn/roll/2025-05-31/doc-ineykzhk8030958.shtml","confidence":"Confident","accessibility":""},{"id":"epoch:9e1f79d1f8e3cd237f54","name":"Eleven Scribe","organization":"ElevenLabs","date":"2025-02-26","domains":["Speech"],"country":"United States of America","reference":"https://elevenlabs.io/blog/meet-scribe","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:e8e869bed58f7626ef3f","name":"Wan 2.1 14B I2V","organization":"Alibaba","date":"2025-02-25","domains":["Video","Vision"],"country":"China","reference":"https://arxiv.org/abs/2503.20314\n","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fc8c0eb0a87644d4ac98","name":"Bailing-Pro-20250225","organization":"Ant Group","date":"2025-02-25","domains":["Language"],"country":"China","reference":"https://rank.opencompass.org.cn/leaderboard-llm/?m=25-04","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:70b848f1f66ddbf621c6","name":"Tianxi-7B","organization":"Lenovo","date":"2025-02-25","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250710112236/https://finance.sina.cn/tech/2025-02-25/detail-inemtimv4245680.d.html?fromtech=1&cid=38745&node_id=38745&oid=%E6%B7%98%E5%AE%9D%E6%99%AE%E6%8B%89%E8%BE%BE%E9%AB%98%E4%BB%BF%E5%8C%85%28%E5%BE%AE%E4%BF%A1198099199%293f5x&vt=4","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:2fc6d2f99e6cbd964fc4","name":"BFS-Prover","organization":"ByteDance","date":"2025-02-25","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2502.03438","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:72a195d8dff2706ebc4c","name":"YandexGPT 5 Pro","organization":"Yandex","date":"2025-02-25","domains":["Language"],"country":"Russia","reference":"https://ya.ru/ai/gpt\n\nhttps://habr.com/ru/companies/yandex/articles/885218/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:ab873f037b189d03b519","name":"YandexGPT 5 Lite","organization":"Yandex","date":"2025-02-25","domains":["Language"],"country":"Russia","reference":"https://ya.ru/ai/gpt\n\nhttps://habr.com/ru/companies/yandex/articles/885218/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:93d739383209b1447f97","name":"Claude 3.7 Sonnet","organization":"Anthropic","date":"2025-02-24","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://www.anthropic.com/news/claude-3-7-sonnet","confidence":"Likely","accessibility":"API access"},{"id":"epoch:12aae2b350eaf0d652b3","name":"Step-Video-T2V","organization":"StepFun","date":"2025-02-24","domains":["Video"],"country":"China","reference":"https://arxiv.org/abs/2502.10248","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5ecccc9455b283befe67","name":"Baigong (\u767e\u5de5)","organization":"Shanghai Lingyi Artificial Intelligence Technology Co., Ltd.","date":"2025-02-21","domains":["Robotics"],"country":"China","reference":"https://web.archive.org/web/20250624070848/https://m.jfdaily.com/wx/detail.do?id=728305","confidence":"Unknown","accessibility":""},{"id":"epoch:40f19aab3fb93be4835c","name":"Helix","organization":"Figure AI","date":"2025-02-20","domains":["Robotics"],"country":"United States of America","reference":"https://www.figure.ai/news/helix","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:573818d456e7a42a9f97","name":"SigLIP 2","organization":"Google DeepMind","date":"2025-02-20","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2502.14786","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2451cc602d6224bae47f","name":"Evo 2 40B","organization":"Arc Institute,Stanford University,Nvidia,Liquid,University of California (UC) Berkeley,Goodfire,Columbia University,University of California San Francisco","date":"2025-02-19","domains":["Biology"],"country":"United States of America,United States of America,United States of America,United States of America,United States of America,United States of America,United States of America","reference":"https://arcinstitute.org/manuscripts/Evo2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:767372a5526df34aba22","name":"Evo 2 7B","organization":"Arc Institute,Stanford University,Nvidia,Liquid,University of California (UC) Berkeley,Goodfire,Columbia University,University of California San Francisco","date":"2025-02-19","domains":["Biology"],"country":"United States of America,United States of America,United States of America,United States of America,United States of America,United States of America,United States of America","reference":"https://arcinstitute.org/manuscripts/Evo2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:26aad64e01e0ca0595cc","name":"Grok-3 mini","organization":"xAI","date":"2025-02-19","domains":["Language"],"country":"United States of America","reference":"https://x.ai/news/grok-3","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:09b1ac81f48df4cd7f28","name":"PaliGemma 2 3B Mix 224","organization":"Google","date":"2025-02-19","domains":["Vision","Multimodal","Language"],"country":"United States of America","reference":"https://developers.googleblog.com/en/introducing-paligemma-2-mix/\n\nhttps://arxiv.org/abs/2412.03555","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:0a604da3d83629a088b9","name":"PaliGemma 2 3B Mix 448","organization":"Google","date":"2025-02-19","domains":["Vision","Multimodal","Language"],"country":"United States of America","reference":"https://developers.googleblog.com/en/introducing-paligemma-2-mix/\n\nhttps://arxiv.org/abs/2412.03555","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:4bccda90fc0562a35773","name":"Qwen2.5-VL-72B","organization":"Alibaba","date":"2025-02-19","domains":["Language","Vision","Multimodal"],"country":"China","reference":"https://arxiv.org/abs/2502.13923","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:09b4b4b91bb6349b48e8","name":"Qwen2.5-VL-7B","organization":"Alibaba","date":"2025-02-19","domains":["Language","Vision","Multimodal"],"country":"China","reference":"https://arxiv.org/abs/2502.13923","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a6e3626a334a0025b4e7","name":"Qwen2.5-VL-3B","organization":"Alibaba","date":"2025-02-19","domains":["Language","Vision","Multimodal"],"country":"China","reference":"https://arxiv.org/abs/2502.13923","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:d0e6f264d799068c49b8","name":"R1 1776","organization":"Perplexity","date":"2025-02-18","domains":["Language"],"country":"United States of America","reference":"https://www.perplexity.ai/hub/blog/open-sourcing-r1-1776","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8f539149ed74697a6d2c","name":"TasiChat (TasiChat\u5927\u6a21\u578b)","organization":"Chengdu Tasi Technology Co., Ltd.","date":"2025-02-18","domains":["Multimodal","Language","Speech"],"country":"China","reference":"https://web.archive.org/web/20250624102052/https://www.stdaily.com/web/gdxw/2025-02/22/content_300290.html","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:0bb118ba49757d5d7929","name":"Brain2Qwerty","organization":"Meta AI,Universite de Technologie de Compi\u00e8gne \u2013 CNRS,Basque Center on Cognition","date":"2025-02-18","domains":["Language"],"country":"United States of America,France,Spain","reference":"https://ai.meta.com/research/publications/brain-to-text-decoding-a-non-invasive-approach-via-typing/","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:ea1cf395c2f390a0ad81","name":"Step-Audio-Chat 130B","organization":"StepFun","date":"2025-02-18","domains":["Speech"],"country":"China","reference":"https://arxiv.org/abs/2502.11946","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b772cb7586411f1de506","name":"Step-1","organization":"StepFun","date":"2025-02-18","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2502.11946","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4227d2ffa9e4651e03f4","name":"Step-Omni","organization":"StepFun","date":"2025-02-18","domains":["Speech","Language","Multimodal","Vision"],"country":"China","reference":"https://arxiv.org/abs/2502.11946","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:1a86e4bd8eb269cfea5e","name":"Grok 3","organization":"xAI","date":"2025-02-17","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://x.ai/blog/grok-3","confidence":"Likely","accessibility":"API access"},{"id":"epoch:dd1bbe6844d6d0ba348e","name":"Mistral Saba","organization":"Mistral AI","date":"2025-02-17","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/mistral-saba","confidence":"Confident","accessibility":"API access"},{"id":"epoch:c9468df8e6181059f3e0","name":"YAYI-Ultra","organization":"Yayi (Wenge)","date":"2025-02-15","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://web.archive.org/web/20250627090629/https://blog.csdn.net/qq_19841021/article/details/145658088","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:a3400f18c23782e46b4a","name":"SkyReels-A1","organization":"Kunlun Inc.","date":"2025-02-15","domains":["Video"],"country":"China","reference":"https://arxiv.org/abs/2502.10841","confidence":"Confident","accessibility":""},{"id":"epoch:d9714a20e22519273f16","name":"BRIA 3.1","organization":"BRIA AI","date":"2025-02-15","domains":["Image generation"],"country":"United States of America","reference":"https://huggingface.co/briaai/BRIA-3.1","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:d738d4fd6481d2b3886b","name":"LLaDA","organization":"Renmin University of China,Ant Group","date":"2025-02-14","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2502.09992","confidence":"Confident","accessibility":""},{"id":"epoch:64a95769bd37ebf999d3","name":"Deephermes 3 Llama 3 8B Preview","organization":"Nous Research","date":"2025-02-14","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/NousResearch/DeepHermes-3-Llama-3-8B-Preview","confidence":"Speculative","accessibility":"Open weights (restricted use)"},{"id":"epoch:74e7398adff56b86ad55","name":"Granite Vision 3.2 2B","organization":"IBM","date":"2025-02-14","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2502.09927","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9970514c9c23f6b796db","name":"Sonar Deep Research","organization":"Perplexity","date":"2025-02-14","domains":["Language"],"country":"United States of America","reference":"https://www.perplexity.ai/hub/blog/introducing-perplexity-deep-research","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:d24fb548778992549d02","name":"Step-Audio-TTS-3B","organization":"StepFun","date":"2025-02-14","domains":["Speech"],"country":"China","reference":"https://huggingface.co/stepfun-ai/Step-Audio-TTS-3B","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9b545e8e8428ab802bcc","name":"OmniHuman-1","organization":"ByteDance","date":"2025-02-13","domains":["Video","Vision","Audio"],"country":"China","reference":"https://arxiv.org/abs/2502.01061","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b9f28d76fb6f47c6107d","name":"OREAL 32B","organization":"Shanghai AI Lab,Shanghai Jiao Tong University,Chinese University of Hong Kong (CUHK),InnoHK","date":"2025-02-10","domains":["Mathematics","Language"],"country":"China,China,Hong Kong,Hong Kong","reference":"https://arxiv.org/abs/2502.06781","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4a87511ac17470fa0023","name":"OREAL 7B","organization":"Shanghai AI Lab,Shanghai Jiao Tong University,Chinese University of Hong Kong (CUHK),InnoHK","date":"2025-02-10","domains":["Mathematics","Language"],"country":"China,China,Hong Kong,Hong Kong","reference":"https://arxiv.org/abs/2502.06781","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:26338973beac88c0d673","name":"Animate Anyone 2","organization":"Alibaba","date":"2025-02-10","domains":["Video"],"country":"China","reference":"https://arxiv.org/abs/2502.06145","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:a9232b8e226b588fb4c0","name":"Zonos-v0.1","organization":"Zyphra","date":"2025-02-10","domains":["Speech"],"country":"United States of America","reference":"https://github.com/Zyphra/Zonos","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f812c402c5b80b3898ad","name":"HAMSTER VLM","organization":"Nvidia,University of Washington,University of Southern California","date":"2025-02-08","domains":["Robotics"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2502.05485\nhttps://hamster-robot.github.io/\nhttps://github.com/liyi14/HAMSTER_beta\nhttps://huggingface.co/yili18/Hamster_dev","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:29a97b10c054797cd752","name":"Goku-8B","organization":"The University of Hong Kong,ByteDance","date":"2025-02-07","domains":["Video","Vision"],"country":"Hong Kong,China","reference":"https://arxiv.org/abs/2502.04896v1","confidence":"Confident","accessibility":""},{"id":"epoch:d924c1250f2c9cb0cd65","name":"Goku+","organization":"The University of Hong Kong,ByteDance","date":"2025-02-07","domains":["Video","Vision"],"country":"Hong Kong,China","reference":"https://saiyan-world.github.io/goku/","confidence":"Confident","accessibility":""},{"id":"epoch:78b5b1f0064c59507660","name":"Gemini 2.0 Flash-Lite","organization":"Google DeepMind","date":"2025-02-05","domains":["Language","Vision","Video","Speech","Multimodal"],"country":"United States of America","reference":"https://deepmind.google/models/gemini/flash-lite/\n\nhttps://storage.googleapis.com/model-cards/documents/gemini-2-flash-lite.pdf","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:c7ebb3e788d0e95d7975","name":"Prithvi-EO-2.0 600M","organization":"IBM Research,NASA,University of Alabama,University of Iceland,Forschungszentrum Julich,Virginia Tech (Virginia Polytechnic Institute and State University),Arizona State University,Oregon State University,Boston University,University of California (UC) Berkeley,Julich Supercomputing Center","date":"2025-02-03","domains":["Earth science"],"country":"United States of America,United States of America,United States of America,Iceland,Germany,United States of America,United States of America,United States of America,United States of America,United States of America,Germany","reference":"https://arxiv.org/abs/2412.02732","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2f8d541a9c139715e4ff","name":"Prithvi-EO-2.0 300M","organization":"IBM Research,NASA,University of Alabama,University of Iceland,Forschungszentrum Julich,Virginia Tech (Virginia Polytechnic Institute and State University),Arizona State University,Oregon State University,Boston University,University of California (UC) Berkeley,Julich Supercomputing Center","date":"2025-02-03","domains":["Earth science"],"country":"United States of America,United States of America,United States of America,Iceland,Germany,United States of America,United States of America,United States of America,United States of America,United States of America,Germany","reference":"https://arxiv.org/abs/2412.02732","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a2a8817484c3f675a342","name":"Sonar Reasoning Pro","organization":"Perplexity","date":"2025-02-01","domains":["Language"],"country":"United States of America","reference":"https://docs.perplexity.ai/models/models/sonar-reasoning-pro","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:fef9254d693266063cfb","name":"o3-mini","organization":"OpenAI","date":"2025-01-31","domains":["Language"],"country":"United States of America","reference":"https://openai.com/index/openai-o3-mini/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:8af77cf34686a353ae2b","name":"s1-32B","organization":"Stanford University,University of Washington,Allen Institute for AI,Contextual AI","date":"2025-01-31","domains":["Language"],"country":"United States of America,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2501.19393","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:85970874180caad54ab0","name":"s1.1","organization":"Stanford University,University of Washington,Allen Institute for AI,Contextual AI","date":"2025-01-31","domains":["Language"],"country":"United States of America,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2501.19393","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:40b13cfadebc57cfbeb8","name":"Mistral Small 3","organization":"Mistral AI","date":"2025-01-30","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/mistral-small-3/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:69ae84a9bab231182d17","name":"Tulu 3 405B","organization":"Allen Institute for AI,University of Washington","date":"2025-01-30","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2411.15124\n\nhttps://allenai.org/blog/tulu-3-405B","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:1f480b104573341fcad6","name":"Sonar Reasoning","organization":"Perplexity","date":"2025-01-29","domains":["Language"],"country":"United States of America","reference":"https://docs.perplexity.ai/models/models/sonar-reasoning","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:a77c58b14b2b30730f85","name":"GPT-4o (Jan 2025)","organization":"OpenAI","date":"2025-01-29","domains":["Multimodal","Language","Audio","Speech","Vision"],"country":"United States of America","reference":"https://help.openai.com/en/articles/9624314-model-release-notes","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:242d41064724df039c16","name":"Qwen2.5-Max","organization":"Alibaba","date":"2025-01-28","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2.5-max/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:12c6527f3b11fc6f88f2","name":"Janus-Pro-7B","organization":"DeepSeek","date":"2025-01-27","domains":["Image generation","Vision","Language","Multimodal"],"country":"China","reference":"https://github.com/deepseek-ai/Janus/blob/main/janus_pro_tech_report.pdf","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:8e4d20ec331aa724f79b","name":"Janus-Pro-1B","organization":"DeepSeek","date":"2025-01-27","domains":["Image generation","Vision","Language","Multimodal"],"country":"China","reference":"https://github.com/deepseek-ai/Janus/blob/main/janus_pro_tech_report.pdf","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:07d2702791c144a41c09","name":"Kokoro v1.0","organization":"hexgrad","date":"2025-01-27","domains":["Speech"],"country":"","reference":"https://huggingface.co/spaces/hexgrad/Kokoro-TTS","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0851dc5448ece9fc2711","name":"Baichuan-Omni-1.5","organization":"Baichuan","date":"2025-01-26","domains":["Multimodal","Language","Speech","Vision","Video","Audio"],"country":"China","reference":"https://arxiv.org/abs/2501.15368","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:8fe03081781017a3fdba","name":"Sonar Pro","organization":"Perplexity","date":"2025-01-25","domains":["Language"],"country":"United States of America","reference":"https://www.perplexity.ai/hub/blog/introducing-the-sonar-pro-api\n\nhttps://docs.perplexity.ai/models/models/sonar-pro","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:98086f1a4408ccd01b79","name":"Computer-Using Agent (CUA)","organization":"OpenAI","date":"2025-01-23","domains":["Vision","Language","Multimodal"],"country":"United States of America","reference":"https://openai.com/index/computer-using-agent/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:37f3e31c01f0a1007ea9","name":"DoMINO","organization":"Nvidia","date":"2025-01-23","domains":["3D modeling"],"country":"United States of America","reference":"https://arxiv.org/abs/2501.13350","confidence":"Unknown","accessibility":""},{"id":"epoch:b488fb961876744d2804","name":"Luma Ray2","organization":"LumaLabs","date":"2025-01-23","domains":["Video"],"country":"United States of America","reference":"https://aws.amazon.com/bedrock/luma-ai/?nc1=h_ls\n\nhttps://lumalabs.ai/ray","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:e7aefc8b06a064e47152","name":"Doubao-1.5-pro","organization":"ByteDance","date":"2025-01-22","domains":["Language"],"country":"China","reference":"https://team.doubao.com/zh/special/doubao_1_5_pro","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:b40e7a109612c54d7c18","name":"Kimi k1.5","organization":"Moonshot","date":"2025-01-22","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://arxiv.org/abs/2501.12599v1","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:29920ddce720f70b65c1","name":"DeepSeek-R1-Distill-Llama-70B","organization":"DeepSeek","date":"2025-01-22","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2501.12948","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c47af3168e3c7732deed","name":"DeepSeek-R1-Distill-Qwen-14B","organization":"DeepSeek","date":"2025-01-22","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2501.12948","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3b72a0b74ea51b4d66aa","name":"DeepSeek-R1-Distill-Qwen-1.5B","organization":"DeepSeek","date":"2025-01-22","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2501.12948","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:da73a813dcc2dbf7ff5e","name":"DeepSeek-R1-Distill-Llama-8B","organization":"DeepSeek","date":"2025-01-22","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2501.12948","confidence":"Speculative","accessibility":"Open weights (restricted use)"},{"id":"epoch:56766c2a2a33e1c698b0","name":"DeepSeek-R1-Distill-Qwen-32B","organization":"DeepSeek","date":"2025-01-22","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2501.12948","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c2b4bdc4c2252d78624c","name":"DeepSeek-R1-Distill-Qwen-7B","organization":"DeepSeek","date":"2025-01-22","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2501.12948","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a053d5a764a5679a021e","name":"gte-modernbert","organization":"Alibaba","date":"2025-01-22","domains":["Language"],"country":"China","reference":"https://huggingface.co/Alibaba-NLP/gte-modernbert-base","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8aaf6aacffaf4bd0bdb8","name":"Hunyuan3D 2.0","organization":"Tencent","date":"2025-01-21","domains":["Language","Vision","3D modeling"],"country":"China","reference":"https://github.com/Tencent/Hunyuan3D-2/blob/main/assets/report/Tencent_Hunyuan3D_2_0.pdf","confidence":"Unknown","accessibility":"Open weights (restricted use)"},{"id":"epoch:cd5d7546f9d55b391cfe","name":"Sonar","organization":"Perplexity","date":"2025-01-21","domains":["Language"],"country":"United States of America","reference":"https://www.perplexity.ai/hub/blog/meet-new-sonar","confidence":"Confident","accessibility":"API access"},{"id":"epoch:9200b4f8baeda487e8c3","name":"DeepSeek-R1","organization":"DeepSeek","date":"2025-01-20","domains":["Language"],"country":"China","reference":"https://api-docs.deepseek.com/news/news250120","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b9cd8d357d6b7abc5f53","name":"DeepSeek-R1-Zero","organization":"DeepSeek","date":"2025-01-20","domains":["Language"],"country":"China","reference":"https://api-docs.deepseek.com/news/news250120","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:cd78e1cac8fe8b493ff3","name":"Eagle 2","organization":"Nvidia,Nanjing University,Tsinghua University,Hong Kong Polytechnic University,Johns Hopkins University,New York University (NYU)","date":"2025-01-20","domains":["Vision","Robotics","Language"],"country":"United States of America,China,China,Hong Kong,United States of America,United States of America","reference":"arxiv.org/abs/2501.14818","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:3738bdceed5cdd432951","name":"Zero-shot Monocular Scene Flow (ZeroMSF)","organization":"Nvidia,Brown University","date":"2025-01-20","domains":["3D modeling","Driving"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2501.10357","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:7743775c0cf6e5ff5387","name":"INTELLECT-MATH","organization":"Prime Intellect","date":"2025-01-17","domains":["Mathematics"],"country":"United States of America","reference":"https://huggingface.co/PrimeIntellect/INTELLECT-MATH\nhttps://www.primeintellect.ai/blog/intellect-math","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:92677212128564d49c00","name":"GPT-4b micro","organization":"OpenAI,Retro Biosciences","date":"2025-01-17","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://www.technologyreview.com/2025/01/17/1110086/openai-has-created-an-ai-model-for-longevity-science/","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:5c1bd5f43a5f02a27514","name":"MiniMax Speech-01-turbo (T2A-01-turbo)","organization":"MiniMax","date":"2025-01-17","domains":["Speech"],"country":"China","reference":"https://platform.minimaxi.com/document/T2A%20V2?key=66719005a427f0c8a5701643","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:f295ba43c85b3d71f45e","name":"MatterGen","organization":"Microsoft Research AI for Science","date":"2025-01-16","domains":["Materials science"],"country":"United States of America","reference":"https://www.nature.com/articles/s41586-025-08628-5","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:df5e56cd284bfadcf8e3","name":"MiniMax Speech-01-HD (T2A-01-HD)","organization":"MiniMax","date":"2025-01-15","domains":["Speech"],"country":"China","reference":"https://www.minimaxi.com/news/t2a-01-hd-%E5%8F%91%E5%B8%83\n\nhttps://platform.minimaxi.com/document/T2A%20V2?key=66719005a427f0c8a5701643","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:afea49ba74184814d63b","name":"Pika 2.1","organization":"Pika Labs","date":"2025-01-15","domains":["Video"],"country":"United States of America","reference":"https://www.linkedin.com/posts/pika-labs_pika-21-is-here-the-details-are-crystal-activity-7289689170954891265-CkzI?","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:04ff01570610955898b0","name":"Unichat-32B-c1","organization":"China Unicom","date":"2025-01-15","domains":["Language"],"country":"China","reference":"https://github.com/UnicomAI/Unichat-32B-c1?tab=readme-ov-file#unichat-qwen25-32b-c1","confidence":"Confident","accessibility":""},{"id":"epoch:a247524b734310832af2","name":"InternLM3","organization":"Shanghai AI Lab","date":"2025-01-15","domains":["Language"],"country":"China","reference":"https://huggingface.co/internlm/internlm3-8b-instruct","confidence":"Confident","accessibility":""},{"id":"epoch:12fb606c2c2f3bced1e5","name":"s1","organization":"Stanford University,University of Washington,Allen Institute for AI,Contextual AI","date":"2025-01-14","domains":["Language"],"country":"United States of America,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2501.19393","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fd642dedcce7a47f50a5","name":"MiniMax-Text-01","organization":"MiniMax","date":"2025-01-14","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2501.08313","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:4ed76505d2e6eb6cfe15","name":"MiniMax-VL-01","organization":"MiniMax","date":"2025-01-14","domains":["Vision","Language","Multimodal"],"country":"China","reference":"https://arxiv.org/abs/2501.08313","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:81192e958ec5ce28fafb","name":"OpenBioLLM-Llama3-70B","organization":"Saama","date":"2025-01-14","domains":["Language","Medicine"],"country":"United States of America","reference":"https://www.saama.com/openbiollm-llama3-saama-medical-llms/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:eee88059bb62545c3113","name":"OpenBioLLM-Llama3-8B","organization":"Saama","date":"2025-01-14","domains":["Language","Medicine"],"country":"United States of America","reference":"https://www.saama.com/openbiollm-llama3-saama-medical-llms/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:c2f2f4f18dcc35ff765f","name":"SenseNova Unified Large Model","organization":"SenseTime","date":"2025-01-13","domains":["Language"],"country":"Hong Kong","reference":"https://www.sensetime.com/en/news-detail/51169317?categoryId=1072","confidence":"Unknown","accessibility":""},{"id":"epoch:7c98c1fe626ae8eb1a00","name":"Stable Point Aware 3D (SPAR3D)","organization":"Stability AI,University of Illinois Urbana-Champaign (UIUC)","date":"2025-01-08","domains":["3D modeling"],"country":"United Kingdom,United States of America","reference":"https://arxiv.org/abs/2501.04689","confidence":"Unknown","accessibility":"Open weights (restricted use)"},{"id":"epoch:d6e2e46e294669fb3369","name":"Cosmos-1.0-\nDiffusion-14B Video2World","organization":"Nvidia","date":"2025-01-07","domains":["Robotics","Vision","Video"],"country":"United States of America","reference":"https://arxiv.org/abs/2501.03575","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:6d69d1bb3e8d28f7675f","name":"Cosmos-Predict1-7b-Video2World","organization":"Nvidia","date":"2025-01-07","domains":["Video","Vision","Robotics"],"country":"United States of America","reference":"https://arxiv.org/abs/2501.03575","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:0179d1a80373f4dce570","name":"voyage-3-large","organization":"Voyage AI","date":"2025-01-07","domains":["Language"],"country":"United States of America","reference":"https://blog.voyageai.com/2025/01/07/voyage-3-large/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:07b99d4cbc42bb6899ce","name":"Cosmos-Predict1-14b-Video2World","organization":"Nvidia","date":"2025-01-07","domains":["Video","Vision","Robotics"],"country":"United States of America","reference":"https://arxiv.org/abs/2501.03575","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:b6b26009d5d54312d123","name":"Tiangong 4.0","organization":"Kunlun Inc.","date":"2025-01-06","domains":["Language"],"country":"China","reference":"https://www.aibase.com/news/14480","confidence":"Unknown","accessibility":""},{"id":"epoch:dc17df2ce9503c6beec8","name":"Eurus-2-7B-PRIME","organization":"Tsinghua University,University of Illinois Urbana-Champaign (UIUC),Shanghai AI Lab,Peking University,Shanghai Jiao Tong University,CUHK Shenzhen Research Institute","date":"2025-01-02","domains":["Mathematics","Language"],"country":"China,United States of America,China,China,China,China","reference":"https://arxiv.org/abs/2502.01456","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:dfac7220bc94dd6d4f6e","name":"OLMo 2 Furious 7B","organization":"Allen Institute for AI,University of Washington,New York University (NYU)","date":"2024-12-31","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2501.00656","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:14eb15efc820c9c710ff","name":"OLMo 2 Furious 13B","organization":"Allen Institute for AI,University of Washington,New York University (NYU)","date":"2024-12-31","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2501.00656","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:cf0dddf2a4de37767528","name":"STORM-B/8","organization":"University of Southern California,Georgia Institute of Technology,Stanford University,Nvidia","date":"2024-12-31","domains":["3D modeling"],"country":"United States of America,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2501.00602\nhttps://github.com/NVlabs/GaussianSTORM\nhttps://jiawei-yang.github.io/STORM","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:dbb78885ebf8caa79f18","name":"Zhaoyan (\u5146\u8a00\u5927\u8bed\u8a00\u6a21\u578b)","organization":"Shanghai Jiao Tong University","date":"2024-12-30","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250624123213/https://cq.cri.cn/2024-03-08/890810f4-7bb4-4450-fba5-a1aaebb44858.html","confidence":"Unknown","accessibility":""},{"id":"epoch:95144373ba5740851973","name":"LTX-Video-0.9.1 2B","organization":"Lightricks","date":"2024-12-30","domains":["Video"],"country":"Israel","reference":"https://arxiv.org/abs/2501.00103","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:4493cf1854ba081ed2c0","name":"HiDream Foundation Model 1.0","organization":"HiDream","date":"2024-12-28","domains":["Video","Image generation","Vision","3D modeling"],"country":"China","reference":"https://www.chinaitcapital.com/detail/126","confidence":"Likely","accessibility":"API access"},{"id":"epoch:f6cd6345b4711fdf31c6","name":"Zhixiang 3.0 (\u667a\u8c61)","organization":"HiDream","date":"2024-12-28","domains":["Video","Image generation"],"country":"China","reference":"https://web.archive.org/web/20250624080243/https://news.sciencenet.cn/htmlnews/2024/12/536672.shtm\nhttps://web.archive.org/web/20250624080711/https://finance.sina.com.cn/roll/2025-04-28/doc-ineustca2189779.shtml","confidence":"Confident","accessibility":"API access"},{"id":"epoch:9fa8b11d05234fd10e5e","name":"LinGan VL (\u4e34\u611fVL)","organization":"Beijing 58 Information Technology","date":"2024-12-27","domains":["Multimodal"],"country":"China","reference":"https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm","confidence":"Unknown","accessibility":""},{"id":"epoch:a824b4f57dd94272bf5a","name":"Xiaomi Pengpai Image (\u5c0f\u7c73\u6f8e\u6e43\u56fe\u50cf)","organization":"Xiaomi Corp","date":"2024-12-27","domains":["Image generation"],"country":"China","reference":"https://web.archive.org/web/20250708152732/https://www.cnfin.com/gs-lb/detail/20241227/4165720_1.html","confidence":"Unknown","accessibility":""},{"id":"epoch:41ea8a2e12ca42327d44","name":"Xiaomi Edge-side Text (\u5c0f\u7c73\u7aef\u4fa7\u6587\u672c)","organization":"Xiaomi Corp","date":"2024-12-27","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250708152732/https://www.cnfin.com/gs-lb/detail/20241227/4165720_1.html","confidence":"Unknown","accessibility":""},{"id":"epoch:271f0a6b06c8bf1a7921","name":"QVQ","organization":"Alibaba","date":"2024-12-25","domains":["Language","Vision","Multimodal"],"country":"China","reference":"https://qwenlm.github.io/blog/qvq-72b-preview/\nhttps://arxiv.org/abs/2502.13923","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:4ace966016470a7bd448","name":"Jueqing LLM (\u89c9\u537f\u5927\u6a21\u578b)","organization":"Suzhou Jueqing Diyu Intelligent Technology","date":"2024-12-25","domains":["Language","Psychology"],"country":"China","reference":"https://web.archive.org/web/20250624095518/https://www.ksrmtzx.com/news/detail/239889","confidence":"Unknown","accessibility":""},{"id":"epoch:1b625c076569ff710e71","name":"Kokoro v0.19","organization":"hexgrad","date":"2024-12-25","domains":["Speech"],"country":"","reference":"https://huggingface.co/spaces/hexgrad/Kokoro-TTS","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b1274c4a792a06e35d9e","name":"DeepSeek-V3","organization":"DeepSeek","date":"2024-12-24","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2412.19437","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:77c890306155810a96f2","name":"OCTAVE 8B","organization":"Hume","date":"2024-12-23","domains":["Speech"],"country":"United States of America","reference":"https://www.hume.ai/blog/introducing-octave","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:18b6be7a23c5df3f7489","name":"OCTAVE 3B","organization":"Hume","date":"2024-12-23","domains":["Speech"],"country":"United States of America","reference":"https://www.hume.ai/blog/introducing-octave","confidence":"Confident","accessibility":"API access"},{"id":"epoch:47fff2b76dcf35b9b978","name":"RMBG v2.0","organization":"BRIA AI","date":"2024-12-23","domains":["Image generation","Vision"],"country":"United States of America","reference":"https://bria.ai/rmbg-for-commercial-use","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:b51b40d3a8fc700c5585","name":"VRMBG 2.0","organization":"BRIA AI","date":"2024-12-23","domains":["Video"],"country":"United States of America","reference":"https://bria.ai/rmbg-for-commercial-use#video","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:954c5928e4c2d3453850","name":"o3","organization":"OpenAI","date":"2024-12-20","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://openai.com/index/introducing-o3-and-o4-mini/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:092b3b5067a83d550966","name":"Gemini 2.0 Flash Thinking","organization":"Google DeepMind,Google","date":"2024-12-19","domains":["Language","Vision","Multimodal"],"country":"United States of America,United States of America","reference":"https://deepmind.google/technologies/gemini/flash-thinking/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:85099029b16de656dc62","name":"SEA-LION V3 Gemma2 9B","organization":"AI Singapore","date":"2024-12-19","domains":["Language"],"country":"Singapore","reference":"https://huggingface.co/aisingapore/gemma2-9b-cpt-sea-lionv3-base","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:0a5c3738bc8b5ecbb05b","name":"SEA-LION V3 Llama3.1 8B","organization":"AI Singapore","date":"2024-12-19","domains":["Language"],"country":"Singapore","reference":"https://huggingface.co/aisingapore/llama3.1-8b-cpt-sea-lionv3-base","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:e21fcef0b522b27af382","name":"SEA-LION V3 Llama3.1 70B","organization":"AI Singapore","date":"2024-12-19","domains":["Language"],"country":"Singapore","reference":"https://huggingface.co/aisingapore/Llama-SEA-LION-v3-70B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:c1abd0e370922a7568ff","name":"Llama 3.1 Typhoon 2 70B","organization":"Typhoon / SCB 10X","date":"2024-12-19","domains":["Language"],"country":"Thailand","reference":"https://arxiv.org/abs/2412.13702","confidence":"Speculative","accessibility":"Open weights (restricted use)"},{"id":"epoch:89d5d7f1f8e2d79516c7","name":"Llama 3.1 Typhoon 2 8B","organization":"Typhoon / SCB 10X","date":"2024-12-19","domains":["Language"],"country":"Thailand","reference":"https://arxiv.org/abs/2412.13702","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:f694d702a47960a80e47","name":"Typhoon 2 7B","organization":"Typhoon / SCB 10X","date":"2024-12-19","domains":["Language"],"country":"Thailand","reference":"https://arxiv.org/abs/2412.13702","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ad98d2ce1e46a8b75eb7","name":"LLama 3.2 Typhoon 2 3B","organization":"Typhoon / SCB 10X","date":"2024-12-19","domains":["Language"],"country":"Thailand","reference":"https://arxiv.org/abs/2412.13702","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:ca63201b2fc68511955a","name":"LLama 3..2 Typhoon 2 1B","organization":"Typhoon / SCB 10X","date":"2024-12-19","domains":["Language"],"country":"Thailand","reference":"https://arxiv.org/abs/2412.13702","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:7d11e3e1c0dc981f1347","name":"Typhoon2-Vision","organization":"Typhoon / SCB 10X","date":"2024-12-19","domains":["Vision","Language","Multimodal"],"country":"Thailand","reference":"https://arxiv.org/abs/2412.13702","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:cbf44b1ec61a50727b64","name":"Typhoon2-Audio","organization":"Typhoon / SCB 10X","date":"2024-12-19","domains":["Audio","Language","Speech","Multimodal"],"country":"Thailand","reference":"https://arxiv.org/abs/2412.13702","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:69c4de2b21163752defd","name":"Kling 1.6 Pro","organization":"Kuaishou Technology","date":"2024-12-19","domains":["Video","Vision"],"country":"China","reference":"https://ir.kuaishou.com/news-releases/news-release-details/kuaishou-kling-ai-unveils-multi-image-reference-feature-further/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:e4e7c32b37ed01b421c1","name":"Granite 3.1 2B","organization":"IBM","date":"2024-12-18","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/ibm-granite/granite-3.1-2b-base","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:35568ed51bdcaf030769","name":"Granite 3.1 8B","organization":"IBM","date":"2024-12-18","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/ibm-granite/granite-3.1-8b-base","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ea20c0ecda540571d242","name":"Eleven Flash v2.5","organization":"ElevenLabs","date":"2024-12-18","domains":["Speech"],"country":"United States of America","reference":"https://elevenlabs.io/blog/meet-flash","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:1468900f935e4203a7c9","name":"Falcon3-7B","organization":"Technology Innovation Institute","date":"2024-12-17","domains":["Language"],"country":"United Arab Emirates","reference":"https://huggingface.co/blog/falcon3","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c080be64a4201821d155","name":"Veo 2","organization":"Google DeepMind","date":"2024-12-16","domains":["Video","Vision"],"country":"United States of America","reference":"https://deepmind.google/technologies/veo/veo-2/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:6f5343d83d745bf89a14","name":"SEA-LION-v1-7B-IT","organization":"AI Singapore","date":"2024-12-16","domains":["Language"],"country":"Singapore","reference":"https://huggingface.co/aisingapore/SEA-LION-v1-7B-IT","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2e64767f08872268b6ee","name":"F5-TTS","organization":"Shanghai Jiao Tong University,University of Cambridge,Geely Automobile Research Institute (Ningbo) Company","date":"2024-12-15","domains":["Speech"],"country":"China,United Kingdom,China","reference":"https://arxiv.org/abs/2410.06885","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:069a7ee5c0a9fcd05101","name":"Jinshi (\u91d1\u77f3\u5927\u6a21\u578b)","organization":"Wuxi Sixiang Digital Intelligence Technology Co., Ltd.","date":"2024-12-15","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250624054958/https://m.thepaper.cn/newsDetail_forward_29934565","confidence":"Unknown","accessibility":""},{"id":"epoch:3e8b1a3588bd5ce4501f","name":"Pika 2.0","organization":"Pika Labs","date":"2024-12-15","domains":["Video"],"country":"United States of America","reference":"https://pikalabs.org/pika-2-0/\n\nhttps://www.linkedin.com/posts/pika-labs_today-we-launched-our-pika-20-model-superior-activity-7273465246919380994-8R40?","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:9bd422515e17a375a0fb","name":"Apollo 3B","organization":"Meta AI,Stanford University","date":"2024-12-13","domains":["Video","Language","Multimodal"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2412.10360","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:196ed7bbe6d0dc88a532","name":"Apollo 1.5B","organization":"Meta AI,Stanford University","date":"2024-12-13","domains":["Video","Language","Multimodal"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2412.10360","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:dc9a138255843764905c","name":"360zhinao2-o1","organization":"360 Security Technology","date":"2024-12-13","domains":["Language"],"country":"China","reference":"https://zhuanlan.zhihu.com/p/19327450022","confidence":"Unknown","accessibility":""},{"id":"epoch:67d1d2434cc4c31d8148","name":"Apollo 7B","organization":"Meta AI,Stanford University","date":"2024-12-13","domains":["Multimodal","Video","Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2412.10360","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2118814a7ff4ba6fbaaf","name":"GigaChat Lite (GigaChat-20B-A3B)","organization":"Sber","date":"2024-12-13","domains":["Language"],"country":"Russia","reference":"https://habr.com/ru/companies/sberdevices/articles/865996/","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:88f3ec4e0e3206b81136","name":"Phi-4","organization":"Microsoft Research","date":"2024-12-12","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2412.08905","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4010fc1658abdb0f697d","name":"Gemini 2.0 Flash","organization":"Google DeepMind,Google","date":"2024-12-11","domains":["Language","Vision","Audio","Speech","Video","Multimodal"],"country":"United States of America,United States of America","reference":"https://blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/#ceo-message","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:fb0b398a313abc216c49","name":"Gemini 2.0 Pro","organization":"Google DeepMind","date":"2024-12-11","domains":["Language","Multimodal","Vision","Video","Audio"],"country":"United States of America","reference":"https://deepmind.google/technologies/gemini/pro/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:731f5cb4406701dc823c","name":"Coconut","organization":"Facebook,University of California San Diego","date":"2024-12-11","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2412.06769","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:892a6f5c845ebdbf9a59","name":"T-Pro","organization":"T-Bank","date":"2024-12-11","domains":["Language"],"country":"Russia","reference":"https://habr.com/ru/companies/tbank/articles/865582/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:491c4929a73df4f896b4","name":"T-Lite","organization":"T-Bank","date":"2024-12-11","domains":["Language"],"country":"Russia","reference":"https://habr.com/ru/companies/tbank/articles/865582/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c619c29e493d9d6b79d3","name":"Sora Turbo","organization":"OpenAI","date":"2024-12-09","domains":["Video","Vision"],"country":"United States of America","reference":"https://openai.com/index/sora-is-here/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:58b67a2cc7ecae6df820","name":"EXAONE 3.5 2.4B","organization":"LG AI Research","date":"2024-12-09","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/abs/2412.04862","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:2c5f5719bd499e322c8f","name":"EXAONE 3.5 7.8B","organization":"LG AI Research","date":"2024-12-09","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/abs/2412.04862","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:3f7adc3999378258ce2e","name":"EXAONE 3.5 32B","organization":"LG AI Research","date":"2024-12-09","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/abs/2412.04862","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:73d6ed9180e113c7a205","name":"xTrimoPGLM - 1B","organization":"","date":"2024-12-09","domains":[],"country":"","reference":"https://arxiv.org/abs/2401.06199","confidence":"Confident","accessibility":""},{"id":"epoch:ddcc0678910db7d3dad1","name":"Grok Image Generation / Aurora","organization":"xAI","date":"2024-12-09","domains":["Image generation"],"country":"United States of America","reference":"https://x.ai/news/grok-image-generation-release","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:a576808d7139ea58c68c","name":"Llama 3.3 70B","organization":"Meta AI","date":"2024-12-06","domains":["Language"],"country":"United States of America","reference":"https://www.llama.com/docs/model-cards-and-prompt-formats/llama3_3/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:188844eaebbed20bf37e","name":"InternVL2_5-26B","organization":"Shanghai AI Lab,SenseTime,Tsinghua University,Nanjing University,Fudan University,Chinese University of Hong Kong (CUHK),Shanghai Jiao Tong University","date":"2024-12-06","domains":["Multimodal","Language","Vision"],"country":"China,Hong Kong,China,China,China,Hong Kong,China","reference":"https://arxiv.org/abs/2412.05271","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:715a42e76a3f6311a176","name":"InternVL2_5-38B","organization":"Shanghai AI Lab,SenseTime,Tsinghua University,Nanjing University,Fudan University,Chinese University of Hong Kong (CUHK),Shanghai Jiao Tong University","date":"2024-12-06","domains":["Multimodal","Language","Vision"],"country":"China,Hong Kong,China,China,China,Hong Kong,China","reference":"https://arxiv.org/abs/2412.05271","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:85c140d610fb8cfa4e8c","name":"InternVL2_5-78B","organization":"Shanghai AI Lab,SenseTime,Tsinghua University,Nanjing University,Fudan University,Chinese University of Hong Kong (CUHK),Shanghai Jiao Tong University","date":"2024-12-06","domains":["Multimodal","Language","Vision"],"country":"China,Hong Kong,China,China,China,Hong Kong,China","reference":"https://arxiv.org/abs/2412.05271","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:7440ce8db55e69e83816","name":"o1","organization":"OpenAI","date":"2024-12-05","domains":["Language","Mathematics","Multimodal"],"country":"United States of America","reference":"https://openai.com/index/introducing-chatgpt-pro/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:a6a435c7c50b177dce3e","name":"NVILA 15B","organization":"Nvidia,Massachusetts Institute of Technology (MIT),University of California (UC) Berkeley,University of California San Diego,University of Washington,Tsinghua University","date":"2024-12-05","domains":["Vision","Language","Multimodal","Video"],"country":"United States of America,United States of America,United States of America,United States of America,United States of America,China","reference":"https://arxiv.org/abs/2412.04468","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:4a06ce10cae57edd73ab","name":"NVILA 8B","organization":"Nvidia,Massachusetts Institute of Technology (MIT),University of California (UC) Berkeley,University of California San Diego,University of Washington,Tsinghua University","date":"2024-12-05","domains":["Vision","Language","Multimodal","Video"],"country":"United States of America,United States of America,United States of America,United States of America,United States of America,China","reference":"https://arxiv.org/abs/2412.04468","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:67dfb7c0f7ac64d53723","name":"Infinity","organization":"ByteDance","date":"2024-12-05","domains":["Image generation"],"country":"China","reference":"https://arxiv.org/abs/2412.04431v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:22094d38a8580e63b407","name":"Pleias 1.0 350m","organization":"PleIAs","date":"2024-12-05","domains":["Language"],"country":"France","reference":"https://huggingface.co/PleIAs/Pleias-350m-Preview","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:29cf8710f9c1a9a1ab86","name":"Pleias 1.0 1.2B","organization":"PleIAs","date":"2024-12-05","domains":["Language"],"country":"France","reference":"https://huggingface.co/PleIAs/Pleias-1.2b-Preview","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2129f266f82c9d4970b7","name":"Genie 2","organization":"Google DeepMind","date":"2024-12-04","domains":["Games","Video","3D modeling"],"country":"United States of America","reference":"https://deepmind.google/discover/blog/genie-2-a-large-scale-foundation-world-model/","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:f22525800c5816a14653","name":"voyage-code-3","organization":"Voyage AI","date":"2024-12-04","domains":["Language"],"country":"United States of America","reference":"https://blog.voyageai.com/2024/12/04/voyage-code-3/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:6023fa03f0df6c839d2f","name":"TokenFlow-XL","organization":"ByteDance","date":"2024-12-04","domains":["Vision","Multimodal","Language"],"country":"China","reference":"https://arxiv.org/abs/2412.03069v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:bb27ae11020dc3484df6","name":"TokenFlow-t2i","organization":"ByteDance","date":"2024-12-04","domains":["Image generation"],"country":"China","reference":"https://arxiv.org/abs/2412.03069v1","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5f78e1c1fe393cc43dfc","name":"Amazon Nova Pro","organization":"Amazon","date":"2024-12-03","domains":["Multimodal","Language","Video","Vision"],"country":"United States of America","reference":"https://aws.amazon.com/es/blogs/aws/introducing-amazon-nova-frontier-intelligence-and-industry-leading-price-performance/\n\nhttps://assets.amazon.science/96/7d/0d3e59514abf8fdcfafcdc574300/nova-tech-report-20250317-0810.pdf","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:c6d1d1490a86f326e730","name":"Amazon Nova Lite","organization":"Amazon","date":"2024-12-03","domains":["Multimodal","Language","Video","Vision"],"country":"United States of America","reference":"https://aws.amazon.com/es/blogs/aws/introducing-amazon-nova-frontier-intelligence-and-industry-leading-price-performance/\n\nhttps://assets.amazon.science/96/7d/0d3e59514abf8fdcfafcdc574300/nova-tech-report-20250317-0810.pdf","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:ea377530843b3aead439","name":"Amazon Nova Micro","organization":"Amazon","date":"2024-12-03","domains":["Language"],"country":"United States of America","reference":"https://aws.amazon.com/es/blogs/aws/introducing-amazon-nova-frontier-intelligence-and-industry-leading-price-performance/\n\nhttps://assets.amazon.science/96/7d/0d3e59514abf8fdcfafcdc574300/nova-tech-report-20250317-0810.pdf","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:bdbbd270a13a9599009c","name":"Luma Photon","organization":"LumaLabs","date":"2024-12-03","domains":["Image generation","Vision"],"country":"United States of America","reference":"https://www.luma-ai.com/luma-photon/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:c08bcf050804e0641bfa","name":"Hunyuan Video","organization":"Tencent","date":"2024-12-03","domains":["Video"],"country":"China","reference":"https://www.arxiv.org/abs/2412.03603","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:9d1877b106eb0a7bf85e","name":"Intelligent Go-Explore (IGE)","organization":"University of British Columbia (UBC),Vector Institute,CIFAR AI Research","date":"2024-12-03","domains":["Games"],"country":"Canada,Canada,Canada","reference":"https://arxiv.org/abs/2405.15143","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0e87f414e021b811c1da","name":"Hailuo I2V-01-Live","organization":"MiniMax,Hailuo AI","date":"2024-12-03","domains":["Video","Vision"],"country":"China,Singapore","reference":"https://hailuoai.video/discover/329354592735850496","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:38e9e87167dcd953b658","name":"Amazon Nova Canvas","organization":"Amazon","date":"2024-12-03","domains":["Image generation","Vision"],"country":"United States of America","reference":"https://docs.aws.amazon.com/ai/responsible-ai/nova-canvas/overview.html\n\nhttps://assets.amazon.science/96/7d/0d3e59514abf8fdcfafcdc574300/nova-tech-report-20250317-0810.pdf","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:a12c51af669af41bb9c3","name":"Luma Photon Flash","organization":"LumaLabs","date":"2024-12-03","domains":["Image generation","Vision"],"country":"United States of America","reference":"https://lumalabs.ai/photon","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:1ccc6691531ca5932401","name":"Poolside Malibu","organization":"Poolside","date":"2024-12-02","domains":["Language"],"country":"United States of America","reference":"https://reinvent.awsevents.com/content/dam/reinvent/2024/slides/aim/AIM133-S_Dive-into-poolside-AI-models-just-for-software-engineering-sponsored-by-poolside.pdf","confidence":"Unknown","accessibility":""},{"id":"epoch:10cadffe32e764bced53","name":"RNAformer","organization":"University of Freiburg","date":"2024-12-01","domains":["Biology"],"country":"Germany","reference":"https://www.biorxiv.org/content/10.1101/2024.02.12.579881v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:1ed4cb58fe3da0865b28","name":"INTELLECT-1","organization":"Prime Intellect,Hugging Face,Arcee AI","date":"2024-11-29","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://github.com/PrimeIntellect-ai/prime/blob/main/INTELLECT_1_Technical_Report.pdf","confidence":"Confident","accessibility":""},{"id":"epoch:22689d983fa0e87c308d","name":"360gpt2-pro","organization":"360 Security Technology","date":"2024-11-29","domains":["Language"],"country":"China","reference":"","confidence":"Unknown","accessibility":""},{"id":"epoch:9f2fe9610eaf643caf49","name":"abab7","organization":"MiniMax","date":"2024-11-29","domains":["Language"],"country":"China","reference":"https://www.minimaxi.com/en/news/abab65-series","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:f19dc14ec2fabab0659a","name":"Zhiyan (\u667a\u8a00)","organization":"4Paradigm","date":"2024-11-29","domains":[],"country":"China","reference":"https://web.archive.org/web/20250624064035/https://xueqiu.com/4558658792/330973753","confidence":"Unknown","accessibility":""},{"id":"epoch:ad6eceab3cc6b4f44a7f","name":"Mingzhi Guwen","organization":"Shanghai Shuheng Information Technology Co., Ltd.","date":"2024-11-29","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250624085542/https://i.ifeng.com/c/8fAlXCUhM2w","confidence":"Unknown","accessibility":""},{"id":"epoch:34b914ec61deb97450b7","name":"MiLi (\u7c73\u7c92)","organization":"CreditEase","date":"2024-11-29","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250624094523/https://www.sohu.com/a/832814445_103207","confidence":"Unknown","accessibility":""},{"id":"epoch:480164c44cf2a38c9e37","name":"HaiYue (\u6d6a\u6f6e\u901a\u7528\u8f6f\u4ef6\u6709\u9650\u516c\u53f8)","organization":"Inspur","date":"2024-11-29","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250707083041/https://cn.chinadaily.com.cn/a/202505/29/WS683804cca310205377035a8e.html\nhttps://web.archive.org/web/20250707083526/https://www.sohu.com/a/800383638_282002","confidence":"Confident","accessibility":""},{"id":"epoch:c26784ed4e13649c7b2d","name":"DeepThought-8B","organization":"Ruliad","date":"2024-11-27","domains":["Language"],"country":"United Kingdom","reference":"https://huggingface.co/ruliad/deepthought-8b-llama-v0.01-alpha","confidence":"Confident","accessibility":"API access"},{"id":"epoch:74d3cf44b367eb16dddb","name":"Ovis1.6-Gemma2-27B","organization":"Alibaba","date":"2024-11-26","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://huggingface.co/AIDC-AI/Ovis1.6-Gemma2-27B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ffa8d06a8ca381e62e05","name":"ControlNet (SD 3.5 Large) Depth","organization":"Stability AI","date":"2024-11-26","domains":["Image generation"],"country":"United Kingdom","reference":"https://stability.ai/news/sd3-5-large-controlnets","confidence":"Unknown","accessibility":"Open weights (restricted use)"},{"id":"epoch:bdf357fcc8b4da604079","name":"ControlNet (SD 3.5 Large) Blur","organization":"Stability AI","date":"2024-11-26","domains":["Image generation"],"country":"United Kingdom","reference":"https://stability.ai/news/sd3-5-large-controlnets","confidence":"Unknown","accessibility":"Open weights (restricted use)"},{"id":"epoch:0e9700d133ac61ad5673","name":"ControlNet (SD 3.5 Large) Canny","organization":"Stability AI","date":"2024-11-26","domains":["Image generation"],"country":"United Kingdom","reference":"https://stability.ai/news/sd3-5-large-controlnets","confidence":"Unknown","accessibility":"Open weights (restricted use)"},{"id":"epoch:219d2ca06d0c472234ba","name":"Fugatto 1","organization":"Nvidia","date":"2024-11-25","domains":["Multimodal","Language","Audio"],"country":"United States of America","reference":"https://research.nvidia.com/publication/2024-11_fugatto-1-foundational-generative-audio-transformer-opus-1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f18ffa487388fbf3c500","name":"ConfRank","organization":"University of Bonn,Institute for Numerical Simulation,Fraunhofer Institute for Algorithms and Scientific Computing","date":"2024-11-24","domains":["Biology"],"country":"Germany,Germany,Germany","reference":"https://pubs.acs.org/doi/abs/10.1021/acs.jcim.4c01524","confidence":"Confident","accessibility":""},{"id":"epoch:ea4709661d9456834619","name":"Fish-Speech 1.5","organization":"Fish Audio","date":"2024-11-24","domains":["Speech"],"country":"United States of America","reference":"https://huggingface.co/fishaudio/fish-speech-1.5","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:db4a618bc1509ae4074e","name":"Hymba","organization":"Nvidia","date":"2024-11-22","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2411.13676","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:676775761b501b6ac411","name":"Tulu 3 (T\u00fclu 3) 70B","organization":"Allen Institute for AI,University of Washington","date":"2024-11-21","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2411.15124\n\nhttps://allenai.org/blog/tulu-3-technical","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:d2fd49234a0300afff6a","name":"Gauss2","organization":"Samsung","date":"2024-11-21","domains":["Multimodal","Language","Vision"],"country":"South Korea","reference":"https://news.samsung.com/global/samsung-electronics-hosts-samsung-developer-conference-korea-2024-unveils-its-improved-gen-ai-model","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:37122fb7df6f1b0d4340","name":"Tulu 3 8B","organization":"Allen Institute for AI,University of Washington","date":"2024-11-21","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2411.15124\n\nhttps://allenai.org/blog/tulu-3-technical","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:ae35ab2aa62b838c6646","name":"DeepSeek-R1-Lite-Preview","organization":"DeepSeek","date":"2024-11-20","domains":["Language"],"country":"China","reference":"https://x.com/deepseek_ai/status/1859200141355536422","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:be970a252873d6002f2e","name":"GPT-4o (Nov 2024)","organization":"OpenAI","date":"2024-11-20","domains":["Multimodal","Language","Audio","Speech","Vision"],"country":"United States of America","reference":"https://help.openai.com/en/articles/9624314-model-release-notes","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:aa5915324dd266586429","name":"Suno v4","organization":"Suno","date":"2024-11-19","domains":["Audio"],"country":"United States of America","reference":"https://suno.com/blog/v4","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:1ac9f9a136347a3e85e7","name":"Boltz-1","organization":"Massachusetts Institute of Technology (MIT), Genesis Therapeutics","date":"2024-11-19","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.11.19.624167v2.abstract","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:61e691d146c9d9e6672d","name":"Pixtral Large","organization":"Mistral AI","date":"2024-11-18","domains":["Multimodal","Language","Vision"],"country":"France","reference":"https://mistral.ai/news/pixtral-large/","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:faa5c9034c96b716e6c1","name":"360Zhinao2-7B","organization":"360 Security Technology","date":"2024-11-18","domains":["Language"],"country":"China","reference":"https://github.com/Qihoo360/360zhinao2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:dff310a7d443a0f55107","name":"BiRNA-BERT","organization":"Bangladesh University of Engineering and Technology,University of California Riverside,Carnegie Mellon University (CMU)","date":"2024-11-18","domains":["Biology"],"country":"India,United States of America,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.07.02.601703v1","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:6509469373c4ee437ecb","name":"SK Telecom Telco","organization":"SK Telecom","date":"2024-11-18","domains":["Language","Multimodal","Vision"],"country":"South Korea","reference":"https://news.sktelecom.com/en/1647","confidence":"Unknown","accessibility":""},{"id":"epoch:14e5ca04c67c3670b0d3","name":"k0-math","organization":"Moonshot","date":"2024-11-16","domains":["Mathematics","Language"],"country":"China","reference":"https://web.archive.org/web/20250124103542/https://www.globaltimes.cn/page/202411/1323248.shtml","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:eb8ad56e9b292afb437d","name":"Gemini-Exp-1114","organization":"Google DeepMind","date":"2024-11-15","domains":["Language"],"country":"United States of America","reference":"https://www.tomsguide.com/ai/google-gemini/google-drops-new-gemini-model-and-it-goes-straight-to-the-top-of-the-llm-leaderboard","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:d8534b0170dda0123308","name":"FaXin","organization":"","date":"2024-11-15","domains":[],"country":"","reference":"https://www.stdaily.com/web/gdxw/2024-11/15/content_259094.html","confidence":"Confident","accessibility":""},{"id":"epoch:9eecb64d925785ce8495","name":"LLaVA-CoT","organization":"Peking University,Tsinghua University,Peng Cheng Laboratory,Alibaba DAMO Academy,Lehigh University","date":"2024-11-15","domains":["Language","Vision","Multimodal"],"country":"China,China,China,China,United States of America","reference":"https://arxiv.org/abs/2411.10440","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:4e9fb0c9fbccc369ffaa","name":"bailing-pro-1120","organization":"Ant Group","date":"2024-11-15","domains":["Language"],"country":"China","reference":"","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:ac755084f15d5bf14132","name":"Qwen2.5-Turbo","organization":"Alibaba","date":"2024-11-15","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2.5-turbo/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:4d7838b4ce2377a3f4c9","name":"Mistral Large 2.1","organization":"Mistral AI","date":"2024-11-15","domains":["Language"],"country":"France","reference":"https://huggingface.co/mistralai/Mistral-Large-Instruct-2411","confidence":"Unknown","accessibility":"Open weights (non-commercial)"},{"id":"epoch:5bc1c73e0419286a6f33","name":"Athene-V2","organization":"Nexusflow","date":"2024-11-14","domains":["Language"],"country":"United States of America","reference":"https://nexusflow.ai/blogs/athene-v2","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:9522d2c83190da764d4c","name":"Gemma2 9B CPT Sahabat-AI","organization":"Indosat,Tech Mahindra,AI Singapore,GoTo","date":"2024-11-14","domains":["Language"],"country":"Indonesia,India,Singapore,Indonesia","reference":"https://sahabat-ai.com/\n\nhttps://huggingface.co/GoToCompany/gemma2-9b-cpt-sahabatai-v1-instruct","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:6eac25d947253f2bd17b","name":"Vidu 1.5","organization":"ShengShu","date":"2024-11-13","domains":["Video","Vision"],"country":"China","reference":"https://www.163.com/dy/article/JGSQ7OOF05566WT8.html\n\nhttps://www.notebookcheck.net/Shengshu-Technology-reveals-a-new-AI-video-generator-to-take-on-OpenAI-s-Sora.917537.0.html","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:7069d81f42dc67c6be9a","name":"Qwen2.5-Coder (32B)","organization":"Alibaba","date":"2024-11-12","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2409.12186","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:13c1fe67803d570cc421","name":"MassiveFold","organization":"Universit\u00e9 de Lille,Link\u00f6ping University,Universite de Technologie de Compi\u00e8gne \u2013 CNRS","date":"2024-11-11","domains":["Biology"],"country":"France,Sweden,France","reference":"https://www.researchsquare.com/article/rs-4319486/v1","confidence":"Unknown","accessibility":""},{"id":"epoch:350b08a1721bfb16a919","name":"SeedEdit","organization":"ByteDance","date":"2024-11-11","domains":["Vision"],"country":"China","reference":"https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ljhwZthlaukjlkulzlp/SeedEdit.pdf\nhttps://team.doubao.com/en/special/seededit","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:97baaf3e3ff4b875a98d","name":"NatureLM-audio","organization":"Earth Species Project","date":"2024-11-11","domains":["Audio"],"country":"United States of America","reference":"https://arxiv.org/abs/2411.07186","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:89aa0ea37481eec1269a","name":"stFormer","organization":"Shanghai Jiao Tong University,Chinese Academy of Sciences","date":"2024-11-09","domains":["Biology"],"country":"China,China","reference":"https://www.biorxiv.org/content/10.1101/2024.09.27.615337v5.abstract","confidence":"Likely","accessibility":""},{"id":"epoch:05e025b7f832e522b6a6","name":"Community Research Earth Digital Intelligence Twin (CREDIT)","organization":"","date":"2024-11-09","domains":[],"country":"","reference":"https://arxiv.org/abs/2411.07814","confidence":"Unknown","accessibility":""},{"id":"epoch:61a4e3fd46b01633f409","name":"Fish-Speech 1.4","organization":"Fish Audio","date":"2024-11-09","domains":["Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2411.01156","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:a17c330b4b94a91d9fad","name":"TeleChat2-7B","organization":"China Telecom","date":"2024-11-08","domains":["Language"],"country":"China","reference":"https://github.com/Tele-AI/TeleChat2","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:70b1640349e6f01ab7f2","name":"TeleChat2-3B","organization":"China Telecom","date":"2024-11-08","domains":["Language"],"country":"China","reference":"https://github.com/Tele-AI/TeleChat2","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:92aa3d60c9267f552b70","name":"Mistral Moderation","organization":"Mistral AI","date":"2024-11-07","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/mistral-moderation","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:021bb7b674daff433e4b","name":"Hunyuan-Large","organization":"Tencent","date":"2024-11-06","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2411.02265","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:f3ef4b4b7925b25651ab","name":"FLUX1.1 [pro] Ultra","organization":"Black Forest Labs","date":"2024-11-06","domains":["Image generation"],"country":"Germany","reference":"https://blackforestlabs.ai/flux-1-1-ultra/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:08defad7f7a84b05b353","name":"FLUX1.1 [pro] Raw","organization":"Black Forest Labs","date":"2024-11-06","domains":["Image generation"],"country":"Germany","reference":"https://blackforestlabs.ai/flux-1-1-ultra/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:b595ce3f0bfc850a08da","name":"OpenPhenom-S/16","organization":"Recursion Pharmaceuticals","date":"2024-11-05","domains":["Biology","Vision"],"country":"United States of America","reference":"https://www.rxrx.ai/phenom#:~:text=We%20call%20this%20model%20Phenom-Beta.%20It%20flexibly%20processes,create%20a%20meaningful%20representation%20of%20the%20input%20image.","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:63bdd772210fb9599373","name":"Llama-3.1-Minitron-4B","organization":"Nvidia","date":"2024-11-04","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2407.14679","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5e3d6fb7346d46c9d77e","name":"Minitron 8B","organization":"Nvidia","date":"2024-11-04","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2407.14679","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:1f416b5c73ad0a85a93d","name":"Minitron 4B","organization":"Nvidia","date":"2024-11-04","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2407.14679","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:45581e3e6357f98fe35c","name":"KwooVa","organization":"Harbin Institute of Technology","date":"2024-11-02","domains":["Vision"],"country":"China","reference":"https://github.com/HIT-Kwoo/KwooVa","confidence":"Confident","accessibility":"API access"},{"id":"epoch:ec2f41541e71fafd3fab","name":"KwooLa","organization":"Harbin Institute of Technology","date":"2024-11-02","domains":["Language"],"country":"China","reference":"https://github.com/HIT-Kwoo/KwooLa","confidence":"Confident","accessibility":"API access"},{"id":"epoch:e80d6d1b93637323ec93","name":"KwooGr","organization":"Harbin Institute of Technology","date":"2024-11-02","domains":[],"country":"China","reference":"https://github.com/HIT-Kwoo/KwooGr","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:53fa4fd8018ecca8ade7","name":"FvFold","organization":"Jeonbuk National University","date":"2024-11-01","domains":["Biology"],"country":"South Korea","reference":"https://www.sciencedirect.com/science/article/pii/S0010482524012137","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4eeacfad0b4463f93466","name":"Uni-Med","organization":"Tsinghua University,Beijing University of Posts and Telecommunications","date":"2024-11-01","domains":[],"country":"China,China","reference":"https://arxiv.org/pdf/2409.17508","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:1981cf103fed71b30e92","name":"SimPO","organization":"Princeton University,University of Virginia","date":"2024-11-01","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2405.14734","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:802148bbb91d6f4b8d43","name":"\u03c00 (pi-zero)","organization":"Physical Intelligence","date":"2024-10-31","domains":["Robotics","Vision"],"country":"United States of America","reference":"https://www.physicalintelligence.company/download/pi0.pdf\nhttps://arxiv.org/abs/2410.24164","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4d1ac9a6879da5d1302c","name":"Aha (\u5927\u6a21\u578b)","organization":"Shanghai Xingzhi Technology Co., Ltd.","date":"2024-10-31","domains":["Multimodal"],"country":"China","reference":"https://blog.moontak.com/id/311916","confidence":"Unknown","accessibility":""},{"id":"epoch:d84a0a5bf6f7ba81f2e4","name":"VASA-1","organization":"Microsoft Research Asia","date":"2024-10-31","domains":["Video","Audio"],"country":"China","reference":"https://arxiv.org/abs/2404.10667","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:46a4adbe88314e3e8965","name":"AstroOne","organization":"Zhijiang Lab","date":"2024-10-30","domains":["Astronomy"],"country":"China","reference":"https://en.zhejianglab.com/news/202501/t20250127_4158.shtml?utm_source=chatgpt.com","confidence":"Confident","accessibility":""},{"id":"epoch:64d00eb0dbae400ca4bf","name":"aiXcoder-7B Base","organization":"Peking University","date":"2024-10-30","domains":["Language"],"country":"China","reference":"https://arxiv.org/html/2410.13187v2","confidence":"Confident","accessibility":""},{"id":"epoch:360a6807ae1ecb646106","name":"Recraft V3","organization":"Recraft","date":"2024-10-30","domains":["Image generation"],"country":"United Kingdom","reference":"https://www.recraft.ai/blog/recraft-introduces-a-revolutionary-ai-model-that-thinks-in-design-language","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:d96cafffb11948d4a67c","name":"Universal-2-TF","organization":"AssemblyAI","date":"2024-10-30","domains":["Speech"],"country":"","reference":"https://www.assemblyai.com/research/universal-2","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:10157abc6887420c59bf","name":"Super XiaoAI","organization":"","date":"2024-10-29","domains":[],"country":"","reference":"https://www.aibase.com/zh/news/12800","confidence":"Unknown","accessibility":""},{"id":"epoch:660551fb88cc6b7de86d","name":"Haiper 2.0","organization":"Haiper","date":"2024-10-29","domains":["Image generation"],"country":"United Kingdom","reference":"https://docs.haiper.ai/api-reference/endpoint/2-0-text-to-image","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:9e089659630a6f1cb45f","name":"Stable Diffusion 3.5 Medium","organization":"Stability AI","date":"2024-10-29","domains":["Image generation"],"country":"United Kingdom","reference":"https://stability.ai/news/introducing-stable-diffusion-3-5","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:22dbc3b6f33c7d180986","name":"Pro-PRIME","organization":"Shanghai Jiao Tong University,Shanghai AI Lab,East China University of Science and Technology,Shanghai Tech University,Guangzhou Inernational Bio Island,Chinese Academy of Sciences,Shanghai Academy of Experimental Medicine","date":"2024-10-28","domains":["Biology"],"country":"China,China,China,China,China,China,China","reference":"https://arxiv.org/abs/2307.12682","confidence":"Confident","accessibility":""},{"id":"epoch:767ff8e835759bfb567d","name":"Doubao-pro","organization":"ByteDance","date":"2024-10-28","domains":["Language"],"country":"China","reference":"https://www.volcengine.com/docs/6360/1264663","confidence":"Likely","accessibility":"API access"},{"id":"epoch:6aab3905b5ab2f4e44b2","name":"Aya Expanse 32B","organization":"Cohere Labs (formerly Cohere for AI)","date":"2024-10-24","domains":["Language"],"country":"Canada","reference":"https://cohere.com/blog/aya-expanse-connecting-our-world\n\nhttps://huggingface.co/CohereForAI/aya-expanse-32b?ref=cohere-ai.ghost.io","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:adac81a5bfe6e362d7e4","name":"Aya Expanse 8B","organization":"Cohere Labs (formerly Cohere for AI)","date":"2024-10-24","domains":["Language"],"country":"Canada","reference":"https://cohere.com/blog/aya-expanse-connecting-our-world\n\nhttps://huggingface.co/CohereForAI/aya-expanse-8b?ref=cohere-ai.ghost.io","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:b9388e9864228922adb7","name":"Bielik 7B","organization":"SpeakLeash,Cyfronet AGH","date":"2024-10-24","domains":["Language"],"country":"Poland,Poland","reference":"https://arxiv.org/abs/2410.18565","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fadd2695961b1c5aeb6a","name":"GigaChat MAX","organization":"Sber","date":"2024-10-24","domains":["Language","Vision","Multimodal"],"country":"Russia","reference":"https://habr.com/ru/companies/sberdevices/articles/855368/\n\nhttps://www.sberbank.ru/ru/sberpress/tekhnologii/article?newsID=9cf4dc40-731d-4e79-849b-e840c9258699&blockID=69b149cd-6db4-45aa-ade1-b6920d771b11&regionID=77&lang=ru&type=NEWS","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:bc46496de7fc232be5c0","name":"YandexGPT 4 Pro","organization":"Yandex","date":"2024-10-24","domains":["Language"],"country":"Russia","reference":"https://yandex.com/company/news/24-10-2024\n\nhttps://habr.com/ru/companies/yandex/articles/852968/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:96d753b02c0334117304","name":"YandexGPT 4 Lite","organization":"Yandex","date":"2024-10-24","domains":["Language"],"country":"Russia","reference":"https://yandex.com/company/news/24-10-2024\n\nhttps://habr.com/ru/companies/yandex/articles/852968/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:c0586b9727c7711bf552","name":"YOLOv11","organization":"Huddersfield University","date":"2024-10-23","domains":["Vision"],"country":"United Kingdom","reference":"https://arxiv.org/html/2410.17725v1","confidence":"Unknown","accessibility":""},{"id":"epoch:84f9563c70594f079031","name":"Mochi 1","organization":"Genmo","date":"2024-10-22","domains":["Video"],"country":"United States of America","reference":"https://www.genmo.ai/blog","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:30b3f6adfca42eeb11b3","name":"NVLM-D 72B","organization":"Nvidia","date":"2024-10-22","domains":["Vision","Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2409.11402","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:e8f8c4c0afc39685ea73","name":"NVLM-H 72B","organization":"Nvidia","date":"2024-10-22","domains":["Vision","Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2409.11402","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:121504cf1ac6efb5dcdc","name":"NVLM-X 72B","organization":"Nvidia","date":"2024-10-22","domains":["Vision","Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2409.11402","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:3e48bef8507947521e33","name":"Claude 3.5 Haiku","organization":"Anthropic","date":"2024-10-22","domains":["Language"],"country":"United States of America","reference":"https://www.anthropic.com/news/3-5-models-and-computer-use","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:3f6b67189f72212eb23b","name":"Stable Diffusion 3.5 Large","organization":"Stability AI","date":"2024-10-22","domains":["Image generation"],"country":"United Kingdom","reference":"https://stability.ai/news/introducing-stable-diffusion-3-5","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:93316da47347a28c5212","name":"Stable Diffusion 3.5 Large Turbo","organization":"Stability AI","date":"2024-10-22","domains":["Image generation"],"country":"United Kingdom","reference":"https://stability.ai/news/introducing-stable-diffusion-3-5","confidence":"Unknown","accessibility":"Open weights (restricted use)"},{"id":"epoch:d4969beb5baa86735ffa","name":"LongVU","organization":"Meta AI,King Abdullah University of Science and Technology (KAUST),Korea University","date":"2024-10-22","domains":["Multimodal","Video","Language"],"country":"United States of America,Saudi Arabia,South Korea","reference":"https://arxiv.org/abs/2410.17434v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e87d57971f0c0bf31846","name":"Granite 3.0 8B","organization":"IBM","date":"2024-10-21","domains":["Language"],"country":"United States of America","reference":"https://github.com/ibm-granite/granite-3.0-language-models/tree/main","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ee0792b0bfc49501447e","name":"Granite 3.0 2B","organization":"IBM","date":"2024-10-21","domains":["Language"],"country":"United States of America","reference":"https://github.com/ibm-granite/granite-3.0-language-models/tree/main","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:65a18ab77d2a505e76e7","name":"Emu3","organization":"Beijing Academy of Artificial Intelligence / BAAI","date":"2024-10-21","domains":["Video","Multimodal","Image generation","Vision","Language"],"country":"China","reference":"https://arxiv.org/abs/2409.18869","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:29e4ef537720426ec0a0","name":"Xunguang","organization":"Alibaba DAMO Academy","date":"2024-10-21","domains":["Video"],"country":"China","reference":"https://www.53ai.com/news/LargeLanguageModel/2024071084710.html\nhttps://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm","confidence":"Unknown","accessibility":""},{"id":"epoch:1d324f20033d9bd1c7c9","name":"Taiyi (\u65f7\u89c6\u592a\u4e59)","organization":"Megvii Inc","date":"2024-10-21","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://web.archive.org/web/20250624062402/https://m.tech.china.com/tech/article/20241101/112024_1596885.html","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:0bd0a3c3543efb21e0a5","name":"Zhiyun Culture LLM (\u667a\u4e91\u6587\u5316\u5927\u6a21\u578b) ","organization":"Xinhua Zhiyun Technology Co., Ltd.","date":"2024-10-21","domains":["Multimodal","Language"],"country":"China","reference":"https://digital.gmw.cn/2025-01/21/content_37811282.htm","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:d40c8de53dd4267c9d0a","name":"Allegro","organization":"Rhymes AI","date":"2024-10-20","domains":["Video"],"country":"United States of America","reference":"https://arxiv.org/abs/2410.15458","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0b5887c2434b1c9eb52e","name":"Depth Anything V2 Giant","organization":"Tik Tok,Hong Kong University","date":"2024-10-20","domains":["3D modeling"],"country":"China,Hong Kong","reference":"https://arxiv.org/abs/2406.09414","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c5d53cc5a4e1185155f7","name":"Depth Anything V2 Large","organization":"Tik Tok,Hong Kong University","date":"2024-10-20","domains":["3D modeling"],"country":"China,Hong Kong","reference":"https://arxiv.org/abs/2406.09414","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:331034993c13913087ec","name":"Sonic","organization":"Cartesia","date":"2024-10-19","domains":["Speech"],"country":"United States of America","reference":"https://cartesia.ai/blog/sonic\n\nhttps://docs.cartesia.ai/2024-11-13/build-with-cartesia/models/tts","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:8695c5b7f3a4c019f776","name":"TeleChat2-35B","organization":"China Telecom","date":"2024-10-18","domains":["Language"],"country":"China","reference":"https://github.com/Tele-AI/TeleChat2","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:16fe2ddf46f2d624d8b4","name":"Janus 1.3B","organization":"DeepSeek,The University of Hong Kong,Peking University","date":"2024-10-17","domains":["Language","Vision","Multimodal"],"country":"China,Hong Kong,China","reference":"https://arxiv.org/abs/2410.13848","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:2b508c6a6a233a5fad06","name":"Yi-Lightning","organization":"01.AI","date":"2024-10-16","domains":["Language"],"country":"China","reference":"https://www.lingyiwanwu.com/en https://platform.lingyiwanwu.com/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:215840d71c5681e797ae","name":"Belle-whisper-larger-v3-turbo-zh","organization":"KE Holdings Inc. (\u201cBeike\u201d)","date":"2024-10-16","domains":["Speech"],"country":"China","reference":"https://huggingface.co/BELLE-2/Belle-whisper-large-v3-turbo-zh","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:feeaed89739957983225","name":"Ministral 3B","organization":"Mistral AI","date":"2024-10-16","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/ministraux/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:a4ca199ff9751a268c7e","name":"Ministral 8B","organization":"Mistral AI","date":"2024-10-16","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/ministraux/","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:da040096546e697c90b4","name":"Marco-o1","organization":"Alibaba","date":"2024-10-16","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2411.14405","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7e641ea755d5f3ac0f55","name":"CHAI-1","organization":"Chai discovery","date":"2024-10-15","domains":["Biology"],"country":"United States of America","reference":"https://www.chaidiscovery.com/blog/introducing-chai-1\nhttps://www.biorxiv.org/content/10.1101/2024.10.10.615955v2","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:88a6ee26650c9eb72651","name":"Pika 1.5","organization":"Pika Labs","date":"2024-10-15","domains":["Video"],"country":"United States of America","reference":"https://pikartai.com/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:c403490d78ff21cd6824","name":"Tiangong SkyPaint","organization":"Kunlun Inc.","date":"2024-10-15","domains":["Image generation"],"country":"China","reference":"https://web.archive.org/web/20250708154400/https://www.digitaling.com/articles/884470.html","confidence":"Unknown","accessibility":""},{"id":"epoch:bb4b444cf581b6c0d674","name":"Firefly Video","organization":"Adobe","date":"2024-10-14","domains":["Video","Vision"],"country":"United States of America","reference":"https://news.adobe.com/news/2024/10/101424-adobe-launches-firefly-video-model","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:bc48e724937cc9071e19","name":"SANA 1.6B","organization":"Nvidia,Massachusetts Institute of Technology (MIT),Tsinghua University","date":"2024-10-14","domains":["Image generation"],"country":"United States of America,United States of America,China","reference":"https://arxiv.org/abs/2410.10629","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:9ab2ab8008be31cd8437","name":"Alibaba-NLP (mGTE)","organization":"Alibaba,Hong Kong Polytechnic University","date":"2024-10-14","domains":["Language"],"country":"China,Hong Kong","reference":"https://arxiv.org/pdf/2407.19669","confidence":"Confident","accessibility":""},{"id":"epoch:eea4e92b13c7c4e9e5ca","name":"MolPath","organization":"Southwest Petroleum University,East China Normal University","date":"2024-10-13","domains":["Biology","Materials science"],"country":"China,China","reference":"https://academic.oup.com/bioinformatics/advance-article/doi/10.1093/bioinformatics/btae574/7818417","confidence":"Likely","accessibility":""},{"id":"epoch:4caebb5c02db876eab43","name":"RNADiffFold","organization":"Hangzhou Institute of Medicine,Zhejiang University (ZJU),University of Chinese Academy of Sciences","date":"2024-10-13","domains":["Biology"],"country":"China,China,China","reference":"https://www.biorxiv.org/content/10.1101/2024.05.28.596177v2.abstract","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:932b28517f6f694da013","name":"PROPERMAB","organization":"Regeneron","date":"2024-10-12","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.10.10.616558v1","confidence":"Confident","accessibility":""},{"id":"epoch:719e1ad2aebbad02900b","name":"Yuel 2","organization":"Pennsylvania State University","date":"2024-10-12","domains":["Biology"],"country":"United States of America","reference":"https://doi.org/10.1101/2024.10.08.617219","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:ec04fc31aabe40f7290f","name":"vScreenML 2.0","organization":"Fox Chase Cancer Center,Temple University School of Pharmacy","date":"2024-10-12","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://doi.org/10.1101/2024.10.08.617248","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:1c13a6b06f58999bfcd0","name":"EvoBind2","organization":"Stockholm University,Science for Life Laboratory","date":"2024-10-12","domains":["Biology"],"country":"Sweden,Sweden","reference":"https://www.biorxiv.org/content/10.1101/2024.06.20.599739v2.abstract","confidence":"Unknown","accessibility":""},{"id":"epoch:cd23abd0386ae3517f33","name":"Llama-3.1-Nemotron-70B-Instruct","organization":"Nvidia,Meta AI","date":"2024-10-12","domains":["Language"],"country":"United States of America,United States of America","reference":"https://catalog.ngc.nvidia.com/orgs/nim/teams/nvidia/containers/llama-3.1-nemotron-70b-instruct","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:acb05bdf51f65bd181c4","name":"ALICE","organization":"Nanhu Brain-Computer Interface Institute,Lingang Laboratory,Medical School of Nantong University,Zhejiang University School of Medicine","date":"2024-10-11","domains":["Biology"],"country":"China,China,China,China","reference":"https://www.biorxiv.org/content/10.1101/2024.10.11.617764v1","confidence":"Confident","accessibility":""},{"id":"epoch:cccc1b03eea8a2d5cbfd","name":"Deep Learning Enabled Discovery of Kinase Drug Targets in Pharos","organization":"West Virginia University,University of New Mexico","date":"2024-10-11","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://doi.org/10.1101/2024.10.08.612754","confidence":"Speculative","accessibility":""},{"id":"epoch:c1b6e9b351532f258c7d","name":"T2V-Turbo-v2","organization":"University of California Santa Barbara (UCSB),University of California Los Angeles (UCLA),Amazon,University of Waterloo","date":"2024-10-11","domains":["Video"],"country":"United States of America,United States of America,United States of America,Canada","reference":"https://arxiv.org/abs/2410.05677","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:84594d7190ef1f9de633","name":"Baichuan-Omni","organization":"Baichuan,Westlake University,Zhejiang University (ZJU)","date":"2024-10-11","domains":["Language","Vision","Multimodal","Audio","Speech","Video"],"country":"China,China,China","reference":"https://arxiv.org/abs/2410.08565v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:eef844cd3d048348e8ec","name":"ProteinChat","organization":"University of California San Diego,BioMap Research,The Scripps Research Institute,Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)","date":"2024-10-10","domains":["Biology"],"country":"United States of America,China,United States of America,United Arab Emirates","reference":"https://www.researchsquare.com/article/rs-4941886/v1","confidence":"Likely","accessibility":""},{"id":"epoch:f4c0bcdc95a47425982c","name":"HaloClass","organization":"University of California Davis,Pt. Jawahar Lal Nehru Memorial Medical College,Purdue University","date":"2024-10-10","domains":["Biology"],"country":"United States of America,India,United States of America","reference":"https://www.researchsquare.com/article/rs-5027369/v1","confidence":"Confident","accessibility":""},{"id":"epoch:9a50dc95e93bc882732d","name":"IEV2MOL","organization":"Tokyo Institute of Technology","date":"2024-10-10","domains":["Biology"],"country":"Japan","reference":"https://pubs.acs.org/doi/full/10.1021/acs.jcim.4c00842","confidence":"Likely","accessibility":""},{"id":"epoch:03aaf29eb1203c80df06","name":"RDT-1B","organization":"Tsinghua University","date":"2024-10-10","domains":["Robotics"],"country":"China","reference":"https://arxiv.org/abs/2410.07864","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7ee0ab4c7a16786eb9ab","name":"ProCALM (Uniref9B)","organization":"Profluent Bio,California Institute of Technology","date":"2024-10-09","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://www.arxiv.org/abs/2410.03634","confidence":"Confident","accessibility":""},{"id":"epoch:c9833a01a3bbaccf8951","name":"SCUBA-D","organization":"University of Science and Technology of China (USTC),Oristruct Biotech Company,iFLYTEK Research","date":"2024-10-09","domains":["Biology"],"country":"China,United Kingdom,China","reference":"https://www.nature.com/articles/s41592-024-02437-w","confidence":"Unknown","accessibility":""},{"id":"epoch:b2c20810575647712393","name":"AF_unmasked","organization":"Link\u00f6ping University,Uppsala University,Stockholm University","date":"2024-10-09","domains":["Biology"],"country":"Sweden,Sweden,Sweden","reference":"https://www.nature.com/articles/s41467-024-52951-w","confidence":"Unknown","accessibility":""},{"id":"epoch:e02f5ebced6e92584668","name":"Palmyra X 004","organization":"Writer","date":"2024-10-09","domains":["Language"],"country":"United States of America","reference":"https://writer.com/engineering/actions-with-palmyra-x-004/","confidence":"","accessibility":"API access"},{"id":"epoch:de1f13ce21c980c910e4","name":"Chirp 2 Speech-to-Text","organization":"Google","date":"2024-10-09","domains":["Speech"],"country":"United States of America","reference":"https://cloud.google.com/speech-to-text/v2/docs/chirp_2-model","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:9cb04482195107af59e0","name":"Genesis","organization":"Ecole Polytechnique F\u00b4ed\u00b4erale de Lausanne (EPFL),Swiss Institute of Bioinformatics,Imperial College London,University of Oxford,Prescient Design","date":"2024-10-08","domains":["Biology"],"country":"Switzerland,Switzerland,United Kingdom,United Kingdom,United States of America","reference":"https://www.cell.com/cell-systems/fulltext/S2405-4712(24)00270-9","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:17ed6f2e1e6f10357593","name":"CLEAN-Contact","organization":"Cleveland Clinic,Kent State University,Pacific Northwest National Laboratory","date":"2024-10-08","domains":["Biology"],"country":"United States of America,United States of America,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.05.14.594148v2.abstract","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b82e288b78e7ff6038ef","name":"SO3LR","organization":"University of Luxembourg,Technische Universitat Berlin,Berlin Institute for the Foundations of Learning and Data,DeepMind,Max Planck Institute for Informatics,Korea University","date":"2024-10-08","domains":["Biology"],"country":"Luxembourg,Germany,Germany,United Kingdom,Germany,South Korea","reference":"https://chemrxiv.org/engage/chemrxiv/article-details/6704263051558a15ef6478b6","confidence":"Likely","accessibility":""},{"id":"epoch:bc17b48a65f2e3c4948f","name":"CogVideoX","organization":"Z.ai (Zhipu AI),Tsinghua University","date":"2024-10-08","domains":["Video"],"country":"China,China","reference":"https://arxiv.org/abs/2408.06072","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:c773b14a44dd7f94dac5","name":"GR-2","organization":"ByteDance","date":"2024-10-08","domains":["Robotics"],"country":"China","reference":"https://arxiv.org/abs/2410.06158v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f083ba4c75f81bbe3a25","name":"Pyramid Flow","organization":"Peking University,Kuaishou Technology,Beijing University of Posts and Telecommunications","date":"2024-10-08","domains":["Video"],"country":"China,China,China","reference":"https://arxiv.org/abs/2410.05954","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ab5457d4c51e348c30b0","name":"Aria","organization":"Rhymes AI","date":"2024-10-08","domains":["Multimodal","Language","Video","Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2410.05993","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1ddba47d7e2ddcdb16cc","name":"MLDD3UTRmRRNAS","organization":"Ginkgo Bioworks","date":"2024-10-07","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.10.07.616676v1","confidence":"Likely","accessibility":"API access"},{"id":"epoch:a20a5709c5cc6d332996","name":"scHyena","organization":"Korea Advanced Institute of Science and Technology (KAIST)","date":"2024-10-04","domains":["Biology"],"country":"South Korea","reference":"https://arxiv.org/abs/2310.02713","confidence":"Confident","accessibility":""},{"id":"epoch:e0f337a0879b33e25861","name":"Movie Gen Video","organization":"Meta AI","date":"2024-10-04","domains":["Video","Vision"],"country":"United States of America","reference":"https://ai.meta.com/static-resource/movie-gen-research-paper","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:5c00ca23c432627ccdc3","name":"Movie Gen Audio","organization":"Meta AI","date":"2024-10-04","domains":["Audio"],"country":"United States of America","reference":"https://ai.meta.com/static-resource/movie-gen-research-paper","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0766b9b687e73fe16e2c","name":"EnzymeFlow","organization":"McGill University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),Hong Kong University of Science and Technology (HKUST),University of Washington,Microsoft Research,DeepMind,Shanghai Jiao Tong University,University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2024-10-01","domains":["Biology"],"country":"Canada,Canada,Hong Kong,United States of America,United States of America,United Kingdom,China,Canada","reference":"https://arxiv.org/abs/2410.00327","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:7fe7c72fa600031cd0ec","name":"BindCraft","organization":"Ecole Polytechnique F\u00b4ed\u00b4erale de Lausanne (EPFL),University of Zurich,University of Lausanne,Massachusetts Institute of Technology (MIT),Visterra Inc,Swiss Federal Institute of Technology","date":"2024-10-01","domains":["Biology"],"country":"Switzerland,Switzerland,Switzerland,United States of America,United States of America,Switzerland","reference":"https://www.biorxiv.org/content/10.1101/2024.09.30.615802v1.abstract","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:b85c0f9fec5531865859","name":"PlasmidGPT","organization":"Harvard University","date":"2024-10-01","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.09.30.615762v1.abstract","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e1d1c7a0f5baac1e29ef","name":"LFM 40B","organization":"Liquid","date":"2024-09-30","domains":["Language"],"country":"United States of America","reference":"https://www.liquid.ai/liquid-foundation-models","confidence":"Confident","accessibility":"API access"},{"id":"epoch:2059dff13ae79eb8dd8e","name":"FragLlama","organization":"YDS Pharmatech","date":"2024-09-30","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.09.28.615626v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:2778489f7a181a345060","name":"Takane","organization":"Fujitsu,Cohere","date":"2024-09-30","domains":["Language"],"country":"Japan,Canada","reference":"https://www.fujitsu.com/global/about/resources/news/press-releases/2024/0930-01.html","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:537681e2789d6e414c9b","name":"PocketFlow","organization":"University of Science and Technology of China (USTC),State Key Laboratory of Cognitive Intelligence,Harvard University","date":"2024-09-29","domains":["Biology"],"country":"China,China,United States of America","reference":"https://arxiv.org/abs/2409.19520","confidence":"Likely","accessibility":""},{"id":"epoch:0054b0d37aa217b76ccc","name":"FlexSBDD","organization":"University of Science and Technology of China (USTC),State Key Laboratory of Cognitive Intelligence,Princeton University","date":"2024-09-29","domains":["Biology"],"country":"China,China,United States of America","reference":"https://arxiv.org/abs/2409.19645","confidence":"Confident","accessibility":""},{"id":"epoch:227ee5b3a486bb145fb1","name":"DFMDock","organization":"Johns Hopkins University","date":"2024-09-28","domains":["Biology"],"country":"United States of America","reference":"https://pmc.ncbi.nlm.nih.gov/articles/PMC11463455/","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4d1f0afac4982d7a84e4","name":"ConoDL","organization":"Chongqing University,Ministry of Natural Resources (China)","date":"2024-09-28","domains":["Biology"],"country":"China,China","reference":"https://www.biorxiv.org/content/10.1101/2024.09.27.614001v1.abstract","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:e55f149a52bbd13e674a","name":"DeepREAD","organization":"Shape Therapeutics","date":"2024-09-28","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.09.27.613923v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:e330eff92e6c56053640","name":"PepNet","organization":"Shandong University","date":"2024-09-28","domains":["Biology"],"country":"China","reference":"https://www.nature.com/articles/s42003-024-06911-1","confidence":"Confident","accessibility":""},{"id":"epoch:174da5b4d07473f1c4c8","name":"Mdgen","organization":"Massachusetts Institute of Technology (MIT)","date":"2024-09-26","domains":["Biology","Materials science"],"country":"United States of America","reference":"https://arxiv.org/abs/2409.17808","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e57deab4afdfc3d310e4","name":"AlphaChip","organization":"Google DeepMind","date":"2024-09-26","domains":["Other"],"country":"United States of America","reference":"https://deepmind.google/discover/blog/how-alphachip-transformed-computer-chip-design/\nhttps://www.nature.com/articles/s41586-024-08032-5","confidence":"Unknown","accessibility":""},{"id":"epoch:deef21460f31ea383333","name":"SPOT","organization":"Heinrich Heine University,Concordia University","date":"2024-09-26","domains":["Biology"],"country":"Germany,Canada","reference":"https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3002807","confidence":"Confident","accessibility":""},{"id":"epoch:cb6265739710d253540f","name":"Loop-Diffusion","organization":"University of Washington","date":"2024-09-26","domains":["Biology"],"country":"United States of America","reference":"https://arxiv.org/abs/2409.18201","confidence":"Confident","accessibility":""},{"id":"epoch:039d1f56df96c3f217bd","name":"RWKV-5 (Eagle) 7B","organization":"RWKV Foundation,EleutherAI,Ohio State University,University of California Santa Barbara (UCSB),Wroclaw Tech (Wroc\u0142aw University of Science and Technology),Guangdong Laboratory of Artificial Intelligence and Digital Economy (Pazhou Lab),New York University (NYU),Harvard University,Contextual AI,University of Chinese Academy of Sciences,University of California Santa Cruz,Tsinghua University,University of Edinburgh,University of British Columbia (UBC),Pennsylvania State University","date":"2024-09-26","domains":["Language"],"country":"Multinational,United States of America,United States of America,United States of America,Poland,China,United States of America,United States of America,United States of America,China,United States of America,China,United Kingdom,Canada,United States of America","reference":"https://arxiv.org/abs/2404.05892","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:59a9169177aa1033f02b","name":"RWKV-6 (Finch) 3B","organization":"RWKV Foundation,EleutherAI,Ohio State University,University of California Santa Barbara (UCSB),Wroclaw Tech (Wroc\u0142aw University of Science and Technology),Guangdong Laboratory of Artificial Intelligence and Digital Economy (Pazhou Lab),New York University (NYU),Harvard University,Contextual AI,University of Chinese Academy of Sciences,University of California Santa Cruz,Tsinghua University,University of Edinburgh,University of British Columbia (UBC),Pennsylvania State University","date":"2024-09-26","domains":["Language"],"country":"Multinational,United States of America,United States of America,United States of America,Poland,China,United States of America,United States of America,United States of America,China,United States of America,China,United Kingdom,Canada,United States of America","reference":"https://arxiv.org/abs/2404.05892","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:10500385697bef8d54a7","name":"RNA-DCGen","organization":"Bangladesh University of Engineering and Technology,University of California Riverside","date":"2024-09-25","domains":["Biology"],"country":"India,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.09.23.614570v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:924a7663d0bdf7c27bbd","name":"ProteinGenerator","organization":"University of Washington,Institute for Protein Design,Georgia Institute of Technology,Microsoft,Heidelberg University","date":"2024-09-25","domains":["Biology"],"country":"United States of America,United States of America,United States of America,United States of America,Germany","reference":"https://www.nature.com/articles/s41587-024-02395-w","confidence":"Unknown","accessibility":""},{"id":"epoch:cef359fc13271662a3f8","name":"Mothra","organization":"Tokyo Institute of Technology","date":"2024-09-25","domains":["Biology"],"country":"Japan","reference":"","confidence":"Likely","accessibility":""},{"id":"epoch:94f3ae44ad8bf9dbe1ad","name":"GeoAB","organization":"Zhejiang University (ZJU),Westlake University","date":"2024-09-25","domains":["Biology"],"country":"China,China","reference":"https://www.biorxiv.org/content/10.1101/2024.05.15.594274v2.abstract","confidence":"Unknown","accessibility":""},{"id":"epoch:f86d472ff30bf58502e6","name":"ChemNet","organization":"University of Washington","date":"2024-09-25","domains":["Biology"],"country":"United States of America","reference":"https://pmc.ncbi.nlm.nih.gov/articles/PMC11463446/","confidence":"Confident","accessibility":""},{"id":"epoch:64e8e71c7e4fa30c45b1","name":"PPFlow","organization":"Zhejiang University (ZJU),Westlake University","date":"2024-09-25","domains":["Biology"],"country":"China,China","reference":"https://www.biorxiv.org/content/10.1101/2024.03.07.583831v5.abstract","confidence":"Likely","accessibility":""},{"id":"epoch:d7619cc93262ac23bdd9","name":"TxGNN","organization":"Harvard Medical School,Harvard-MIT Program in Health Sciences and Technology,The Mount Sinai Hospital (New York),Broad Institute,Harvard Data Science Initiative,Harvard University,Stanford University","date":"2024-09-25","domains":["Biology"],"country":"United States of America,United States of America,United States of America,United States of America,United States of America,United States of America,United States of America","reference":"https://www.nature.com/articles/s41591-024-03233-x","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:25b138744963ae7903ea","name":"SeaMoon","organization":"Sorbonne University,Universit\u00e9 Grenoble Alpes,Institut Universitaire de France (IUF)","date":"2024-09-25","domains":["Biology"],"country":"France,France,France","reference":"https://www.biorxiv.org/content/10.1101/2024.09.23.614585v1.abstract","confidence":"Likely","accessibility":""},{"id":"epoch:5bcff1c1dceca1fc4af2","name":"KnoMol","organization":"Zhejiang University (ZJU),Jiangsu University of Technology,Zhejiang University School of Medicine","date":"2024-09-25","domains":["Biology"],"country":"China,China,China","reference":"https://pubs.acs.org/doi/abs/10.1021/acs.jcim.4c01092","confidence":"Confident","accessibility":""},{"id":"epoch:5822eb9c1820f8e341ff","name":"CodonMPNN","organization":"Harvard Medical School,Massachusetts Institute of Technology (MIT)","date":"2024-09-25","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2409.17265","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:87f60eaeddb542ef2bd3","name":"Molmo 72B","organization":"Allen Institute for AI,University of Washington","date":"2024-09-25","domains":["Language","Vision","Multimodal"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2409.17146","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:915400054017fa833b34","name":"dnaGrinder","organization":"Hong Kong Polytechnic University","date":"2024-09-24","domains":["Biology"],"country":"Hong Kong","reference":"https://arxiv.org/abs/2409.15697","confidence":"Confident","accessibility":""},{"id":"epoch:0a186bbee3a3ef5b1bc5","name":"Protein-Mamba","organization":"Rensselaer Polytechnic Institute,Stanford University,University of Minnesota,Korea Advanced Institute of Science and Technology (KAIST),University of Illinois Urbana-Champaign (UIUC)","date":"2024-09-24","domains":["Biology"],"country":"United States of America,United States of America,United States of America,South Korea,United States of America","reference":"https://arxiv.org/abs/2409.14617","confidence":"Confident","accessibility":""},{"id":"epoch:f761f7d4648f179c644d","name":"AFP-Deep","organization":"Nanjing University,Yangzhou University","date":"2024-09-24","domains":["Biology"],"country":"China,China","reference":"https://ieeexplore.ieee.org/document/10691918/authors#authors","confidence":"Confident","accessibility":""},{"id":"epoch:d271e2ee6ce074545c38","name":"ProtBFN","organization":"InstaDeep","date":"2024-09-24","domains":["Biology"],"country":"United Kingdom","reference":"https://www.biorxiv.org/content/10.1101/2024.09.24.614734v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:375ecbd3c670b7c2676e","name":"Llama 3.2 11B","organization":"Meta AI","date":"2024-09-24","domains":["Multimodal","Vision","Language"],"country":"United States of America","reference":"https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:ff1055388484de06f19c","name":"Llama 3.2 90B","organization":"Meta AI","date":"2024-09-24","domains":["Multimodal","Vision","Language"],"country":"United States of America","reference":"https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:bbb9004e9d3e3db8379e","name":"Llama 3.2 1B","organization":"Meta AI","date":"2024-09-24","domains":["Language"],"country":"United States of America","reference":"https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:43a7df5b5f90a5a1bcab","name":"Llama 3.2 3B","organization":"Meta AI","date":"2024-09-24","domains":["Language"],"country":"United States of America","reference":"https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:c0c7cbcf57ef50d57766","name":"DrugTar","organization":"Isfahan University of Technology","date":"2024-09-24","domains":["Biology"],"country":"Iran (Islamic Republic of)","reference":"https://www.biorxiv.org/content/10.1101/2024.09.21.614218v1.abstract","confidence":"Likely","accessibility":""},{"id":"epoch:975585972f822e1c2924","name":"Thermostable protein design","organization":"Indraprastha Institute of Information Technology\nDelhi","date":"2024-09-24","domains":["Biology"],"country":"India","reference":"https://www.biorxiv.org/content/10.1101/2024.09.21.614294v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:f477116777265d90a6ce","name":"Jaeger","organization":"University Medicine Greifswald,Utrecht University,Friedrich Schiller University Jena","date":"2024-09-24","domains":["Biology"],"country":"Germany,Netherlands,Germany","reference":"https://www.biorxiv.org/content/10.1101/2024.09.24.612722v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:4b865c7421472c66f31d","name":"Importance of higher-order epistasis in large protein sequence-function relationships","organization":"University of Florida","date":"2024-09-24","domains":["Biology"],"country":"United States of America","reference":"https://pmc.ncbi.nlm.nih.gov/articles/PMC11463489/","confidence":"Confident","accessibility":""},{"id":"epoch:9387c350eec21d5739a7","name":"MTDP","organization":"Chinese University of Hong Kong (CUHK),City University of Hong Kong","date":"2024-09-24","domains":["Biology"],"country":"Hong Kong,Hong Kong","reference":"https://academic.oup.com/bioinformatics/article/40/9/btae567/7772445","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:d99034ac4654bb0462ed","name":"AlphaMut","organization":"Indian Institute of Science Education and Research","date":"2024-09-24","domains":["Biology"],"country":"India","reference":"https://www.biorxiv.org/content/10.1101/2024.09.21.614241v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:34a490c961c2601cae44","name":"PixelDance","organization":"ByteDance","date":"2024-09-24","domains":["Video","Vision"],"country":"China","reference":"https://pixeldance.io/ (LINK BROKEN)","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:91ea78a9069dc9bfe545","name":"ByteDance Seaweed","organization":"ByteDance","date":"2024-09-24","domains":["Video"],"country":"China","reference":"https://kr-asia.com/bytedance-enters-ai-video-race-with-doubaos-pixeldance-and-seaweed","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:0985bac24a051c60c9a2","name":"ProteinSetTransformer","organization":"University of Wisconsin Madison","date":"2024-09-23","domains":["Biology"],"country":"United States of America","reference":"https://www.researchsquare.com/article/rs-4844047/v1","confidence":"Confident","accessibility":""},{"id":"epoch:1df76121f8527bfb09ef","name":"PreAlgPro","organization":"Shanghai Ocean University","date":"2024-09-23","domains":["Biology"],"country":"China","reference":"https://www.sciencedirect.com/science/article/abs/pii/S014181302406570X?via%3Dihub","confidence":"Confident","accessibility":""},{"id":"epoch:750b7cc03dc1af3442f7","name":"PocketGen","organization":"University of Science and Technology of China (USTC),Hefei Comprehensive National Science Center,Harvard University,Broad Institute,Harvard Data Science Initiative","date":"2024-09-23","domains":["Biology"],"country":"China,China,United States of America,United States of America,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.02.25.581968v4.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:b16c8b51cde975a84ac0","name":"TAWFN","organization":"Northeastern University (China)","date":"2024-09-23","domains":["Biology"],"country":"China","reference":"https://academic.oup.com/bioinformatics/article/40/10/btae571/7766190","confidence":"Confident","accessibility":""},{"id":"epoch:feeb56e954daf755dc00","name":"Spark 4.0","organization":"iFlytek","date":"2024-09-23","domains":["Language"],"country":"China","reference":"https://www.iflytek.com/en/news-events/news/218.html","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:84d1d48aeb7d1a2c48c5","name":"AMPLIFY","organization":"Chandar Research Lab,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),Amgen,Polytechnique Montreal,CIFAR AI Research","date":"2024-09-23","domains":["Biology"],"country":"Canada,Canada,United States of America,Canada,Canada","reference":"https://www.biorxiv.org/content/10.1101/2024.09.23.614603v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d353ea1142815cdf5b65","name":"MoEFold2D","organization":"George Washington University","date":"2024-09-22","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.09.18.613732v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8719e0044290593e9dc9","name":"IgGM","organization":"Chinese Academy of Sciences,University of Chinese Academy of Sciences,Tencent","date":"2024-09-22","domains":["Biology"],"country":"China,China,China","reference":"https://www.biorxiv.org/content/10.1101/2024.09.19.613838v1.abstract","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:83e58ea2e87b06cc2d4f","name":"Kongtian Lingmou 3.0","organization":"","date":"2024-09-22","domains":[],"country":"","reference":"https://www.stdaily.com/web/gdxw/2024-09/22/content_232969.html","confidence":"Confident","accessibility":""},{"id":"epoch:9b1422ff31b0b8fdeb92","name":"Kling 1.5 Pro","organization":"Kuaishou Technology","date":"2024-09-22","domains":["Video","Vision"],"country":"China","reference":"https://www.zeniteq.com/blog/kling-1-5-ai-video-generator-is-finally-here-with-major-upgrades","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:24451396552d160aad27","name":"BTFBS","organization":"Nanjing Agricultural University","date":"2024-09-22","domains":["Biology"],"country":"China","reference":"https://www.biorxiv.org/content/10.1101/2024.09.19.613986v1","confidence":"Likely","accessibility":""},{"id":"epoch:a843d92a377c08d90196","name":"PepINVENT","organization":"AstraZeneca,Chalmers University of Technology","date":"2024-09-21","domains":["Biology"],"country":"United Kingdom,Sweden,Sweden","reference":"https://arxiv.org/abs/2409.14040","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d9923aea76bf0f3d110b","name":"ExSelfRL","organization":"Soochow University","date":"2024-09-20","domains":["Biology"],"country":"Taiwan","reference":"https://www.sciencedirect.com/science/article/pii/S0957417424022772","confidence":"Confident","accessibility":""},{"id":"epoch:b3b7ee8c3f03a88ccab4","name":"Telechat2-115B","organization":"China Telecom","date":"2024-09-20","domains":["Language"],"country":"China","reference":"https://huggingface.co/Tele-AI/TeleChat2-115B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:ac129cf4f39b493a4cec","name":"Prithvi WxC","organization":"IBM Research,University of Alabama,Stanford University,Colorado State University,Oak Ridge National Laboratory,NASA","date":"2024-09-20","domains":["Earth science"],"country":"United States of America,United States of America,United States of America,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2409.13598","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ca56c6bb974f96f713ac","name":"Qwen2.5-72B","organization":"Alibaba","date":"2024-09-19","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2.5/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:bf11c362d056f96a3f67","name":"EITLEM-Kinetics","organization":"Beijing University of Chemical Technology","date":"2024-09-19","domains":["Biology"],"country":"China","reference":"https://www.cell.com/chem-catalysis/abstract/S2667-1093(24)00266-5","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4829719699ed230be230","name":"pKALM","organization":"Hokkaido University","date":"2024-09-19","domains":["Biology"],"country":"Japan","reference":"https://www.biorxiv.org/content/10.1101/2024.09.16.613101v1.abstract","confidence":"Speculative","accessibility":""},{"id":"epoch:cc575170cba1a9155957","name":"GeoSeqBuilder","organization":"Peking University","date":"2024-09-19","domains":["Biology"],"country":"China","reference":"https://onlinelibrary.wiley.com/doi/abs/10.1002/ange.202411461?casa_token=oCX9uTFcpvQAAAAA%3AiGc_jXpH-dzeUXL0LgRnySqIJorSnOaPwaPGpjf8PbI9etcMdXXoXzwrEOX-wrQdbRsjcAEkQO0C6w","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:5158087cce3769f7e6ef","name":"McMLP","organization":"Harvard Medical School,University of Illinois Urbana-Champaign (UIUC),Harvard TH Chan School of Public Health","date":"2024-09-19","domains":["Biology"],"country":"United States of America,United States of America,United States of America","reference":"https://pmc.ncbi.nlm.nih.gov/articles/PMC10054958/","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:d6513ab5b0c7fec0ca77","name":"Qwen2.5 Instruct (7B)","organization":"Alibaba","date":"2024-09-19","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2.5/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3cd2e63ebab00594e59e","name":"Qwen2.5 Instruct (72B)","organization":"Alibaba","date":"2024-09-19","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2.5/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:0999f97e50fd5a34bf14","name":"Qwen2.5-3B","organization":"Alibaba","date":"2024-09-19","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2.5-llm/","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:880b93fc19c0e6e6dbd5","name":"Qwen2.5-7B","organization":"Alibaba","date":"2024-09-19","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2.5/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9719d12dd9cac22087f4","name":"Qwen2.5-1.5B","organization":"Alibaba","date":"2024-09-19","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2.5-llm/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f9685715553e4f0db1f0","name":"Qwen2.5-14B","organization":"Alibaba","date":"2024-09-19","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2.5/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:10dbd1c5d0326edfa7b6","name":"Qwen2.5-Math-7B-Base","organization":"Alibaba","date":"2024-09-19","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2.5-math/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a4173af20c262e8548ab","name":"Oryx 34B","organization":"Tsinghua University,Tencent,Nanyang Technological University","date":"2024-09-19","domains":["Multimodal","Vision","3D modeling","Video","Language"],"country":"China,China,Singapore","reference":"https://arxiv.org/abs/2409.12961v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e805435a7cbb2857bbb0","name":"Oryx 7B","organization":"Tsinghua University,Tencent,Nanyang Technological University","date":"2024-09-19","domains":["Multimodal","Vision","Language"],"country":"China,China,Singapore","reference":"https://arxiv.org/abs/2409.12961v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:59c7fef4d3c183ce417b","name":"Qwen2.5 Instruct (32B)","organization":"Alibaba","date":"2024-09-19","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2.5/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:eea630894545f1989023","name":"Qwen2.5-Math-1.5B","organization":"Alibaba","date":"2024-09-19","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2.5-math/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d2d2a1b5b8c6d2a41dd9","name":"Qwen2-VL-72B","organization":"Alibaba","date":"2024-09-18","domains":["Language","Vision","Multimodal"],"country":"China","reference":"https://arxiv.org/abs/2409.12191","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:d52f99f0e902e4f9c2cd","name":"Qwen2-VL-2B","organization":"Alibaba","date":"2024-09-18","domains":["Language","Vision","Multimodal"],"country":"China","reference":"https://arxiv.org/abs/2409.12191","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a113cb751693e63c34ce","name":"Qwen2-VL-7B","organization":"Alibaba","date":"2024-09-18","domains":["Language","Vision","Multimodal"],"country":"China","reference":"https://arxiv.org/abs/2409.12191","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d136d9a558e2b9cd828b","name":"Qwen2.5-Coder (7B)","organization":"Alibaba","date":"2024-09-18","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2409.12186","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:38175a201dca6c5e701d","name":"Qwen2.5-Coder (1.5B)","organization":"Alibaba","date":"2024-09-18","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2409.12186","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:aedf732511e9d45c05fd","name":"RoseTTAFold2-Lite","organization":"University of Washington,University of Texas Southwest Medical Center,Seoul National University,Massachusettes General Hospital,Harvard Medical School,Broad Institute","date":"2024-09-18","domains":["Biology"],"country":"United States of America,United States of America,South Korea,United States of America,United States of America,United States of America","reference":"https://www.nature.com/articles/s41564-024-01791-x","confidence":"Unknown","accessibility":""},{"id":"epoch:4fa0c22ece81c6c7f785","name":"AtomFlow","organization":"Peking University,Chinese University of Hong Kong (CUHK),Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Universit\u00e9 de Montr\u00e9al,HEC Montreal,CIFAR AI Research","date":"2024-09-18","domains":["Biology"],"country":"China,Hong Kong,Canada,Canada,Canada,Canada","reference":"https://arxiv.org/abs/2409.12080","confidence":"Unknown","accessibility":""},{"id":"epoch:55dac0d8625c8f0bbc0a","name":"GraphEC","organization":"Sun Yat-sen University,National Supercomputing Center in Shenzhen,Chongqing University,Key Laboratory of Machine Intelligence and Advanced Computing","date":"2024-09-18","domains":["Biology"],"country":"China,China,China,China","reference":"https://www.nature.com/articles/s41467-024-52533-w","confidence":"Confident","accessibility":""},{"id":"epoch:f179334b92426a3e46a1","name":"ProtENN2","organization":"European Bioinformatics Institute,University of Cambridge,Google Research","date":"2024-09-18","domains":["Biology"],"country":"Multinational,United Kingdom,United Kingdom,United States of America","reference":"https://academic.oup.com/bioinformaticsadvances/article/4/1/vbae132/7760237","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:5f9799d2db707f6c498b","name":"Whale Bioacoustics Model","organization":"Google Research,National Oceanic and Atmospheric Administration (NOAA),Oregon State University","date":"2024-09-18","domains":["Audio"],"country":"United States of America,United States of America,United States of America","reference":"https://research.google/blog/whistles-songs-boings-and-biotwangs-recognizing-whale-vocalizations-with-ai/\n\nhttps://www.frontiersin.org/journals/marine-science/articles/10.3389/fmars.2024.1394695/full","confidence":"Unknown","accessibility":"Open weights (non-commercial)"},{"id":"epoch:b69699ea0f0302d41db3","name":"1X World Model","organization":"1X","date":"2024-09-17","domains":["Robotics","Video"],"country":"United States of America","reference":"https://www.1x.tech/discover/1x-world-model","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:bdab79d877a37c838075","name":"Mistral Small v24.09","organization":"Mistral AI","date":"2024-09-17","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/september-24-release/","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:42775daaf0ed763db769","name":"AminoAcid-0","organization":"Ginkgo Bioworks","date":"2024-09-17","domains":["Biology"],"country":"United States of America","reference":"https://government.ginkgo.bio/resources/blog/aminoacid-0-aa-0-a-protein-llm-trained-with-2-billion-proprietary-sequences","confidence":"Likely","accessibility":"API access"},{"id":"epoch:c86e70d9bec1fdd67315","name":"OmniGen","organization":"Beijing Academy of Artificial Intelligence / BAAI","date":"2024-09-17","domains":["Image generation"],"country":"China","reference":"https://arxiv.org/abs/2409.11340","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:68b90b9d008135367bd2","name":"DeepRelax","organization":"National University of Singapore,Sun Yat-sen University,Peking University,China Medical University Hospital,Asia university,Guangdong L-Med Biotechnology Company","date":"2024-09-17","domains":["Biology"],"country":"Singapore,China,China,China,Taiwan,China","reference":"https://www.nature.com/articles/s41467-024-52378-3","confidence":"Unknown","accessibility":""},{"id":"epoch:19d40ca75683108b9d55","name":"ProTeM","organization":"Zhejiang Lab,Zhejiang University (ZJU),Huazhong University of Science and Technology,Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)","date":"2024-09-17","domains":["Biology"],"country":"China,China,China,United Arab Emirates","reference":"https://link.springer.com/chapter/10.1007/978-3-031-72353-7_10","confidence":"Confident","accessibility":""},{"id":"epoch:69c77afdbab8b595ef74","name":"Pixtral 12B","organization":"Mistral AI","date":"2024-09-17","domains":["Vision","Language","Multimodal"],"country":"France","reference":"https://mistral.ai/news/pixtral-12b/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:746bdb8393e1dc932882","name":"Qwen2.5-32B","organization":"Alibaba","date":"2024-09-17","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2.5/ ","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2192a63a2bf55e0cd8d6","name":"RNA language models predict mutations that improve RNA function","organization":"NERSC, Lawrence Berkeley National Laboratory,University of California San Francisco,University of California (UC) Berkeley","date":"2024-09-16","domains":["Biology"],"country":"United States of America,United States of America,United States of America","reference":"https://pmc.ncbi.nlm.nih.gov/articles/PMC11014562/","confidence":"Confident","accessibility":""},{"id":"epoch:3b36e35f21cce147e65c","name":"DeepUrfold","organization":"University of Virginia","date":"2024-09-16","domains":["Biology"],"country":"United States of America","reference":"https://www.nature.com/articles/s41467-024-52020-2","confidence":"Confident","accessibility":""},{"id":"epoch:689ef13b087c06fa47dd","name":"Ovis1.6-Gemma2-9B","organization":"Alibaba","date":"2024-09-16","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://huggingface.co/AIDC-AI/Ovis1.6-Gemma2-9B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e9df4cb33730458fe8e9","name":"Playground v3","organization":"Playground","date":"2024-09-16","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2409.10695","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:504c4482c0e725e3dbb6","name":"RNAdiffusion","organization":"Princeton University,Tsinghua University,Stanford University","date":"2024-09-15","domains":["Biology"],"country":"United States of America,China,United States of America","reference":"https://arxiv.org/abs/2409.09828","confidence":"Likely","accessibility":""},{"id":"epoch:b77724888bbfac770dd4","name":"UdanDTI","organization":"Tsinghua University","date":"2024-09-15","domains":["Biology"],"country":"China","reference":"https://www.biorxiv.org/content/10.1101/2024.09.12.612771v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:fa47771a11b4e13f0106","name":"Luma Ray 1.6","organization":"LumaLabs","date":"2024-09-15","domains":["Video","Vision"],"country":"United States of America","reference":"https://www.runcomfy.com/playground/luma/luma-ray-1-6","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:961faf23bb503db4455c","name":"ProtRNA","organization":"Fudan University,Shanghai AI Lab","date":"2024-09-14","domains":["Biology"],"country":"China,China","reference":"https://www.biorxiv.org/content/10.1101/2024.09.10.612218v1.abstract","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f6baaf8e20667382cd88","name":"PLTNUM","organization":"Kyoto University,National Institute of Biomedical Innovation,RIKEN","date":"2024-09-14","domains":["Biology"],"country":"Japan,Japan,Japan","reference":"https://www.biorxiv.org/content/10.1101/2024.09.10.612367v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:3c01de79b10518d20457","name":"LEGO","organization":"Chinese Academy of Sciences,Beijing Academy of Artificial Intelligence / BAAI,University of Chinese Academy of Sciences","date":"2024-09-14","domains":["Biology"],"country":"China,China,China","reference":"https://www.biorxiv.org/content/10.1101/2024.09.10.612249v1","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:b8a4ddd276e02c6d83c0","name":"MolSnapper","organization":"University of Oxford","date":"2024-09-14","domains":["Biology"],"country":"United Kingdom","reference":"https://www.biorxiv.org/content/10.1101/2024.03.28.586278v2.abstract","confidence":"Unknown","accessibility":""},{"id":"epoch:08911986d2767fa44eab","name":"NeoaPred","organization":"South China University of Technology,Jinan University","date":"2024-09-14","domains":["Biology"],"country":"China,China","reference":"https://academic.oup.com/bioinformatics/article/40/9/btae547/7758065","confidence":"Confident","accessibility":""},{"id":"epoch:b14f33cffa105e317acb","name":"YOLOv8 (reCAPTCHA fine-tuned)","organization":"ETH Zurich","date":"2024-09-13","domains":["Vision"],"country":"Switzerland","reference":"https://arxiv.org/abs/2409.08831","confidence":"Confident","accessibility":""},{"id":"epoch:daec6591bb4cf21d49df","name":"JURA Bio Model","organization":"JURA Bio,New York University (NYU),Harvard Medical School,Columbia University","date":"2024-09-13","domains":["Biology"],"country":"United States of America,United States of America,United States of America,United States of America","reference":"https://static1.squarespace.com/static/66db3a552002432ff6a39dae/t/66e45d8b7ed2bf1a517e03c1/1726242191877/JURA_VariationalSynthesis.pdf","confidence":"Confident","accessibility":""},{"id":"epoch:286510009fd272d6e96e","name":"IDPFold","organization":"Shandong University,BioMap Research,Fuzhou University,Shanghai Jiao Tong University","date":"2024-09-13","domains":["Biology"],"country":"China,China,China,China","reference":"https://www.biorxiv.org/content/10.1101/2024.05.05.592611.abstract","confidence":"Likely","accessibility":""},{"id":"epoch:c457dead213a434fcce2","name":"Novae","organization":"CentraleSupelec,Gustave Roussy,Universit\u00e9 Paris Cit\u00e9","date":"2024-09-13","domains":["Biology"],"country":"France,France,France","reference":"https://www.biorxiv.org/content/10.1101/2024.09.09.612009v1.abstract","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:39fb4f6e1a5398003b03","name":"CodonTransformer","organization":"Vector Institute,University of Toronto,Universit\u00e9 Paris Cit\u00e9","date":"2024-09-13","domains":["Biology"],"country":"Canada,Canada,France","reference":"https://www.biorxiv.org/content/10.1101/2024.09.13.612903v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:ff96ae963d580e522207","name":"Text2Protein","organization":"University of California San Diego,Brown University","date":"2024-09-13","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://www.researchsquare.com/article/rs-4868665/v1","confidence":"Confident","accessibility":""},{"id":"epoch:bf448cfddb3427a77e4e","name":"o1-mini","organization":"OpenAI","date":"2024-09-12","domains":["Language"],"country":"United States of America","reference":"https://openai.com/index/learning-to-reason-with-llms/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:6253401b1be1637a4e3b","name":"o1-preview","organization":"OpenAI","date":"2024-09-12","domains":["Language","Mathematics","Biology"],"country":"United States of America","reference":"https://openai.com/index/introducing-openai-o1-preview/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:d2e661535b3519bc7274","name":"4D Diffusion for Dynamic Protein Structure Prediction with Reference Guided Motion Alignment","organization":"Fudan University,Shanghai Academy of Artificial Intelligence for Science,Nanjing University","date":"2024-09-12","domains":["Biology"],"country":"China,China,China","reference":"https://arxiv.org/abs/2408.12419","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:89732237cd764c6fe4f4","name":"Demostart (in progress)","organization":"Google DeepMind","date":"2024-09-12","domains":["Robotics"],"country":"United States of America","reference":"https://arxiv.org/abs/2409.06613","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:9b2e73296d2062d02448","name":"DataGemma","organization":"Google","date":"2024-09-12","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2409.13741","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:2b841703061cebebd0d3","name":"E2 TTS","organization":"Microsoft","date":"2024-09-12","domains":["Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2406.18009","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:9192d769f247ac9da918","name":"Automated design of multi-target ligands by generative deep learning","organization":"Goethe University Frankfurt,Fraunhofer Institute for Translational Medicine and Pharmacology,Ludwig Maximilian University of Munich","date":"2024-09-11","domains":["Biology"],"country":"Germany,Germany,Germany","reference":"https://www.nature.com/articles/s41467-024-52060-8","confidence":"Confident","accessibility":""},{"id":"epoch:56e50afad8d9231a33f0","name":"Solar Pro","organization":"Upstage","date":"2024-09-11","domains":["Language"],"country":"South Korea","reference":"https://www.upstage.ai/news/solar-pro-preview","confidence":"Confident","accessibility":"API access"},{"id":"epoch:eaa56ca09545a70680a2","name":"Galaxy (\u661f\u6c49Galaxy\u5927\u6a21\u578b)","organization":"Dahan Software / Hanweb","date":"2024-09-11","domains":["Multimodal","Language","Speech","Vision","Video"],"country":"China","reference":"https://www.hanweb.com/xw/art/2025/art_3be9d8ef79ef4a62b457652b2eb62384.html\nhttps://web.archive.org/web/20250624090839/https://www.hanweb.com/xw/art/2024/art_6fda569f72a7402591fe1d8a51ab7322.html","confidence":"Confident","accessibility":""},{"id":"epoch:0d50b075ebe71623ca78","name":"EVI 2","organization":"Hume","date":"2024-09-11","domains":["Speech"],"country":"United States of America","reference":"https://www.hume.ai/blog/introducing-evi2\n\nhttps://dev.hume.ai/docs/empathic-voice-interface-evi/configuration/evi-version","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:90849b71cf7ae78e01c5","name":"GenMS","organization":"Google DeepMind","date":"2024-09-10","domains":["Materials science"],"country":"United States of America","reference":"https://arxiv.org/abs/2409.06762","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:7647864dfbf5ff7f88e8","name":"KinoML","organization":"Charit\u00e9-Universit\u00e4tsmedizin Berlin,Saarland University,Memorial Sloan Kettering Cancer Center","date":"2024-09-10","domains":["Biology"],"country":"Germany,Germany,United States of America","reference":"https://pmc.ncbi.nlm.nih.gov/articles/PMC11419124/","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:00d8f7376694e812ad0f","name":"CPDiffusion","organization":"Shanghai Jiao Tong University,University of New South Wales,University of Cambridge,Shanghai AI Lab","date":"2024-09-10","domains":["Biology"],"country":"China,Australia,United Kingdom,China","reference":"https://www.nature.com/articles/s41421-024-00728-2","confidence":"Confident","accessibility":""},{"id":"epoch:7b67d5b6fa67e0cc4a8d","name":"MolPhenix","organization":"Valence Labs,University of British Columbia (UBC),Vector Institute,University of Toronto,University of Montreal / Universit\u00e9 de Montr\u00e9al,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms)","date":"2024-09-10","domains":["Biology"],"country":"Canada,Canada,Canada,Canada,Canada,Canada","reference":"https://arxiv.org/abs/2409.08302","confidence":"Confident","accessibility":""},{"id":"epoch:d0494d586ed6d9854765","name":"DiffForce","organization":"University of Cambridge,Shanghai Jiao Tong University,University of New South Wales","date":"2024-09-09","domains":["Biology"],"country":"United Kingdom,China,Australia","reference":"https://arxiv.org/abs/2406.05832","confidence":"Unknown","accessibility":""},{"id":"epoch:4ac0caf9bd7d0089a5f2","name":"ProteinMPNN-DDG","organization":"Peptone","date":"2024-09-09","domains":["Biology"],"country":"United Kingdom","reference":"https://www.biorxiv.org/content/10.1101/2024.06.15.599145v4.abstract","confidence":"Unknown","accessibility":""},{"id":"epoch:9d956f8fada6fb377ae0","name":"ESMIF-DDG","organization":"Peptone","date":"2024-09-09","domains":["Biology"],"country":"United Kingdom","reference":"https://www.biorxiv.org/content/10.1101/2024.06.15.599145v4.abstract","confidence":"Unknown","accessibility":""},{"id":"epoch:11fcdd52964f6731b869","name":"PDFII","organization":"Sichuan University","date":"2024-09-09","domains":["Biology"],"country":"China","reference":"https://www.liebertpub.com/doi/abs/10.1089/cmb.2024.0506","confidence":"Unknown","accessibility":""},{"id":"epoch:c1535564e2a29c616577","name":"EpiScan","organization":"Sun Yat-sen University,Guangzhou National Laboratory","date":"2024-09-09","domains":["Biology"],"country":"China,China","reference":"https://www.nature.com/articles/s41540-024-00432-7","confidence":"Confident","accessibility":""},{"id":"epoch:88cf21ac913d814ad92d","name":"VespaG","organization":"Technical University of Munich,Sorbonne University,Institute for Advanced Study,Universit\u00e9 Paris Cit\u00e9,Institut Universitaire de France (IUF)","date":"2024-09-09","domains":["Biology"],"country":"Germany,France,United States of America,France,France","reference":"https://www.biorxiv.org/content/10.1101/2024.04.24.590982v2.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:2a13f9b4d8dda90b20b7","name":"MMAPLE","organization":"City University of New York,Cornell University","date":"2024-09-09","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://www.nature.com/articles/s42003-024-06797-z","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:56dc4a7cfb0573ac7dd7","name":"AbGPT","organization":"Carnegie Mellon University (CMU)","date":"2024-09-09","domains":["Biology"],"country":"United States of America","reference":"https://arxiv.org/abs/2409.06090","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4a38face3e356d0b1ed6","name":"BetterBodies","organization":"University of Freiburg,Collaborative Research Institute Intelligent Oncology ,BrainLinks-BrainTools","date":"2024-09-09","domains":["Biology"],"country":"Germany,Germany,Germany","reference":"https://arxiv.org/abs/2409.16298","confidence":"Confident","accessibility":""},{"id":"epoch:ef80c87b294214ad884c","name":"RiboCode","organization":"Sun Yat-sen University,Rhegen Biotechnology,Chinese Academy of Sciences","date":"2024-09-08","domains":["Biology"],"country":"China,China,China","reference":"https://www.biorxiv.org/content/10.1101/2024.09.06.611590v1","confidence":"Confident","accessibility":""},{"id":"epoch:90be8ebc4f9be282b5c7","name":"ALOHA Unleashed","organization":"Google DeepMind","date":"2024-09-08","domains":["Robotics"],"country":"United States of America","reference":"https://aloha-unleashed.github.io/assets/aloha_unleashed.pdf\nhttps://arxiv.org/abs/2410.13126","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c319b945697d04f5c89a","name":"DDGemb","organization":"University of Bologna","date":"2024-09-07","domains":["Biology"],"country":"Italy","reference":"https://www.biorxiv.org/content/10.1101/2024.09.05.611455v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:d270a679c50082acee83","name":"ESM-DBP","organization":"Hunan University","date":"2024-09-07","domains":["Biology"],"country":"China","reference":"https://www.nature.com/articles/s41467-024-52293-7","confidence":"Confident","accessibility":""},{"id":"epoch:9a9b2faf81714359dd53","name":"MPDF","organization":"Chinese University of Hong Kong (CUHK),Lanzhou University,Zhejiang Lab,Zhejiang University (ZJU)","date":"2024-09-07","domains":["Biology"],"country":"Hong Kong,China,China,China","reference":"https://arxiv.org/abs/2409.05916","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:765b5614f87abb57ec14","name":"Aristotle","organization":"Harmonic","date":"2024-09-07","domains":["Language"],"country":"United States of America","reference":"https://harmonic.fun/news#blog-post-1-link","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:22507184f5d43e0a6d2f","name":"DeepSeek-V2.5","organization":"DeepSeek","date":"2024-09-06","domains":["Language"],"country":"China","reference":"https://huggingface.co/deepseek-ai/DeepSeek-V2.5","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:c7b09a39f17076deb208","name":"GearBind","organization":"BioGeometry,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Universit\u00e9 de Montr\u00e9al,Fudan University,HEC Montreal","date":"2024-09-06","domains":["Biology"],"country":"China,Canada,Canada,China,Canada","reference":"https://www.nature.com/articles/s41467-024-51563-8","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fe6ace81e637cf273231","name":"ProtSSN","organization":"Shanghai Jiao Tong University,East China University of Science and Technology,Shanghai AI Lab","date":"2024-09-06","domains":["Biology"],"country":"China,China,China","reference":"https://www.biorxiv.org/content/10.1101/2023.12.01.569522v3.abstract","confidence":"Likely","accessibility":""},{"id":"epoch:01b74efb36192f4185ac","name":"SAAMBE-MEM","organization":"Clemson University,Central China Normal University","date":"2024-09-06","domains":["Biology"],"country":"United States of America,China","reference":"https://academic.oup.com/bioinformatics/article/40/9/btae544/7750356","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:3f4cb7853be499a80a70","name":"xLAM-8x22B","organization":"Salesforce","date":"2024-09-06","domains":["Language"],"country":"United States of America","reference":"https://www.salesforce.com/blog/xlam-large-action-models/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:f6bb9df9e5c13536ff57","name":"MiniCPM3.0","organization":"Mianbi Intelligence","date":"2024-09-06","domains":["Language"],"country":"China","reference":"https://www.aibase.com/news/11581","confidence":"Confident","accessibility":""},{"id":"epoch:ae987a7cb02732c9cf75","name":"AlphaProteo","organization":"Google DeepMind","date":"2024-09-05","domains":["Biology"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphaproteo-generates-novel-proteins-for-biology-and-health-research/AlphaProteo2024.pdf","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:c6ed8784428c518e30db","name":"\u00b5Former","organization":"Microsoft Research AI for Science","date":"2024-09-05","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/10.1101/2023.11.16.565910v3.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:983d8e2c26e49033cc68","name":"EnOpt","organization":"University of Pittsburgh","date":"2024-09-05","domains":["Biology"],"country":"United States of America","reference":"https://www.nature.com/articles/s41598-024-71699-3","confidence":"Unknown","accessibility":""},{"id":"epoch:e7de9f8953ee1b942efb","name":"Hunyuan Turbo","organization":"Tencent","date":"2024-09-05","domains":["Language"],"country":"China","reference":"https://zhidx.com/p/442250.html","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:18d1a4c39f60b45b9db0","name":"Xinchen Lingo","organization":"West Lake Xinchen / Xinchen AI / \u897f\u6e56\u5fc3\u8fb0\uff08\u676d\u5dde\uff09\u79d1\u6280\u6709\u9650\u516c\u53f8","date":"2024-09-05","domains":["Speech"],"country":"China","reference":"https://en.tmtpost.com/news/7227890#:~:text=27%2C%202024-,China's%20First%20End%2Dto%2DEnd%20General%20Speech%20Model%2C%20Xincheng,LIKE6\n\nhttps://www.dostor.com/p/89004.html","confidence":"Unknown","accessibility":""},{"id":"epoch:76b4c7046f04c7e14fa5","name":"Harrison.rad.1","organization":"Harrison.ai","date":"2024-09-05","domains":["Vision","Medicine","Language","Multimodal"],"country":"Australia","reference":"https://harrison.ai/news/harrison-ai-launches-world-leading-ai-model-to-transform-healthcare/","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:1ba10e3baeacc24c6720","name":"CHIEF","organization":"Harvard Medical School,Massachusetts Institute of Technology (MIT),Sichuan University,Sun Yat-sen University,Shenzhen Maternity & Child Healthcare Hospital,Chongqing University Cancer Hospital,Harvard University,University of Pennsylvania,Cedars-Sinai Medical Center,Broad Institute,Dana-Farber Cancer Institute,Brigham and Women's Hospital,Tencent,Massachusettes General Hospital,Pennsylvania State University,Jinan University,Stanford University","date":"2024-09-04","domains":["Biology","Vision","Medicine"],"country":"United States of America,United States of America,China,China,China,China,United States of America,United States of America,United States of America,United States of America,United States of America,United States of America,China,United States of America,United States of America,China,United States of America","reference":"https://www.nature.com/articles/s41586-024-07894-z","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:e1a670fda0b500073c0c","name":"Integrating Deep Learning and Synthetic Biology: A Co-Design Approach for Enhancing Gene Expression via N-Terminal Coding Sequences","organization":"National University of Singapore,Jiangnan University","date":"2024-09-04","domains":["Biology"],"country":"Singapore,China","reference":"https://pubs.acs.org/doi/abs/10.1021/acssynbio.4c00371","confidence":"Confident","accessibility":""},{"id":"epoch:b7a31a10680e37aa3a3c","name":"MolMVC","organization":"Central South University,Singapore Agency for Science","date":"2024-09-04","domains":["Biology"],"country":"China,Singapore","reference":"https://academic.oup.com/bioinformatics/article/40/Supplement_2/ii190/7749083","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:ed34f50ba12dcbea8db5","name":"OLMoE","organization":"Allen Institute for AI,Contextual AI,University of Washington,Princeton University","date":"2024-09-03","domains":["Language"],"country":"United States of America,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2409.02060v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e63082bf4be01b04b51b","name":"DrugCLIP","organization":"Tsinghua University,Tsinghua-Peiking Center for Life Sciences,Peking University,Beijing Academy of Artificial Intelligence / BAAI","date":"2024-09-03","domains":["Biology"],"country":"China,China,China,China","reference":"https://www.biorxiv.org/content/10.1101/2024.09.02.610777v1?rss=1","confidence":"Confident","accessibility":""},{"id":"epoch:27bf07bd883f7d988fc5","name":"Alphaflow","organization":"Massachusetts Institute of Technology (MIT)","date":"2024-09-02","domains":["Biology"],"country":"United States of America","reference":"https://arxiv.org/abs/2402.04845","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ee9ed843fe4ba7c0b6ab","name":"ChatMol","organization":"Tsinghua University,PingAn Technology,Beijing University of Posts and Telecommunications","date":"2024-09-02","domains":["Biology"],"country":"China,China,China","reference":"https://academic.oup.com/bioinformatics/article/40/9/btae534/7747661","confidence":"Speculative","accessibility":"Open weights (non-commercial)"},{"id":"epoch:5fbd5d8a71fa9002df5c","name":"ParetoDrug","organization":"Chinese University of Hong Kong (CUHK),Zhejiang Lab,Zhejiang University (ZJU),Huawei Noah's Ark Lab","date":"2024-09-02","domains":["Biology"],"country":"Hong Kong,China,China,China","reference":"https://www.nature.com/articles/s42003-024-06746-w","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:667131a1cfe0f284de2b","name":"DTI-LM","organization":"University of Central Florida","date":"2024-09-02","domains":["Biology"],"country":"United States of America","reference":"https://academic.oup.com/bioinformatics/article-abstract/40/9/btae533/7747660","confidence":"Confident","accessibility":""},{"id":"epoch:799401f70fee987b88af","name":"ESMFlow","organization":"Massachusetts Institute of Technology (MIT)","date":"2024-09-02","domains":["Biology"],"country":"United States of America","reference":"https://arxiv.org/abs/2402.04845","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6294a1ea6b648e817d55","name":"Moonshot-v1","organization":"Moonshot","date":"2024-09-01","domains":["Language"],"country":"China","reference":"","confidence":"Unknown","accessibility":""},{"id":"epoch:ea3fbe9c12e544afdaa7","name":"MiniMax Video-01","organization":"MiniMax","date":"2024-08-31","domains":["Video","Vision"],"country":"China","reference":"https://www.minimaxi.com/en/news/video-01","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:4de14f16f4fcb2319392","name":"HelixFold3","organization":"Tecorigin LTD,Tsinghua University,Baidu","date":"2024-08-30","domains":["Biology"],"country":"China,China,China","reference":"https://arxiv.org/abs/2408.16975","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:79aa24da004721ff6095","name":"RENNAISSANCE","organization":"Ecole Polytechnique F\u00b4ed\u00b4erale de Lausanne (EPFL),University of Cambridge,Harvard Medical School","date":"2024-08-30","domains":["Biology"],"country":"Switzerland,United Kingdom,United States of America","reference":"https://www.nature.com/articles/s41929-024-01220-6.pdf","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6aa7cc5470bebae98d5a","name":"NV-Embed-v2","organization":"Nvidia","date":"2024-08-30","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/nvidia/NV-Embed-v2","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:daef624f1e7b9232ef2e","name":"Shuka-1","organization":"Sarvam","date":"2024-08-30","domains":["Speech"],"country":"India","reference":"https://huggingface.co/sarvamai/shuka-1\n\nhttps://web.archive.org/web/20250523013746/https://www.sarvam.ai/blogs/shuka-v1","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:7e6c25c05b12067b0e46","name":"OPUS-Design","organization":"Fudan University,Shanghai AI Lab,Harcam Biomedicines","date":"2024-08-29","domains":["Biology"],"country":"China,China,China","reference":"https://www.biorxiv.org/content/10.1101/2024.08.20.608889v2.abstract","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:312ae6aba84baaa5a8b0","name":"DrugFormer","organization":"University of Florida,University of Texas Health Science Center,H. Lee Moffitt Cancer Center and Research Institute","date":"2024-08-29","domains":["Biology"],"country":"United States of America,United States of America,United States of America","reference":"https://onlinelibrary.wiley.com/doi/full/10.1002/advs.202405861","confidence":"Likely","accessibility":""},{"id":"epoch:1783f6ad5cc8e6716660","name":"GLM-4-Plus","organization":"Z.ai (Zhipu AI)","date":"2024-08-29","domains":["Language"],"country":"China","reference":"https://bigmodel.cn/dev/howuse/glm-4","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:bc41fcb1f342ee1ea9d7","name":"Hairuo","organization":"Inspur","date":"2024-08-29","domains":["Language"],"country":"China","reference":"https://encloud.inspur.com/hairuo/cp/index.html\n\nhttps://www.eguizhou.gov.cn/guiyang/2024-08/29/c_1016827.htm","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:ee19142a5be7e359b9c8","name":"Bielik-11B-v2","organization":"SpeakLeash,Cyfronet AGH","date":"2024-08-28","domains":["Language"],"country":"Poland,Poland","reference":"https://huggingface.co/speakleash/Bielik-11B-v2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7aa44557e68bbb3bab33","name":"GLA Transformer 1.3B","organization":"MIT-IBM Watson AI Lab,Massachusetts Institute of Technology (MIT)","date":"2024-08-27","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2312.06635","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0a29f0e072590034f566","name":"GLA Transformer 340M","organization":"MIT-IBM Watson AI Lab,Massachusetts Institute of Technology (MIT)","date":"2024-08-27","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2312.06635","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:29dda5a7c49c17526cd6","name":"Pharia-1-LLM-7B","organization":"Aleph Alpha","date":"2024-08-26","domains":["Language"],"country":"Germany","reference":"https://huggingface.co/Aleph-Alpha/Pharia-1-LLM-7B-control","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:db9cafd757a9f19ae99b","name":"DISTRO","organization":"Nous Research","date":"2024-08-26","domains":["Language"],"country":"United States of America","reference":"https://github.com/NousResearch/DisTrO/blob/main/A_Preliminary_Report_on_DisTrO.pdf","confidence":"Confident","accessibility":""},{"id":"epoch:21706cd9b9cc1f6fea6e","name":"Xinyu","organization":"Zhejiang University (ZJU),Institute for Advanced Algorithms Research,Northeastern University (China),China Telecom,State Key Laboratory of Media Convergence Production Technology and Systems","date":"2024-08-23","domains":["Language"],"country":"China,China,China,China,China","reference":"https://arxiv.org/abs/2408.11609","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:69981e1745f45cde7501","name":"Jamba 1.5-Large","organization":"AI21 Labs","date":"2024-08-22","domains":["Language"],"country":"Israel","reference":"https://arxiv.org/abs/2408.12570\nhttps://www.ai21.com/blog/announcing-jamba-model-family\nhttps://huggingface.co/ai21labs/AI21-Jamba-1.5-Large","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:a38d10e7a4d3e6cf96d9","name":"EvolMPNN","organization":"Aarhus university","date":"2024-08-22","domains":["Biology"],"country":"Denmark","reference":"https://link.springer.com/chapter/10.1007/978-3-031-70368-3_24","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:ef634478170253900c38","name":"GameNGen","organization":"Google Research,Google DeepMind,Tel Aviv University","date":"2024-08-22","domains":["Games"],"country":"United States of America,United States of America,Israel","reference":"https://arxiv.org/abs/2408.14837","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:68882a18d2b518ef6377","name":"Shanhai 2.0","organization":"","date":"2024-08-22","domains":[],"country":"","reference":"https://www.doit.com.cn/p/517117.html","confidence":"Unknown","accessibility":""},{"id":"epoch:974a27983e3bf71ae6f1","name":"GITIII","organization":"Yale School of Public Health","date":"2024-08-22","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.08.21.608964v1","confidence":"Unknown","accessibility":""},{"id":"epoch:81d0c4cd0a2e221fea1a","name":"Jamba 1.5 Mini","organization":"AI21 Labs","date":"2024-08-22","domains":["Language"],"country":"Israel","reference":"https://arxiv.org/abs/2408.12570","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:4577cf762cb0dc2fd9ef","name":"CoPRA","organization":"Tsinghua University,University College London (UCL),Monash University,Beijing University of Posts and Telecommunications","date":"2024-08-21","domains":["Biology"],"country":"China,United Kingdom,Australia,China","reference":"https://arxiv.org/abs/2409.03773","confidence":"Likely","accessibility":""},{"id":"epoch:82cac074d21b5dd100eb","name":"Xinyi Video 2.0","organization":"","date":"2024-08-21","domains":[],"country":"","reference":"https://tech.chinadaily.com.cn/a/202408/21/WS66c5a501a310b35299d37c8d.html","confidence":"Unknown","accessibility":""},{"id":"epoch:fc042ff42a2fa9f0ce1b","name":"Ideogram v2","organization":"Ideogram","date":"2024-08-21","domains":["Image generation"],"country":"Canada","reference":"https://about.ideogram.ai/2.0","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:0f6b001d52ba116be404","name":"Kosmos-2.5","organization":"Microsoft","date":"2024-08-21","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2309.11419","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:12481267f85e9b4e2260","name":"AntiFormer","organization":"University of Florida,Sichuan University,Shihezi University,University of Macau,University of Texas Health Science Center","date":"2024-08-20","domains":["Biology"],"country":"United States of America,China,China,Macao,United States of America","reference":"https://academic.oup.com/bib/article/25/5/bbae403/7736247","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:90adc2738613a896be45","name":"Rethinking Molecular Design: Integrating Latent Variable and Auto-Regressive Models for Goal Directed Generation","organization":"ETH Zurich,University of Zurich,ETH AI Center","date":"2024-08-19","domains":["Biology"],"country":"Switzerland,Switzerland,Switzerland","reference":"https://arxiv.org/abs/2409.00046","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:434a1adb16bd16d687de","name":"LongVILA-7B","organization":"Nvidia,Massachusetts Institute of Technology (MIT),University of California (UC) Berkeley,UT Austin","date":"2024-08-19","domains":["Multimodal","Video","Language"],"country":"United States of America,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2408.10188","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:4b8baf87a727fb650dbf","name":"Hybrid-Phi-Mamba-1.5B","organization":"Carnegie Mellon University (CMU),Mohamed bin Zayed University of Artificial Intelligence (MBZUAI),Cartesia","date":"2024-08-19","domains":["Language"],"country":"United States of America,United Arab Emirates,United States of America","reference":"https://arxiv.org/abs/2408.10189","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:eb435dd31e29fee2c86f","name":"TransFew","organization":"University of Missouri","date":"2024-08-17","domains":["Biology"],"country":"United States of America","reference":"https://academic.oup.com/bioinformaticsadvances/article/4/1/vbae120/7735316","confidence":"Confident","accessibility":""},{"id":"epoch:e5a2dfcf7034063b3898","name":"CLR_ESP","organization":"Kansas State University","date":"2024-08-16","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.08.13.607829v1","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:d5de6e40061f38215f92","name":"xGen-MM (BLIP-3)","organization":"Salesforce Research,University of Washington","date":"2024-08-16","domains":["Multimodal","Vision","Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/pdf/2408.08872v1","confidence":"Confident","accessibility":""},{"id":"epoch:25218cecf6978165482c","name":"Transformers in music recommendation","organization":"Google Research","date":"2024-08-16","domains":["Recommendation"],"country":"United States of America","reference":"https://research.google/blog/transformers-in-music-recommendation/","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:a4b3674d160011001136","name":"DeepSeek-Prover-V1.5","organization":"DeepSeek","date":"2024-08-15","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2408.08152","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:989334e5a68c04b6b0f8","name":"Gen-3 Alpha Turbo","organization":"Runway","date":"2024-08-15","domains":["Video","Vision"],"country":"United States of America","reference":"https://runwayml.com/news/introducing-the-runway-api\n\nhttps://help.runwayml.com/hc/en-us/articles/30266515017875-Creating-with-Text-Image-to-Video-on-Gen-3-Alpha-and-Turbo","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:e87a5ae24aa01f614965","name":"Hermes 3 405B","organization":"Nous Research","date":"2024-08-14","domains":["Language"],"country":"United States of America","reference":"https://nousresearch.com/wp-content/uploads/2024/08/Hermes-3-Technical-Report.pdf","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:249c0532cb4e60e6e3a4","name":"Hermes 3 70B","organization":"Nous Research","date":"2024-08-14","domains":["Language"],"country":"United States of America","reference":"https://nousresearch.com/wp-content/uploads/2024/08/Hermes-3-Technical-Report.pdf","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:eecd35c973c011bf339b","name":"Hermes 3 8B","organization":"Nous Research","date":"2024-08-14","domains":["Language"],"country":"United States of America","reference":"https://nousresearch.com/wp-content/uploads/2024/08/Hermes-3-Technical-Report.pdf","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:2a63d02012be27ba3979","name":"Zhiye LLM (\u6d6a\u6f6e\u77e5\u4e1a\u5927\u6a21\u578b)","organization":"Inspur","date":"2024-08-14","domains":["Multimodal","Language","Vision","Video"],"country":"China","reference":"https://web.archive.org/web/20250708145724/https://www.sohu.com/a/813472344_121332524","confidence":"Unknown","accessibility":""},{"id":"epoch:b8991944737cb69cdb11","name":"Grok-2","organization":"xAI","date":"2024-08-13","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://x.ai/blog/grok-2","confidence":"Confident","accessibility":"API access"},{"id":"epoch:d22fe2a964cb7d314714","name":"Grok-2 mini","organization":"xAI","date":"2024-08-13","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://x.ai/blog/grok-2","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:bad9b5a160286e89c77d","name":"Agent Q","organization":"Stanford University","date":"2024-08-13","domains":["Other"],"country":"United States of America","reference":"https://arxiv.org/abs/2408.07199","confidence":"Confident","accessibility":""},{"id":"epoch:42beba81f08d09ea6a71","name":"Saaras v1","organization":"Sarvam","date":"2024-08-13","domains":["Speech"],"country":"India","reference":"https://www.maginative.com/article/sarvam-ai-launches-open-weights-sarvam-2b-model-and-suite-of-ai-products-for-indian-languages/#:~:text=Saaras%201.0%20and%20Saarika:%20Speech,and%20chief%20editor%20of%20Maginative.","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:834c2ec7fbfd9cb2e968","name":"Cosine Genie","organization":"Cosine","date":"2024-08-12","domains":["Language"],"country":"United Kingdom","reference":"https://cosine.sh/blog/genie-technical-report","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:1837802f03ef2a081518","name":"Falcon Mamba","organization":"Technology Innovation Institute","date":"2024-08-12","domains":["Language"],"country":"United Arab Emirates","reference":"https://huggingface.co/tiiuae/falcon-mamba-7b","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:9e99b4ba0f2365e7f91b","name":"P-LLama3","organization":"University of Siena","date":"2024-08-12","domains":["Biology"],"country":"Italy","reference":"https://arxiv.org/abs/2408.06396","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:b7f052dc292ba2ad85db","name":"P-Mistral","organization":"University of Siena","date":"2024-08-12","domains":["Biology"],"country":"Italy","reference":"https://arxiv.org/abs/2408.06396","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4ec7497d76de2b00129d","name":"P-Llama2","organization":"University of Siena","date":"2024-08-12","domains":["Biology"],"country":"Italy","reference":"https://arxiv.org/abs/2408.06396","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:56163aea4ded77d5f877","name":"P-gemma","organization":"University of Siena","date":"2024-08-12","domains":["Biology"],"country":"Italy","reference":"https://arxiv.org/abs/2408.06396","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5e233641c2d1e4e03000","name":"OuteTTS-0.1-350M","organization":"","date":"2024-08-12","domains":["Speech"],"country":"","reference":"https://outeai.com/blog/outetts-0.1-350m","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:605eaaa53c930feceec1","name":"PepMLM","organization":"Duke University,Cornell University,McMaster University","date":"2024-08-11","domains":["Biology"],"country":"United States of America,United States of America,Canada","reference":"https://pmc.ncbi.nlm.nih.gov/articles/PMC10593082/","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:5bb421a2fbc335ff685f","name":"Qwen2-Math-1.5B","organization":"Alibaba","date":"2024-08-09","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2-math/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5f967c6218d2293bca6d","name":"Qwen2-Math-72B","organization":"Alibaba","date":"2024-08-09","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2-math/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:ba257fc6506f983f481d","name":"Qwen2-Math-7B","organization":"Alibaba","date":"2024-08-09","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2-math/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a57bc1d202f7c7745a17","name":"Tencent Search LLM (\u817e\u8baf\u641c\u7d22\u5927\u6a21\u578b)","organization":"Tencent","date":"2024-08-09","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250707092435/https://finance.sina.cn/tech/2023-11-19/detail-imzvamrc5289106.d.html","confidence":"Unknown","accessibility":""},{"id":"epoch:13331eb43ee1577de344","name":"MooER","organization":"Moore Threads","date":"2024-08-09","domains":["Speech"],"country":"China","reference":"https://arxiv.org/pdf/2408.05101","confidence":"Confident","accessibility":""},{"id":"epoch:b2cef55301191d478587","name":"LTM-2-mini","organization":"Magic","date":"2024-08-09","domains":["Language"],"country":"United States of America","reference":"https://magic.dev/blog/100m-token-context-windows","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4e1d53daca6b0dd60dc3","name":"3-ensemble of Self-ensembles on CIFAR-100","organization":"Google DeepMind","date":"2024-08-08","domains":["Vision"],"country":"United States of America","reference":"https://www.arxiv.org/abs/2408.05446","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:caeef4720a2af0d84ca9","name":"EXAONE 3.0","organization":"LG AI Research","date":"2024-08-07","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/abs/2408.03541","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:1fb2360285ae8332cba1","name":"Table Tennis Agent","organization":"Google DeepMind","date":"2024-08-07","domains":["Robotics"],"country":"United States of America","reference":"https://deepmind.google/research/publications/107741/\nhttps://arxiv.org/abs/2408.03906","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:55609d1e87139638a55e","name":"MiniCPM-V 2.6","organization":"OpenBMB (Open Lab for Big Model Base)","date":"2024-08-06","domains":["Vision","Language","Video"],"country":"China","reference":"https://huggingface.co/openbmb/MiniCPM-V-2_6","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:068659a4b8438d95f010","name":"LLaVA-OV-7B","organization":"ByteDance,Nanyang Technological University,Chinese University of Hong Kong (CUHK),Hong Kong University of Science and Technology (HKUST)","date":"2024-08-06","domains":["Multimodal","Video","Vision","Language"],"country":"China,Singapore,Hong Kong,Hong Kong","reference":"https://arxiv.org/abs/2408.03326","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fd4575d9141de1af2025","name":"LLaVA-OV-72B","organization":"ByteDance,Nanyang Technological University,Chinese University of Hong Kong (CUHK),Hong Kong University of Science and Technology (HKUST)","date":"2024-08-06","domains":["Multimodal","Vision","Language","Video"],"country":"China,Singapore,Hong Kong,Hong Kong","reference":"https://arxiv.org/abs/2408.03326","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2f5a43123e7a59aa0b45","name":"GPT-4o (Aug 2024)","organization":"OpenAI","date":"2024-08-06","domains":["Multimodal","Language","Audio","Speech","Vision"],"country":"United States of America","reference":"https://help.openai.com/en/articles/9624314-model-release-notes","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:32e6385feaf55def7555","name":"Jais-70B ","organization":"G42,Inception G42","date":"2024-08-05","domains":["Language"],"country":"United Arab Emirates,United Arab Emirates","reference":"https://www.g42.ai/resources/news/g42-launches-jais-70b-and-20-other-ai-models-champion-arabic-natural-language-processing","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:80b2822308d50c6b0ab7","name":"CXR Foundation","organization":"Google","date":"2024-08-02","domains":["Vision","Medicine"],"country":"United States of America","reference":"https://developers.google.com/health-ai-developer-foundations/cxr-foundation","confidence":"Unknown","accessibility":"Open weights (restricted use)"},{"id":"epoch:c57d45e0b7601035f440","name":"H-optimus-0","organization":"","date":"2024-08-02","domains":["Vision","Biology","Medicine"],"country":"","reference":"https://huggingface.co/bioptimus/H-optimus-0","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:32cf68d27a0e2822932e","name":"Flux.1 [pro]","organization":"Black Forest Labs","date":"2024-08-01","domains":["Image generation"],"country":"Germany","reference":"https://blackforestlabs.ai/announcing-black-forest-labs/ ","confidence":"Confident","accessibility":"API access"},{"id":"epoch:c32d410dd7490d2effcb","name":"Flux.1 [dev]","organization":"Black Forest Labs","date":"2024-08-01","domains":["Image generation"],"country":"Germany","reference":"https://blackforestlabs.ai/announcing-black-forest-labs/ ","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:39a7adb081e536938171","name":"OmniParser: Interactable Region Detection Model","organization":"Microsoft Research","date":"2024-08-01","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2408.00203","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9bf765ddd9ad6f139944","name":"OmniParser: Icon Description Model","organization":"Microsoft Research","date":"2024-08-01","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2408.00203","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ec9947af79e895b1ff64","name":"BaiLing TTS","organization":"Ant Group","date":"2024-08-01","domains":["Speech"],"country":"China","reference":"https://arxiv.org/abs/2408.00284v1","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:559b95d21ee7f1888c9d","name":"EPInformer","organization":"The University of Hong Kong,Harvard Medical School","date":"2024-08-01","domains":["Biology"],"country":"Hong Kong,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.08.01.606099v1","confidence":"Confident","accessibility":""},{"id":"epoch:633be05500695671882c","name":"InternLM2.5","organization":"Shanghai AI Lab","date":"2024-08-01","domains":["Language"],"country":"China","reference":"https://github.com/InternLM/InternLM/blob/main/model_cards/internlm2.5_20b.md","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:d775294aa04fca56cb71","name":"Midjourney V6.1","organization":"Midjourney","date":"2024-07-30","domains":["Image generation"],"country":"United States of America","reference":"https://updates.midjourney.com/version-6-1/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:969ac374f0b67c1508b2","name":"Llama SEA-LION V2 8B","organization":"AI Singapore","date":"2024-07-30","domains":["Language"],"country":"Singapore","reference":"https://huggingface.co/aisingapore/Llama-SEA-LION-v2-8B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:849695ddf59a7449f8bf","name":"AFM-on-device","organization":"Apple","date":"2024-07-29","domains":["Language"],"country":"United States of America","reference":"https://machinelearning.apple.com/research/apple-intelligence-foundation-language-models","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:36f6d09db23de6432aa0","name":"AFM-server","organization":"Apple","date":"2024-07-29","domains":["Language"],"country":"United States of America","reference":"https://machinelearning.apple.com/research/apple-intelligence-foundation-language-models","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:10085d66eda73731fc36","name":"RNACG","organization":"Tsinghua University","date":"2024-07-29","domains":["Biology"],"country":"China","reference":"https://arxiv.org/abs/2407.19838","confidence":"Confident","accessibility":""},{"id":"epoch:4cace3c2e98c2f8dc6a7","name":"Segment Anything Model 2","organization":"Meta AI","date":"2024-07-29","domains":["Vision","Video"],"country":"United States of America","reference":"https://ai.meta.com/research/publications/sam-2-segment-anything-in-images-and-videos/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:69bb4be3fa0206c7b297","name":"Florence-2-B (base)","organization":"Microsoft","date":"2024-07-29","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2311.06242","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:8c80f157d82ac0ee77f7","name":"Florence-2-L (large)","organization":"Microsoft","date":"2024-07-29","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2311.06242","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:d13db1926b6a3685878e","name":"Yi Vision","organization":"01.AI","date":"2024-07-29","domains":["Vision"],"country":"China","reference":"https://platform.01.ai/docs#models-and-pricing","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:3b7e7bf78855bf9180f1","name":"ProRNA3D-Single","organization":"Virginia Tech (Virginia Polytechnic Institute and State University)","date":"2024-07-28","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.07.27.605468v1.abstract","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f2b412338ef652b70169","name":"SaulLM-large","organization":"Equall.ai","date":"2024-07-28","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2407.19584","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:08f8510762764cca20c3","name":"NIO World Model (\u851a\u6765\u5927\u6a21\u578b)","organization":"NIO","date":"2024-07-27","domains":["Robotics","Vision"],"country":"China","reference":"https://web.archive.org/web/20250603031645/https://www.nio.cn/smart-technology/20241120002\nhttps://web.archive.org/web/20240729110402/https://finance.sina.com.cn/roll/2024-07-29/doc-incfuvpe5809920.shtml","confidence":"Unknown","accessibility":""},{"id":"epoch:d3acdf48e72cf04c5590","name":"AlphaProof","organization":"Google DeepMind","date":"2024-07-25","domains":["Mathematics"],"country":"United States of America","reference":"https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/\n\nhttps://www.nature.com/articles/s41586-025-09833-y \n\nhttps://static-content.springer.com/esm/art%3A10.1038%2Fs41586-025-09833-y/MediaObjects/41586_2025_9833_MOESM1_ESM.pdf","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:8997eb14130a94f81686","name":"AlphaGeometry 2","organization":"Google DeepMind","date":"2024-07-25","domains":["Mathematics","Language"],"country":"United States of America","reference":"https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:f7770e1fd48abd334602","name":"X-Portrait","organization":"ByteDance","date":"2024-07-25","domains":["Video","Vision"],"country":"China","reference":"https://arxiv.org/abs/2403.15931v4","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7144dd31d1263e7bac28","name":"SearchGPT","organization":"OpenAI","date":"2024-07-25","domains":["Search"],"country":"United States of America","reference":"https://openai.com/index/searchgpt-prototype/","confidence":"Unknown","accessibility":""},{"id":"epoch:2c7a221f3c1d99f79d56","name":"Mistral Large 2","organization":"Mistral AI","date":"2024-07-24","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/mistral-large-2407/","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:f46b39fbc84850e3da6d","name":"Stable Video 4D (SV4D)","organization":"Stability AI,Northeastern University","date":"2024-07-24","domains":["Vision","Video","3D modeling"],"country":"United Kingdom,United States of America","reference":"https://sv4d.github.io/static/sv4d_technical_report.pdf","confidence":"Unknown","accessibility":"Open weights (restricted use)"},{"id":"epoch:129c420d8063d931e313","name":"PrePR-CT","organization":"King Abdullah University of Science and Technology (KAUST),Karolinska Institute","date":"2024-07-24","domains":["Biology"],"country":"Saudi Arabia,Sweden","reference":"https://www.biorxiv.org/content/10.1101/2024.07.24.604816v1","confidence":"Unknown","accessibility":""},{"id":"epoch:f6477f9bef7f308889db","name":"Palmyra Fin","organization":"Writer","date":"2024-07-24","domains":["Language"],"country":"United States of America","reference":"https://writer.com/llms/palmyra-fin/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:d3a0576ff56f01eedb26","name":"Palmyra Med","organization":"Writer","date":"2024-07-24","domains":["Language","Medicine"],"country":"United States of America","reference":"https://writer.com/blog/palmyra-med-fin-models/","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:d0a864019739ca4cd295","name":"Llama 3.1-405B","organization":"Meta AI","date":"2024-07-23","domains":["Language"],"country":"United States of America","reference":"https://ai.meta.com/research/publications/the-llama-3-herd-of-models/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:f308d164737f2eb230c4","name":"Llama 3.1-70B","organization":"Meta AI","date":"2024-07-23","domains":["Language"],"country":"United States of America","reference":"https://ai.meta.com/research/publications/the-llama-3-herd-of-models/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:9e299770191e84ee79fc","name":"Llama 3.1-8B","organization":"Meta AI","date":"2024-07-23","domains":["Language"],"country":"United States of America","reference":"https://ai.meta.com/research/publications/the-llama-3-herd-of-models/","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:36e105563a60915d34a9","name":"OutfitAnyone","organization":"Alibaba,University of Science and Technology of China (USTC)","date":"2024-07-23","domains":["Image generation","Vision"],"country":"China,China","reference":"https://arxiv.org/abs/2407.16224","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:f5f364421191b151862f","name":"CarrotAI (CarrotAI\u5927\u6a21\u578b)","organization":"Jiangsu Huizhi Intelligent Digital Technology Co., Ltd.","date":"2024-07-23","domains":["Language"],"country":"China","reference":"https://huggingface.co/CarrotAI/OpenCarrot-Mix-7B\nhttps://web.archive.org/web/20250624124710/https://www.aiyzh.com/aishow/651/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9f6464d649a5f2faada3","name":"PepPrCLIP","organization":"Duke University,Cornell University,Sanford Burnham Prebys Institute","date":"2024-07-22","domains":["Biology"],"country":"United States of America,United States of America,United States of America","reference":"https://pubmed.ncbi.nlm.nih.gov/39091799/","confidence":"Speculative","accessibility":"Open weights (non-commercial)"},{"id":"epoch:6299f172dd3d0912a1bb","name":"OV-DINO","organization":"Guangzhou AI Public Computing Center,Sun Yat-sen University,Meituan Inc,Pengcheng Lab","date":"2024-07-22","domains":["Vision"],"country":"China,China,China,China","reference":"https://arxiv.org/abs/2407.07844","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:8465517ed764bc8af410","name":"DCLM 7B","organization":"Apple","date":"2024-07-20","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/apple/DCLM-7B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:798a379e22a93cb861c1","name":"Zhiwei (\u77e5\u5fae\u5927\u6a21\u578b)","organization":"Beijing Weimeng Chuangke Network Technology","date":"2024-07-20","domains":["Language"],"country":"China","reference":"https://tech.chinadaily.com.cn/a/202407/23/WS669f556da3107cd55d26c344.html","confidence":"Unknown","accessibility":""},{"id":"epoch:48ff9ffb9e1d2b253b1f","name":"Athene-70B","organization":"Nexusflow","date":"2024-07-19","domains":["Language"],"country":"United States of America","reference":"https://nexusflow.ai/blogs/athene","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:0093bfe8f7a33015cd4c","name":"Grounding Dino L","organization":"Tsinghua University,International Digital Economy Academy,Hong Kong University of Science and Technology (HKUST),Chinese University of Hong Kong (CUHK),Microsoft Research,South China University of Technology","date":"2024-07-19","domains":["Vision"],"country":"China,Hong Kong,Hong Kong,United States of America,China","reference":"https://arxiv.org/abs/2303.05499v5","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e83b4be7736108882e2f","name":"Eleven Turbo v2.5","organization":"ElevenLabs","date":"2024-07-19","domains":["Speech"],"country":"United States of America","reference":"https://elevenlabs.io/blog/introducing-turbo-v2-5","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:94b529238a70d7b11396","name":"Mistral NeMo","organization":"Mistral AI","date":"2024-07-18","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/mistral-nemo/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1ff8973d136601dd8839","name":"GPT-4o mini","organization":"OpenAI","date":"2024-07-18","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:0ab84a589bf0a6949b97","name":"PathOrchestra","organization":"","date":"2024-07-18","domains":[],"country":"","reference":"https://www.stdaily.com/index/kejixinwen/202407/ab05ac54f2ac472ea13d63c108045847.shtml (LINK BROKEN)","confidence":"Unknown","accessibility":""},{"id":"epoch:e6ff304abffdc62f1fca","name":"ChatLing (\u7075\u7280\u5927\u6a21\u578b)","organization":"Beijing 58 Information Technology","date":"2024-07-18","domains":["Language"],"country":"China","reference":"https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm","confidence":"Unknown","accessibility":""},{"id":"epoch:12602381b87952aa2307","name":"Zhizhe Qianwen (\u667a\u8005\u5343\u95ee\u5927\u8bed\u8a00\u6a21\u578b)","organization":"Beijing Zhijing Yunchuang Technology Co., Ltd.","date":"2024-07-18","domains":["Robotics","Multimodal"],"country":"China","reference":"https://web.archive.org/web/20250624064843/https://www.tianyushuke.com/xwmt/2025/2025-02-20/863.html (LINK BROKEN)","confidence":"Unknown","accessibility":""},{"id":"epoch:38373a4c0c2337c58773","name":"Beiruan (\u5317\u8f6f\u5927\u6a21\u578b)","organization":"Beijing Beida Software Engineering Co., Ltd.","date":"2024-07-18","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250624122629/https://finance.sina.com.cn/tech/roll/2025-02-18/doc-inekwqmi3993419.shtml","confidence":"Unknown","accessibility":""},{"id":"epoch:fddd0d5c1b0ee93be3ca","name":"YiYuan (\u58f9\u5143\u5927\u6a21\u578b)","organization":"SoundAI","date":"2024-07-18","domains":["Speech"],"country":"China","reference":"https://web.archive.org/web/20250624130703/https://wallstreetcn.com/articles/3692686","confidence":"Unknown","accessibility":""},{"id":"epoch:9e1903d7c77e77d8a3d3","name":"Xingchen Multimodal Model (\u4e2d\u7535\u4fe1\u4eba\u5de5\u667a\u80fd\u79d1\u6280\uff08\u5317\u4eac\uff09\u6709\u9650\u516c\u53f8)","organization":"China Telecom","date":"2024-07-18","domains":["Multimodal","Language","Vision","Image generation"],"country":"China","reference":"https://web.archive.org/web/20250707091218/https://www.sohu.com/a/735936748_100263313\n\nhttps://www.ithome.com/0/847/821.htm","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:949b11940ffed1db8b31","name":"Cygnet","organization":"Gray Swan","date":"2024-07-16","domains":["Language"],"country":"United States of America","reference":"https://www.grayswan.ai/product/cygnet\nhttps://www.grayswan.ai/news/gray-swan-launch","confidence":"Confident","accessibility":"API access"},{"id":"epoch:2148f44e9122b636e5cb","name":"Codestral Mamba","organization":"Mistral AI","date":"2024-07-16","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/codestral-mamba/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:40e40cdfd008c4435301","name":"DeepL LLM","organization":"DeepL","date":"2024-07-16","domains":["Language"],"country":"Germany","reference":"https://www.deepl.com/en/blog/next-gen-language-model","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:cb27ffa4225bbb677451","name":"Mathstral","organization":"Mistral AI","date":"2024-07-16","domains":["Mathematics","Language"],"country":"France","reference":"https://huggingface.co/mistralai/Mathstral-7B-v0.1\nhttps://mistral.ai/news/mathstral","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:95bda5d63c6522fd78b9","name":"LLaVA-NeXT-32B-Qwen","organization":"LMMs-Lab","date":"2024-07-16","domains":["Multimodal","Language"],"country":"","reference":"https://huggingface.co/lmms-lab/llava-next-qwen-32b","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:f0474b0dfae22df882cc","name":"SmolLM-1.7B","organization":"","date":"2024-07-16","domains":["Language"],"country":"","reference":"https://huggingface.co/blog/smollm","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d6dbf9b23d4639ed589f","name":"Nebula (\u661f\u4e91\u5927\u6a21\u578b)","organization":"ZTE","date":"2024-07-16","domains":["Language"],"country":"China","reference":"https://www.zte.com.cn/global/about/news/zte-launches-the-nebula-telecom-large-model-opening-a-new-paradigm-for-highly-autonomous-networks.html\nhttps://web.archive.org/web/20250707085310/https://www.zte.com.cn/china/about/magazine/zte-technologies/2025/6/2/2.html","confidence":"Confident","accessibility":""},{"id":"epoch:ee74473bb5eb09e76beb","name":"ScribblePrompt-SAM","organization":"Massachusetts Institute of Technology (MIT)","date":"2024-07-16","domains":["Vision","Biology","Medicine"],"country":"United States of America","reference":"https://arxiv.org/abs/2312.07381","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3b47d3eb77d42c636666","name":"Qwen2-Audio","organization":"Alibaba","date":"2024-07-15","domains":["Multimodal","Language","Audio"],"country":"China","reference":"https://arxiv.org/abs/2407.10759","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8608e116f163bfbeb85d","name":"OmniGenome","organization":"University of Exeter","date":"2024-07-15","domains":["Biology"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2407.11242","confidence":"Likely","accessibility":""},{"id":"epoch:24c2f0393a23a1230377","name":"PARM","organization":"Oncode Institute,UMC Utrecht,Netherlands Cancer Institute,University of Groningen,Radboud University Medical Center","date":"2024-07-15","domains":["Biology"],"country":"Netherlands,Netherlands,Netherlands,Netherlands,Netherlands","reference":"https://www.biorxiv.org/content/10.1101/2024.07.09.602649v2","confidence":"Unknown","accessibility":""},{"id":"epoch:f2eb78a08ca973543ef0","name":"InternVL2-Llama3-76B","organization":"Shanghai AI Lab","date":"2024-07-15","domains":["Multimodal","Language","Vision","Video"],"country":"China","reference":"https://huggingface.co/OpenGVLab/InternVL2-Llama3-76B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:5fd43e8e5da5dac89a0f","name":"MP4","organization":"310.ai","date":"2024-07-15","domains":["Biology"],"country":"United States of America","reference":"https://310.ai/docs/design/mp","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:a245945f8e3a2aa32c8d","name":"HelixProtX","organization":"Baidu","date":"2024-07-12","domains":["Biology"],"country":"China","reference":"https://arxiv.org/abs/2407.09274","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:da97d42f3e77db7693b7","name":"Deep learning linking mechanistic models to single-cell transcriptomics data reveals transcriptional bursting in response to DNA damage","organization":"Sun Yat-sen University,University of California Irvine,Guangdong Provincial People's Hospital,Guangdong Academy of Medical Sciences","date":"2024-07-12","domains":["Biology"],"country":"China,United States of America,China,China","reference":"https://www.biorxiv.org/content/10.1101/2024.07.10.602845v1","confidence":"Confident","accessibility":""},{"id":"epoch:68c0fdceb701d1ef8326","name":"PaliGemma","organization":"Google DeepMind","date":"2024-07-10","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2407.07726v1","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:e4f422af5da262addf68","name":"OpenDiLoCo 150M","organization":"Prime Intellect","date":"2024-07-10","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2407.07852","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2ad410f8d9dfa8dfc153","name":"OpenDiLoCo 1.1B","organization":"Prime Intellect","date":"2024-07-10","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2407.07852","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:90e0170297d6de206b3a","name":"MAP-Neo","organization":"University of Waterloo,01.AI,Wuhan University","date":"2024-07-10","domains":["Language"],"country":"Canada,China,China","reference":"https://arxiv.org/abs/2405.19327","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9cc219db09598f76c8c4","name":"SenseChat 5.5","organization":"SenseTime","date":"2024-07-06","domains":["Multimodal","Language","Vision"],"country":"Hong Kong","reference":"https://www.sensetime.com/en/news-detail/51168278?categoryId=1072","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:c99f8bce65547e4c13ed","name":"Precious3GPT","organization":"Insilico Medicine AI,Harvard Medical School","date":"2024-07-05","domains":["Biology"],"country":"Hong Kong,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.07.25.605062v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:403284f779be1059580f","name":"ColPali","organization":"Illuin Technology,Equall.ai,University Paris-Saclay,ETH Zurich","date":"2024-07-05","domains":["Vision"],"country":"France,United States of America,France,Switzerland","reference":"https://arxiv.org/abs/2407.01449","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:7ad7ad472dac5713f99e","name":"InternVL2 26B","organization":"Shanghai AI Lab","date":"2024-07-04","domains":["Multimodal","Language","Vision","Video"],"country":"China","reference":"https://internvl.github.io/blog/2024-07-02-InternVL-2.0/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:242e287e9014902f026f","name":"InternVL2-40B","organization":"Shanghai AI Lab","date":"2024-07-04","domains":["Multimodal","Language","Vision","Video"],"country":"China","reference":"https://internvl.github.io/blog/2024-07-02-InternVL-2.0/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6b85a1c45a07ce30eda4","name":"DualNetGO","organization":"Hong Kong University of Science and Technology (HKUST)","date":"2024-07-03","domains":["Biology"],"country":"Hong Kong","reference":"https://www.biorxiv.org/content/10.1101/2023.11.29.569192v2","confidence":"Confident","accessibility":""},{"id":"epoch:d58783562ac1b0c76b93","name":"Smaug-34B","organization":"Abacus AI","date":"2024-07-03","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2402.13228","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9874c2e43a8bfb79caf9","name":"Smaug-72B","organization":"Abacus AI","date":"2024-07-03","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2402.13228","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e9c196f5db0c0c542ba7","name":"Llama-SEA-LION-v2-8B-IT","organization":"AI Singapore","date":"2024-07-01","domains":["Language"],"country":"Singapore","reference":"https://huggingface.co/aisingapore/Llama-SEA-LION-v2-8B-IT","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:cde4608a22b93556a8e7","name":"RiboDiffusion","organization":"Beihang University,Nanjing University,Chinese University of Hong Kong (CUHK)","date":"2024-06-28","domains":["Biology"],"country":"China,China,Hong Kong","reference":"https://academic.oup.com/bioinformatics/article/40/Supplement_1/i347/7700903","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7cc806837331e0f4bb23","name":"ChatBit","organization":"Beijing Institute of Technology,Academy of Military Science,Minzu University of China","date":"2024-06-28","domains":["Language"],"country":"China,China,China","reference":"https://d.wanfangdata.com.cn/periodical/gfkj202403008","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:4161f480f0a6eaff9bad","name":"Ernie 4.0 Turbo","organization":"Baidu","date":"2024-06-28","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://www.reuters.com/technology/artificial-intelligence/baidu-launches-upgraded-ai-model-says-user-base-hits-300-mln-2024-06-28/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:c70c731c1ef4aaf9b131","name":"CriticGPT","organization":"OpenAI","date":"2024-06-27","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2407.00215","confidence":"Unknown","accessibility":""},{"id":"epoch:ef9dee46e2ff5db168d4","name":"Index-1.9B","organization":"Shanghai Kuanyu Digital Technology Co., Ltd. (Bilibili)","date":"2024-06-27","domains":["Language"],"country":"China","reference":"https://huggingface.co/IndexTeam/Index-1.9B\nhttps://web.archive.org/web/20250710105520/https://www.bilibili.com/opus/942741894140526673","confidence":"Confident","accessibility":""},{"id":"epoch:f25681ca527ebaaff4ac","name":"Molecular Diffusion Models with Virtual Receptors","organization":"Verily Research","date":"2024-06-26","domains":["Biology"],"country":"United States of America","reference":"https://arxiv.org/abs/2406.18330","confidence":"Confident","accessibility":""},{"id":"epoch:06e9981815aa5ca73298","name":"ESM3 (98B)","organization":"EvolutionaryScale,University of California (UC) Berkeley","date":"2024-06-25","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://www.evolutionaryscale.ai/blog/esm3-release ","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e824bd79778c04a84fff","name":"ESM3-open-small","organization":"EvolutionaryScale,University of California (UC) Berkeley","date":"2024-06-25","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://www.evolutionaryscale.ai/blog/esm3-release ","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:eb7009ffb5a3953d2274","name":"Flexi-JEST++","organization":"Google DeepMind","date":"2024-06-25","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2406.17711v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:5f6e6bc88e93bce257c7","name":"JEST++","organization":"Google DeepMind","date":"2024-06-25","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2406.17711v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:5317c868d9c8b76a7adb","name":"JEST-L++","organization":"DeepMind","date":"2024-06-25","domains":["Vision"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2406.17711v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4115ee4b3408de347f9e","name":"Gemma 2 9B","organization":"Google DeepMind","date":"2024-06-24","domains":["Language"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/gemma/gemma-2-report.pdf","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:a879844bbcb73ece0b37","name":"Gemma 2 27B","organization":"Google DeepMind","date":"2024-06-24","domains":["Language"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/gemma/gemma-2-report.pdf","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:67eebbd79742b955fb28","name":"Gemma 2 2B","organization":"Google DeepMind","date":"2024-06-24","domains":["Language"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/gemma/gemma-2-report.pdf","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:013b269e8588b269ee69","name":"DiffPALM","organization":"Ecole Polytechnique F\u00b4ed\u00b4erale de Lausanne (EPFL),Swiss Institute of Bioinformatics","date":"2024-06-24","domains":["Biology"],"country":"Switzerland,Switzerland","reference":"https://www.pnas.org/doi/abs/10.1073/pnas.2311887121","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:d88a799ca43e5fece941","name":"BADGER","organization":"Nvidia,University of California (UC) Berkeley","date":"2024-06-24","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2406.16821","confidence":"Likely","accessibility":""},{"id":"epoch:cf48e846282bfa0bf7d7","name":"Cambrian-1-34B","organization":"New York University (NYU)","date":"2024-06-24","domains":["Multimodal","Vision","Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2406.16860","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:dfbdfd6147a0a6346033","name":"Cambrian-1-13B","organization":"New York University (NYU)","date":"2024-06-24","domains":["Multimodal","Vision","Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2406.16860","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:797fcffdf8fc96881232","name":"Cambrian-1-8B","organization":"New York University (NYU)","date":"2024-06-24","domains":["Multimodal","Vision","Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2406.16860","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:248a9767e7a7dad34502","name":"Code Droid","organization":"Factory","date":"2024-06-22","domains":["Language"],"country":"United States of America","reference":"https://www.factory.ai/news/code-droid-technical-report","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:f4e09e2ef0a82f616156","name":"Pangu 5.0","organization":"","date":"2024-06-21","domains":[],"country":"","reference":"https://www.huawei.com/en/media-center/transform/15-5/04-building-large-ai-models-best-suited-to-industry-needs","confidence":"Unknown","accessibility":""},{"id":"epoch:38c172e384f0f7e9c387","name":"Claude 3.5 Sonnet","organization":"Anthropic","date":"2024-06-20","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://www-cdn.anthropic.com/fed9cc193a14b84131812372d8d5857f8f304c52/Model_Card_Claude_3_Addendum.pdf","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:69113460452a40758ce8","name":"Hermes 2 Theta Llama-3 70B","organization":"Nous Research,Arcee AI","date":"2024-06-20","domains":["Language"],"country":"United States of America,United States of America","reference":"https://huggingface.co/NousResearch/Hermes-2-Theta-Llama-3-70B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:37d97a65b1a32866c108","name":"RNA-FrameFlow","organization":"National University of Singapore,Prescient Design,University of Missouri,University of Cambridge","date":"2024-06-19","domains":["Biology"],"country":"Singapore,United States of America,United States of America,United Kingdom","reference":"https://arxiv.org/abs/2406.13839","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:41572a2b799a5e453f8e","name":"MPNNsol","organization":"Ecole Polytechnique F\u00b4ed\u00b4erale de Lausanne (EPFL),University at Buffalo,University of Washington,Massachusetts Institute of Technology (MIT)","date":"2024-06-19","domains":["Biology"],"country":"Switzerland,United States of America,United States of America,United States of America","reference":"https://www.nature.com/articles/s41586-024-07601-y#Sec30","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0c8ac953acd51bc62f5b","name":"GLM-4V-9B","organization":"Z.ai (Zhipu AI),Tsinghua University","date":"2024-06-18","domains":["Language","Multimodal","Vision"],"country":"China,China","reference":"https://arxiv.org/abs/2406.12793","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:c6baaf34f247982d6a21","name":"DeepSeek-Coder-V2 236B","organization":"DeepSeek","date":"2024-06-17","domains":["Language"],"country":"China","reference":"https://github.com/deepseek-ai/DeepSeek-Coder-V2","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:538340948a5c056e0ec9","name":"Gen-3 Alpha","organization":"Runway","date":"2024-06-17","domains":["Video","Vision"],"country":"United States of America","reference":"https://runwayml.com/research/introducing-gen-3-alpha\n\nhttps://help.runwayml.com/hc/en-us/articles/30266515017875-Creating-with-Text-Image-to-Video-on-Gen-3-Alpha-and-Turbo","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:129eec9c479f29a37420","name":"Ovis-7B","organization":"Alibaba,Nanjing University","date":"2024-06-17","domains":["Multimodal","Language","Vision"],"country":"China,China","reference":"https://arxiv.org/abs/2405.20797","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5fc35e13d4dc6405252f","name":"PRISM-1","organization":"Wayve","date":"2024-06-17","domains":["Image generation","Video","3D modeling"],"country":"United Kingdom","reference":"https://wayve.ai/thinking/prism-1/","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:3be6dd1ddb57ef4736f4","name":"MindEye2","organization":"Stability AI,Medical AI Research Center (MedARC),Princeton University,University of Minnesota,University of Sydney,University of Waterloo","date":"2024-06-15","domains":["Medicine","Vision","Image generation"],"country":"United Kingdom,United States of America,United States of America,United States of America,Australia,Canada","reference":"https://arxiv.org/abs/2403.11207","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:40087617b34e7bf99226","name":"JIUTIAN-139MoE","organization":"China Mobile","date":"2024-06-15","domains":["Language"],"country":"China","reference":"https://gitee.com/CMCC-jiutian/JIUTIAN-139MoE?skip_mobile=true","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5561b57a82d844471b14","name":"Nemotron-4 340B","organization":"Nvidia","date":"2024-06-14","domains":["Language"],"country":"United States of America","reference":"https://blogs.nvidia.com/blog/nemotron-4-synthetic-data-generation-llm-training/ ","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:530a5c15c8c3d55d43e0","name":"PLaMo-100B","organization":"Preferred Networks Inc","date":"2024-06-14","domains":["Language"],"country":"Japan","reference":"https://tech.preferred.jp/ja/blog/plamo-100b/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:54712f3cd3f2e280e9b6","name":"DigiRL","organization":"University of California (UC) Berkeley,University of Illinois Urbana-Champaign (UIUC),Google DeepMind","date":"2024-06-14","domains":["Vision","Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2406.11896","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9f392e3e736f95211984","name":"OpenVLA","organization":"Stanford University,University of California (UC) Berkeley,Toyota Research Institute,Google DeepMind,Massachusetts Institute of Technology (MIT),Physical Intelligence","date":"2024-06-13","domains":["Robotics","Vision","Language"],"country":"United States of America,United States of America,United States of America,United States of America,United States of America,United States of America","reference":"https://openvla.github.io/\nhttps://arxiv.org/abs/2406.09246","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7c1a53fd131e6e7d22a5","name":"Animate Anyone","organization":"Alibaba","date":"2024-06-13","domains":["Video"],"country":"China","reference":"https://arxiv.org/abs/2311.17117","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:0328b6fb4ba2e8f6c88e","name":"Mamba2-Hybrid","organization":"Nvidia","date":"2024-06-12","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2406.07887","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5034043db104e9717196","name":"ProteinReDiff","organization":"FPT Software AI Center,University of Chicago,Indiana State University","date":"2024-06-12","domains":["Biology"],"country":"Vietnam,United States of America,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.04.17.589997v3.abstract","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0a994c5a92980fb94d40","name":"Stable Diffusion 3 Medium","organization":"Stability AI","date":"2024-06-12","domains":["Image generation"],"country":"United Kingdom","reference":"https://stability.ai/news/stable-diffusion-3-medium\nhttps://huggingface.co/stabilityai/stable-diffusion-3-medium","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:983574214a4356a2fbea","name":"Luma Dream Machine","organization":"LumaLabs","date":"2024-06-12","domains":["Video","Vision"],"country":"United States of America","reference":"https://lumalabs.ai/dream-machine","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:a44521f3282f07595312","name":"Shutterstock ImageAI","organization":"Databricks","date":"2024-06-12","domains":["Image generation"],"country":"United States of America","reference":"https://www.databricks.com/company/newsroom/press-releases/introducing-shutterstock-imageai-powered-databricks-image","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:b2802f66e0be7a875780","name":"Megrez-3B-Omni","organization":"Infinigence AI,Tsinghua University,Shanghai Jiao Tong University","date":"2024-06-12","domains":["Multimodal","Language","Vision","Speech"],"country":"China,China,China","reference":"https://github.com/infinigence/Infini-Megrez-Omni/blob/main/assets/Megrez_Omni_Technical_Report.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:21d6a03c76e6c6352f2e","name":"Phoenix 1.0 Ultra","organization":"Leonardo AI","date":"2024-06-12","domains":["Image generation"],"country":"Australia","reference":"https://docs.leonardo.ai/docs/generate-images-using-leonardo-phoenix-model","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:1f6e0d0e31a614e74d08","name":"Samba 3.8B","organization":"Microsoft,University of Illinois Urbana-Champaign (UIUC)","date":"2024-06-11","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2406.07522","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4e8aa0d9c7f69a4e78a6","name":"TiTok-L","organization":"ByteDance,Technical University of Munich","date":"2024-06-11","domains":["Image generation"],"country":"China,Germany","reference":"https://arxiv.org/abs/2406.07550","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9f9adb8121b00dabd8f2","name":"YiSu","organization":"","date":"2024-06-11","domains":[],"country":"","reference":"https://www.stdaily.com/index/kejixinwen/202406/bdac079567b240708883899b883433d0.shtml (LINK BROKEN)","confidence":"Unknown","accessibility":""},{"id":"epoch:a6e33187bb6f282cd797","name":"Kling","organization":"Kuaishou Technology","date":"2024-06-10","domains":["Video","Vision"],"country":"China","reference":"https://klingai.io/\nhttps://ir.kuaishou.com/news-releases/news-release-details/kuaishou-unveils-comprehensive-ai-models-reshaping-content-and","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:9e2268b103d2c804cb77","name":"Qwen2-72B","organization":"Alibaba","date":"2024-06-07","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2/ \nhttps://arxiv.org/abs/2407.10671 ","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2b92ee16180462a83967","name":"Qwen2-7B","organization":"Alibaba","date":"2024-06-07","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2/ \nhttps://arxiv.org/abs/2407.10671 ","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3268f05e4cd25c85ec7a","name":"Qwen2-57B-A14B","organization":"Alibaba","date":"2024-06-07","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2/ \nhttps://arxiv.org/abs/2407.10671 ","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:33642105678947520beb","name":"Qwen2-1.5B","organization":"Alibaba","date":"2024-06-07","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2/ \nhttps://arxiv.org/abs/2407.10671 ","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c08aafaa690c142ed67d","name":"Qwen2-0.5B","organization":"Alibaba","date":"2024-06-07","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen2/ \nhttps://arxiv.org/abs/2407.10671 ","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5275c7ed669ae1dbc62b","name":"Audioseal","organization":"Facebook AI Research","date":"2024-06-06","domains":["Audio","Speech"],"country":"United States of America,France","reference":"https://arxiv.org/abs/2401.17264","confidence":"Confident","accessibility":""},{"id":"epoch:45951890b1480a7acc85","name":"ProTrek","organization":"Westlake University","date":"2024-06-03","domains":["Biology"],"country":"China","reference":"https://www.biorxiv.org/content/10.1101/2024.05.30.596740v2.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:4ff67455ff8ca364a4a0","name":"Prot2Token","organization":"University of Missouri,Politecnico di Milano","date":"2024-06-03","domains":["Biology"],"country":"United States of America,Italy","reference":"https://www.biorxiv.org/content/10.1101/2024.05.31.596915v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:8bf55fe38e501adbb73d","name":"MiniCPM-2.4B","organization":"Tsinghua University,ModelBest","date":"2024-06-03","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2404.06395","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:03ba73083c16dd0c008e","name":"MiniCPM-3-4B","organization":"Tsinghua University,ModelBest","date":"2024-06-03","domains":["Language"],"country":"China,China","reference":"https://huggingface.co/openbmb/MiniCPM3-4B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:27c827ca771331a0059a","name":"MiniCPM-1.2B","organization":"Tsinghua University,ModelBest","date":"2024-06-03","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2404.06395","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:6b66973f0e07d9fe7225","name":"MULAN","organization":"AIRI Artificial Intelligence Research Institute,Skolkovo Institute of Science and Technology,Belozersky Institute of Physio-Chemical Biology,Ligand Pro","date":"2024-06-02","domains":["Biology"],"country":"Russia,Russia,Russia,Russia","reference":"https://www.biorxiv.org/content/10.1101/2024.05.30.596565v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5edf297098b9df41f955","name":"DRGN-AI","organization":"Stanford University,SLAC National Laboratory,Princeton University,Columbia University","date":"2024-06-02","domains":["Biology"],"country":"United States of America,United States of America,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.05.30.596729v1","confidence":"Confident","accessibility":""},{"id":"epoch:100d46dce9458e5e71a4","name":"Xingchen Voice Model (\u661f\u8fb0\u8bed\u97f3\u5927\u6a21\u578b)","organization":"China Telecom","date":"2024-06-01","domains":["Speech"],"country":"China","reference":"https://www.chinadaily.com.cn/a/202406/06/WS66611799a31082fc043cb30a.html","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:14966091e08dfa23e5c1","name":"KeTu (Kolors)","organization":"Kuaishou Technology","date":"2024-05-31","domains":["Image generation"],"country":"China","reference":"https://github.com/Kwai-Kolors/Kolors/blob/master/imgs/Kolors_paper.pdf\n\nhttps://ir.kuaishou.com/news-releases/news-release-details/kuaishou-unveils-comprehensive-ai-models-reshaping-content-and","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:c9ac5cd6e3ceba89e724","name":"Mamba 2, 2.7B","organization":"Princeton University,Carnegie Mellon University (CMU)","date":"2024-05-31","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2405.21060","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e9f7c978022374661a45","name":"Granite 20B","organization":"IBM Research","date":"2024-05-31","domains":["Language"],"country":"United States of America","reference":"https://www.ibm.com/downloads/documents/us-en/10a99803c92fdb35","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:293f3b572f232d12a719","name":"sgRNAGen","organization":"Beijing Institute of Technology","date":"2024-05-31","domains":["Biology"],"country":"China","reference":"https://www.biorxiv.org/content/10.1101/2024.05.30.596019v1","confidence":"Confident","accessibility":""},{"id":"epoch:2dea2c964a271c6527af","name":"Ark LLM (\u65b9\u821f\u5927\u6a21\u578b)","organization":"Zhuhai Wujiefangzhou Intelligent Technology","date":"2024-05-31","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250624082004/https://gzdaily.dayoo.com/h5/html5/2024-06/03/content_875_859819.htm","confidence":"Confident","accessibility":""},{"id":"epoch:1344ef0634478db5b3af","name":"FoldFlow2","organization":"Dreamfold,University of Montreal / Universit\u00e9 de Montr\u00e9al,McGill University,University of Oxford","date":"2024-05-30","domains":["Biology"],"country":"Canada,Canada,Canada,United Kingdom","reference":"https://arxiv.org/abs/2405.20313","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:86cea1d2f4e491505cd7","name":"CLAY","organization":"Shanghai Tech University,Deemos Technology,Huazhong University of Science and Technology","date":"2024-05-30","domains":["3D modeling","Vision"],"country":"China,China,China","reference":"https://arxiv.org/abs/2406.13897v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4d87bd15040f2ee2f06f","name":"Codestral","organization":"Mistral AI","date":"2024-05-29","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/codestral/","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:a501b2f9bbecfbf071ed","name":"ChunkLlama2-13B","organization":"Alibaba,The University of Hong Kong,Fudan University","date":"2024-05-29","domains":["Language"],"country":"China,Hong Kong,China","reference":"https://arxiv.org/abs/2402.17463","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:24aacfb5486a3b0bb6d6","name":"Aurora","organization":"Microsoft Research","date":"2024-05-28","domains":["Earth science"],"country":"United States of America","reference":"https://arxiv.org/abs/2405.13063","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:5172362981ecfc6ace20","name":"Nanbeige2-16B-Chat","organization":"Nanbeige LLM Lab","date":"2024-05-28","domains":["Language"],"country":"China","reference":"https://huggingface.co/Nanbeige/Nanbeige2-16B-Chat","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:686724efba8f77215682","name":"TimeGPT-1","organization":"Nixtla","date":"2024-05-27","domains":["Other"],"country":"United States of America","reference":"https://arxiv.org/abs/2310.03589","confidence":"Confident","accessibility":""},{"id":"epoch:0b9a18593f01845eecbe","name":"NV-Embed-v1","organization":"Nvidia","date":"2024-05-27","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2405.17428","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:60e10b7e6b997b9005a0","name":"LUXIA-21.4B-Alignment","organization":"SaltLux","date":"2024-05-27","domains":["Language"],"country":"South Korea","reference":"https://huggingface.co/saltlux/luxia-21.4b-alignment-v1.2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b4e6463c49a07cf64de6","name":"Zamba2-7B","organization":"Zyphra","date":"2024-05-26","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2405.16712","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:be93db6c037addcf2354","name":"Llama-3-70B-Synthia-v3.5","organization":"","date":"2024-05-26","domains":["Language"],"country":"","reference":"https://huggingface.co/migtissera/Llama-3-70B-Synthia-v3.5","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:bc990d5ddbe31343b924","name":"Xingchen Jianwei Security Model (\u661f\u8fb0\u00b7\u89c1\u5fae\u5b89\u5168\u5927 \u6a21\u578b)","organization":"China Telecom","date":"2024-05-25","domains":["Language","Cybersecurity"],"country":"China","reference":"https://web.archive.org/web/20250707074238/https://baby.ifeng.com/c/8jLTtwCy4WS\nhttps://web.archive.org/web/20250707074510/https://finance.sina.cn/2024-05-27/detail-inawrsqu7093303.d.html","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2161666627f3cbb3ee2e","name":"Genie 2 (bio)","organization":"Columbia University,Rutgers University","date":"2024-05-24","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2405.15489","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:639c2fd452f46bbb45c0","name":"LLPS","organization":"InstaDeep","date":"2024-05-24","domains":["Biology"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2405.15840","confidence":"Confident","accessibility":""},{"id":"epoch:9d7579ceaddc6da9e630","name":"UniGPT-mMed","organization":"","date":"2024-05-24","domains":[],"country":"","reference":"https://www.163.com/dy/article/IDFC8PJJ0518C51R.html","confidence":"Unknown","accessibility":""},{"id":"epoch:f89279ea82f774aabd19","name":"OMNI-EPIC","organization":"Imperial College London,University of British Columbia (UBC)","date":"2024-05-24","domains":["Language"],"country":"United Kingdom,Canada","reference":"https://arxiv.org/abs/2405.15568","confidence":"Confident","accessibility":""},{"id":"epoch:48f2101fc827e01a5167","name":"YOLOv10-X","organization":"Tsinghua University","date":"2024-05-23","domains":["Vision"],"country":"China","reference":"https://arxiv.org/abs/2405.14458","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6621c3ecaca8d45b75e1","name":"Baichuan4","organization":"Baichuan","date":"2024-05-22","domains":[],"country":"China","reference":"https://zhidx.com/p/426526.html","confidence":"Unknown","accessibility":""},{"id":"epoch:a3663118e97600ccce82","name":"360Zhinao-7B","organization":"360 Security Technology","date":"2024-05-22","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2405.13386","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:32d5860d65aa706ae301","name":"ProtT3","organization":"National University of Singapore,University of Science and Technology of China (USTC),Hokkaido University","date":"2024-05-21","domains":["Biology"],"country":"Singapore,China,Japan","reference":"https://arxiv.org/abs/2405.12564","confidence":"Likely","accessibility":""},{"id":"epoch:1c38d62490e334e1ac42","name":"ALLaM adapted13B","organization":"Saudi Data and Artificial Intelligence Authority","date":"2024-05-21","domains":["Language"],"country":"Saudi Arabia","reference":"https://arxiv.org/abs/2407.15390\nhttps://huggingface.co/ALLaM-AI/ALLaM-7B-Instruct-previehttps://www.middleeastainews.com/p/sdaias-allam-arabic-llm-live-on-watsonxw","confidence":"Confident","accessibility":"API access"},{"id":"epoch:afe652b54e155cd4e6aa","name":"ALLaM\u00a0adapted 70B","organization":"Saudi Data and Artificial Intelligence Authority","date":"2024-05-21","domains":["Language"],"country":"Saudi Arabia","reference":"https://arxiv.org/abs/2407.15390","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:962de66904b2b69234b4","name":"ALLaM 7B","organization":"Saudi Data and Artificial Intelligence Authority","date":"2024-05-21","domains":["Language"],"country":"Saudi Arabia","reference":"https://arxiv.org/abs/2407.15390","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:20a8bfa68c133071778c","name":"ALLaM 34B","organization":"Saudi Data and Artificial Intelligence Authority","date":"2024-05-21","domains":["Language"],"country":"Saudi Arabia","reference":"https://arxiv.org/abs/2407.15390","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:46aeb4ddde60386b1ea9","name":"GLM-4 (0520)","organization":"Z.ai (Zhipu AI)","date":"2024-05-20","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2406.12793","confidence":"Likely","accessibility":"API access"},{"id":"epoch:416a71ad18b22253c962","name":"Diamond","organization":"University of Geneva,University of Edinburgh,Microsoft Research","date":"2024-05-20","domains":["Games"],"country":"Switzerland,United Kingdom,United States of America","reference":"https://arxiv.org/abs/2405.12399","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:51fb11ed966e521cd916","name":"Octo-Base","organization":"University of California (UC) Berkeley,Stanford University,Carnegie Mellon University (CMU),DeepMind","date":"2024-05-20","domains":["Robotics"],"country":"United States of America,United States of America,United States of America,United Kingdom","reference":"https://arxiv.org/abs/2405.12213 ","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8ccd94f195afc962371d","name":"Octo-Small","organization":"University of California (UC) Berkeley,Stanford University,Carnegie Mellon University (CMU),DeepMind","date":"2024-05-20","domains":["Robotics"],"country":"United States of America,United States of America,United States of America,United Kingdom","reference":"https://arxiv.org/abs/2405.12213 ","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:15eb6fbb061df8ba1f12","name":"HelixFold","organization":"Baidu","date":"2024-05-17","domains":["Biology"],"country":"China","reference":"https://arxiv.org/abs/2404.10260v2","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:da19134bf1c2b7ff17cf","name":"ProSST","organization":"Shanghai Jiao Tong University,Shanghai AI Lab,East China University of Science and Technology","date":"2024-05-17","domains":["Biology"],"country":"China,China,China","reference":"https://www.biorxiv.org/content/10.1101/2024.04.15.589672v3.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:2b145991182e463413ca","name":"Chameleon-34B","organization":"Facebook AI Research","date":"2024-05-16","domains":["Multimodal","Image generation","Language","Vision"],"country":"United States of America,France","reference":"https://arxiv.org/abs/2405.09818v1","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:eb2ac1b6a5f622db31c8","name":"FragLlama: Next-fragment prediction for molecular design","organization":"Facebook AI Research","date":"2024-05-16","domains":["Multimodal","Image generation","Vision","Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/2405.09818v1","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:90a08227dc488f1d8070","name":"LBSTER","organization":"Prescient Design,Genentech","date":"2024-05-15","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.05.14.594108v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:9e18e6f8cf034ce35cac","name":"Doubao-lite","organization":"ByteDance","date":"2024-05-15","domains":["Language"],"country":"China","reference":"https://www.volcengine.com/docs/6360/1264663","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:a6980d18f2ce0f4878d6","name":"Doubao Role-Playing Model","organization":"ByteDance","date":"2024-05-15","domains":["Language"],"country":"China","reference":"https://www.volcengine.com/docs/6360/1264663","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:e0d526cc73bd38a19dd5","name":"Doubao Text-to-Speech Model","organization":"ByteDance","date":"2024-05-15","domains":["Speech"],"country":"China","reference":"https://www.volcengine.com/docs/6360/1264663","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:5c929a44dd27a3351a99","name":"Doubao Voice Cloning Model","organization":"ByteDance","date":"2024-05-15","domains":["Speech"],"country":"China","reference":"https://www.volcengine.com/docs/6360/1264663","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:dcd0db1b400c248a8d5c","name":"Doubao Speech Recognition Model","organization":"ByteDance","date":"2024-05-15","domains":["Speech"],"country":"China","reference":"https://www.volcengine.com/docs/6360/1264663","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:ee16f0f1714129ff9cd3","name":"Doubao Text-to-Image Model","organization":"ByteDance","date":"2024-05-15","domains":["Image generation"],"country":"China","reference":"https://www.volcengine.com/docs/6360/1264663","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:b0f547e4bb315ced374b","name":"Doubao Function Call Model","organization":"ByteDance","date":"2024-05-15","domains":["Language"],"country":"China","reference":"https://www.volcengine.com/docs/6360/1264663","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:6e32e6182c9fe238c96a","name":"Doubao Vectorization Model","organization":"ByteDance","date":"2024-05-15","domains":["Language"],"country":"China","reference":"https://www.volcengine.com/docs/6360/1264663","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:69bd578e05bd95811ed4","name":"Doubao Image-to-Image generation model","organization":"ByteDance","date":"2024-05-15","domains":["Image generation","Vision"],"country":"China","reference":"https://www.volcengine.com/product/doubao","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:a97f652da1f9691e3136","name":"Doubao Real-Time translation model","organization":"ByteDance","date":"2024-05-15","domains":["Language","Speech"],"country":"China","reference":"https://www.volcengine.com/product/doubao","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:6b6744969b696c71f0a6","name":"Doubao Video Generation model","organization":"ByteDance","date":"2024-05-15","domains":["Video"],"country":"China","reference":"https://www.volcengine.com/product/doubao","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:e6ca428bc55ed4190812","name":"VILA1.5-40B","organization":"Nvidia,Massachusetts Institute of Technology (MIT)","date":"2024-05-15","domains":["Multimodal","Language","Vision","Video"],"country":"United States of America,United States of America","reference":"https://huggingface.co/Efficient-Large-Model/VILA1.5-40b","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:9aa991cd14840353064c","name":"VILA-7B","organization":"Nvidia,Massachusetts Institute of Technology (MIT)","date":"2024-05-15","domains":["Multimodal","Language","Vision","Video"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2312.07533","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0b01a1c506ad729994bb","name":"Veo","organization":"Google DeepMind","date":"2024-05-14","domains":["Video","Vision"],"country":"United States of America","reference":"https://deepmind.google/technologies/veo/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:01aafdb9418f59959d2c","name":"Imagen 3","organization":"Google DeepMind","date":"2024-05-14","domains":["Image generation"],"country":"United States of America","reference":"https://deepmind.google/technologies/imagen-3/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:ad275cbb14e391aa4e44","name":"LearnLM-Tutor","organization":"Google Research,Google DeepMind,Google,Arizona State University,Lund University,University of Oxford","date":"2024-05-14","domains":["Language"],"country":"United States of America,United States of America,United States of America,United States of America,Sweden,United Kingdom","reference":"https://storage.googleapis.com/deepmind-media/LearnLM/LearnLM_paper.pdf","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:d75cc7f6804655a1ced2","name":"Hunyuan-DiT","organization":"Tencent","date":"2024-05-14","domains":["Image generation"],"country":"China","reference":"https://arxiv.org/abs/2405.08748","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:6a6baf6f5e80dc4fde33","name":"BRIA 2.3","organization":"BRIA AI","date":"2024-05-14","domains":["Image generation"],"country":"United States of America","reference":"https://huggingface.co/briaai/BRIA-2.3","confidence":"Unknown","accessibility":"Open weights (non-commercial)"},{"id":"epoch:67a77e65807451753432","name":"Yi-Large","organization":"01.AI","date":"2024-05-13","domains":["Language"],"country":"China","reference":"","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:ba8e12e81b14364c8486","name":"GPT-4o","organization":"OpenAI","date":"2024-05-13","domains":["Multimodal","Language","Audio","Speech","Vision"],"country":"United States of America","reference":"https://openai.com/index/hello-gpt-4o/ \nhttps://openai.com/index/gpt-4o-system-card/","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:0b7ac999e3e78f32e6c5","name":"Yi-1.5-34B","organization":"01.AI","date":"2024-05-13","domains":["Language"],"country":"China","reference":"https://huggingface.co/01-ai/Yi-1.5-34B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:a9529e18ed91ab61ad1c","name":"Yi-1.5-9B","organization":"01.AI","date":"2024-05-13","domains":["Language"],"country":"China","reference":"https://huggingface.co/01-ai/Yi-1.5-9B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:303710727d08168f36b6","name":"Digivio (\u8fea\u667a\u4f1f\u5965DIGIVIO)","organization":"Shanghai Digivio Information Technology Co., Ltd.","date":"2024-05-13","domains":["Language"],"country":"China","reference":"","confidence":"Unknown","accessibility":""},{"id":"epoch:9a65df9b6297f73be396","name":"Yuanshi (\u5143\u77f3\u5927\u6a21\u578b)","organization":"Beijing Yuanshi Technology Co., Ltd.","date":"2024-05-13","domains":["Language"],"country":"China","reference":"","confidence":"Unknown","accessibility":""},{"id":"epoch:6138b5a6bec03a7691c8","name":"Xiaoyu Q&A","organization":"Xiaoyu Intelligence Information Technology (Yunnan) Co., Ltd","date":"2024-05-13","domains":["Language","Audio","Speech"],"country":"China","reference":"https://xiaoyuzhineng.com/","confidence":"Unknown","accessibility":""},{"id":"epoch:e76b2dc0dbb71dfcbd98","name":"Xinyuan (\u5fc3\u5143\u5927\u6a21\u578b)","organization":"Beijing Lituo Feiyuan Technology Co., Ltd.,Cylingo Group","date":"2024-05-13","domains":["Psychology","Language"],"country":"China,China","reference":"https://web.archive.org/web/20250624131015/https://www.sohu.com/a/892391591_121284943","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:688b2f6fbde9dfb40f19","name":"Lantu (\u84dd\u56fe\u5927\u6a21\u578b)","organization":"Beijing Bitauto Interactive Advertising Company Limited","date":"2024-05-13","domains":[],"country":"China","reference":"","confidence":"Unknown","accessibility":""},{"id":"epoch:d2819ee4421e21cc71ff","name":"llama-3-airoboros-70b-3.3","organization":"","date":"2024-05-12","domains":["Language"],"country":"","reference":"https://huggingface.co/jondurbin/airoboros-70b-3.3","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:c1e14dcad20d9d1fb003","name":"MoLeR","organization":"Microsoft Research,Novartis","date":"2024-05-12","domains":["Biology"],"country":"United States of America,Switzerland","reference":"https://arxiv.org/abs/2103.03864","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2629d02abd9d444960ca","name":"Fugaku-LLM","organization":"Tohoku University,CyberAgent,Tokyo Institute of Technology,Fujitsu,RIKEN,Nagoya University,Kotoba Technologies","date":"2024-05-10","domains":["Language"],"country":"Japan,Japan,Japan,Japan,Japan,Japan,Japan","reference":"https://www.fujitsu.com/global/about/resources/news/press-releases/2024/0510-01.html","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b3305bc29d9b3e878b78","name":"Gemini 1.5 Flash","organization":"Google DeepMind","date":"2024-05-10","domains":["Multimodal","Language","Vision","Audio"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/gemini/gemini_v1_5_report.pdf","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:36064600213ea5f84a57","name":"Gemini 1.5 Flash 8B","organization":"Google DeepMind","date":"2024-05-10","domains":["Multimodal","Language","Vision","Audio"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/gemini/gemini_v1_5_report.pdf","confidence":"Confident","accessibility":"API access"},{"id":"epoch:17dfa2d65dc460a29225","name":"MatterSim (M3GNet - MatterSim-v1.0.0-5M)","organization":"Microsoft Research AI for Science","date":"2024-05-10","domains":["Materials science"],"country":"United States of America","reference":"https://arxiv.org/abs/2405.04967","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a106c8c5c6accdd7286f","name":"MatterSim (Grpaphomer)","organization":"Microsoft Research AI for Science","date":"2024-05-10","domains":["Materials science"],"country":"United States of America","reference":"https://arxiv.org/abs/2405.04967","confidence":"Speculative","accessibility":""},{"id":"epoch:f38990d602e256866130","name":"Falcon 2 11B","organization":"Technology Innovation Institute","date":"2024-05-09","domains":["Language"],"country":"United Arab Emirates","reference":"https://huggingface.co/tiiuae/falcon-11B ","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:f699f63554d101f5c804","name":"XPT \u201cXiao Model\u201d (\u6653\u6a21\u578bXPT)","organization":"Chengdu Xiaoduo Technology Co., Ltd.","date":"2024-05-09","domains":[],"country":"China","reference":"https://web.archive.org/web/20250624091917/https://36kr.com/p/2789769142715269","confidence":"Unknown","accessibility":""},{"id":"epoch:f28451d126a7d1faafaa","name":"AlphaFold 3","organization":"Google DeepMind,Isomorphic Labs","date":"2024-05-08","domains":["Biology"],"country":"United States of America,United Kingdom","reference":"https://www.nature.com/articles/s41586-024-07487-w","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:57ce05ace92771b31735","name":"BiosimDock","organization":"DeepOrigin","date":"2024-05-08","domains":["Medicine","Biology"],"country":"United States of America","reference":"https://www.deeporigin.com/blog/we-spill-the-beans-deep-origins-ai-and-physics-based-models-for-drug-discovery","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:61221de5286ff36d7318","name":"Emu2","organization":"Beijing Academy of Artificial Intelligence / BAAI,Tsinghua University,Peking University","date":"2024-05-08","domains":["Multimodal","Language","Vision","Image generation"],"country":"China,China,China","reference":"https://arxiv.org/abs/2312.13286","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:3adbe9eef9c755279e8a","name":"DeepSeek-V2 (MoE-236B)","organization":"DeepSeek","date":"2024-05-07","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2405.04434 \nhttps://github.com/deepseek-ai/DeepSeek-V2 ","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:8ad4a0740a999fb9d958","name":"xLSTM 1.4B","organization":"Johannes Kepler University Linz","date":"2024-05-07","domains":["Language"],"country":"Austria","reference":"https://arxiv.org/abs/2405.04517","confidence":"Confident","accessibility":""},{"id":"epoch:60c1458966844392a4db","name":"Amazon Titan Text Premier","organization":"Amazon","date":"2024-05-07","domains":["Language"],"country":"United States of America","reference":"https://aws.amazon.com/ru/about-aws/whats-new/2024/05/amazon-titan-text-premier-amazon-bedrock/\n\nhttps://aws.amazon.com/blogs/aws/build-rag-and-agent-based-generative-ai-applications-with-new-amazon-titan-text-premier-model-available-in-amazon-bedrock/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:f29aadff65291d233633","name":"Med-Gemini-2D","organization":"Google DeepMind,Google Research","date":"2024-05-06","domains":["Medicine","Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2405.03162","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:0a97efc09ba4709ad6c2","name":"Med-Gemini-3D","organization":"Google DeepMind,Google Research","date":"2024-05-06","domains":["Medicine","Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2405.03162","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:009268648039e2427396","name":"Microsoft MAI-1 (2024 unreleased)","organization":"Microsoft","date":"2024-05-06","domains":["Language"],"country":"United States of America","reference":"https://www.theinformation.com/articles/meet-mai-1-microsoft-readies-new-ai-model-to-compete-with-google-openai\n\nhttps://em360tech.com/tech-articles/what-mai-1-deep-dive-microsofts-gpt-4-rival","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:7947d43c337cda696af6","name":"Soccer Robot","organization":"Google DeepMind,University College London (UCL)","date":"2024-05-03","domains":["Robotics"],"country":"United States of America,United Kingdom","reference":"https://arxiv.org/abs/2405.02425","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:957d9101431d2fe08bee","name":"MetaMath 70B","organization":"University of Cambridge,Southern University of Science and Technology (SUSTech),Hong Kong University of Science and Technology (HKUST),Huawei Noah's Ark Lab,Alan Turing Institute,Max Planck Institute for Intelligent Systems","date":"2024-05-03","domains":["Language"],"country":"United Kingdom,China,Hong Kong,China,United Kingdom,Germany","reference":"https://arxiv.org/abs/2309.12284","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:82b2e2c68674a1f7fb2e","name":"MetaMath 7B (LLaMa finetune)","organization":"University of Cambridge,Southern University of Science and Technology (SUSTech),Hong Kong University of Science and Technology (HKUST),Huawei Noah's Ark Lab,Alan Turing Institute,Max Planck Institute for Intelligent Systems","date":"2024-05-03","domains":["Language"],"country":"United Kingdom,China,Hong Kong,China,United Kingdom,Germany","reference":"https://arxiv.org/abs/2309.12284","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:172df0a33112173f268d","name":"MetaMath 7B (Mistral finetune)","organization":"University of Cambridge,Southern University of Science and Technology (SUSTech),Hong Kong University of Science and Technology (HKUST),Huawei Noah's Ark Lab,Alan Turing Institute,Max Planck Institute for Intelligent Systems","date":"2024-05-03","domains":["Language"],"country":"United Kingdom,China,Hong Kong,China,United Kingdom,Germany","reference":"https://arxiv.org/abs/2309.12284","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f437b226238c45924497","name":"Idefics2","organization":"Hugging Face,Sorbonne University","date":"2024-05-03","domains":["Multimodal","Language","Vision"],"country":"United States of America,France","reference":"https://arxiv.org/abs/2405.02246","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:bc4cfaba0e6012492743","name":"VILA1.5-13B","organization":"Nvidia,Massachusetts Institute of Technology (MIT)","date":"2024-05-03","domains":["Multimodal","Language","Vision","Video"],"country":"United States of America,United States of America","reference":"https://huggingface.co/Efficient-Large-Model/VILA1.5-13b\nhttps://github.com/NVlabs/VILA/tree/bbc609baf326b1b49b93450b48edc516db3737fc/scripts/v1_5/release/13b\nhttps://developer.nvidia.com/blog/visual-language-models-on-nvidia-hardware-with-vila/\n","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:503c94173a7fce0cb744","name":"OpenELM-1.1B","organization":"Apple","date":"2024-05-02","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.14619","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:fe69cac0693fedb73330","name":"OpenELM-3B","organization":"Apple","date":"2024-05-02","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.14619","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:b1226f614784c78c1ede","name":"OpenELM-450M","organization":"Apple","date":"2024-05-02","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.14619","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:8641500b4b71e7a368d7","name":"OpenELM-270M","organization":"Apple","date":"2024-05-02","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.14619","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:87cfbed5126efcfc3e61","name":"Med-Gemini-M 1.5","organization":"Google DeepMind,Google Research","date":"2024-05-01","domains":["Medicine","Vision","Video","Language","Multimodal"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2404.18416","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:f7638766907296688826","name":"GenCast","organization":"Google DeepMind","date":"2024-05-01","domains":["Earth science"],"country":"United States of America","reference":"https://arxiv.org/abs/2312.15796","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:070662fcaaf5b46ace76","name":"Multi-Token Prediction 7B","organization":"Facebook AI Research","date":"2024-04-30","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/2404.19737","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:ccbb6f2a69181720325a","name":"Multi-Token Prediction 13B","organization":"Facebook AI Research","date":"2024-04-30","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/2404.19737","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:522e125fd8c2ac76ae49","name":"DiffPepBuilder","organization":"Peking University","date":"2024-04-30","domains":["Biology"],"country":"China","reference":"https://arxiv.org/abs/2405.00128","confidence":"Confident","accessibility":""},{"id":"epoch:022af3b1eb6e56786cca","name":"Amazon Q Developer","organization":"Amazon","date":"2024-04-30","domains":["Language"],"country":"United States of America","reference":"https://aws.amazon.com/blogs/aws/amazon-q-developer-now-generally-available-includes-new-capabilities-to-reimagine-developer-experience/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:40f77fa57be2913046c9","name":"SeedLLM","organization":"","date":"2024-04-29","domains":["Speech"],"country":"","reference":"https://www.sciencedirect.com/science/article/abs/pii/S1674205225001728","confidence":"Confident","accessibility":""},{"id":"epoch:dfc4bac8e44a77095f18","name":"TAIDE-LX-7B","organization":"National Science and Technology Council","date":"2024-04-29","domains":["Language"],"country":"Taiwan","reference":"https://huggingface.co/taide/TAIDE-LX-7B","confidence":"Confident","accessibility":""},{"id":"epoch:61275bfaed3bbbe7f360","name":"TAIDE LX-13B","organization":"National Science and Technology Council","date":"2024-04-29","domains":["Language"],"country":"Taiwan","reference":"https://en.taide.tw/","confidence":"Confident","accessibility":""},{"id":"epoch:ea9640f82146c615c591","name":"Llama 3-TAIDE-LX-8B-Chat-Alpha1","organization":"National Science and Technology Council","date":"2024-04-29","domains":["Language"],"country":"Taiwan","reference":"https://huggingface.co/taide/Llama3-TAIDE-LX-8B-Chat-Alpha1","confidence":"Confident","accessibility":""},{"id":"epoch:3b7899a6e60fe431bbde","name":"InternVL1.5","organization":"Shanghai AI Lab,SenseTime,Tsinghua University,Nanjing University,Fudan University,Chinese University of Hong Kong (CUHK)","date":"2024-04-29","domains":["Multimodal","Vision","Language"],"country":"China,Hong Kong,China,China,China,Hong Kong","reference":"https://arxiv.org/abs/2404.16821","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c74d589353b9d6453d99","name":"Swallow","organization":"Tokyo Institute of Technology","date":"2024-04-27","domains":["Language"],"country":"Japan","reference":"https://arxiv.org/abs/2404.17790","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:2d68e75fb675e0175c01","name":"Vidu","organization":"Tsinghua University,ShengShu","date":"2024-04-27","domains":["Video"],"country":"China,China","reference":"https://www.stdaily.com/index/kejixinwen/202408/927f046f4fb743f994726fb08c7e3cbf.shtml\n\nhttps://arxiv.org/abs/2405.04233","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:91057db70c342c75e489","name":"Insights into Human Harmony (\u6d1e\u89c1\u4eba\u548c)","organization":"Zhejiang Lianxin Digital Co., Ltd.","date":"2024-04-27","domains":["Multimodal","Psychology"],"country":"China","reference":"https://www.iyiou.com/news/202404191064563","confidence":"Unknown","accessibility":""},{"id":"epoch:af5a04b97f5f90d26709","name":"Qwen1.5-110B","organization":"Alibaba","date":"2024-04-25","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen1.5-110b/?ref=upstract.com","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5992f728aa010c969a66","name":"Beyond ESM2: Graph-Enhanced Protein Sequence Modeling with Efficient","organization":"Huazhong University of Science and Technology,Fudan University,Northwestern Polytechnical University","date":"2024-04-24","domains":["Biology"],"country":"China,China,China","reference":"https://arxiv.org/abs/2404.15805","confidence":"Confident","accessibility":""},{"id":"epoch:b51e5efcecf62745535f","name":"Arctic","organization":"Snowflake","date":"2024-04-24","domains":["Language"],"country":"United States of America","reference":"https://www.snowflake.com/en/blog/arctic-open-efficient-foundation-language-models-snowflake/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1a8b93734b9115df5388","name":"NEC cotomi","organization":"NEC Corporation","date":"2024-04-24","domains":["Language"],"country":"Japan","reference":"https://www.nec.com/en/press/202404/global_20240424_01.html","confidence":"Unknown","accessibility":""},{"id":"epoch:5860c97a1fe2eb437daa","name":"Yuanjing LLM (\u8054\u901a\u5143\u666f\u5927\u6a21\u578b)","organization":"China Unicom","date":"2024-04-24","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250708142523/http://www.xinhuanet.com/info/20240526/9332a929aa3a43ccb50fb64c18abce93/c.html\nhttps://web.archive.org/web/20250708142743/https://www.stdaily.com/web/gdxw/2025-01/03/content_282553.html\nhttps://web.archive.org/web/20250708142937/https://www.hiascend.com/marketplace/solution/detail/2181","confidence":"Unknown","accessibility":""},{"id":"epoch:7012de0eaa272c475bbc","name":"phi-3-mini 3.8B","organization":"Microsoft","date":"2024-04-23","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.14219","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:598f2e4d5ad19e2f8598","name":"phi-3-medium 14B","organization":"Microsoft","date":"2024-04-23","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.14219","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:46b055bc0aa1d828a41f","name":"SenseChat 5.0","organization":"SenseTime","date":"2024-04-23","domains":["Language"],"country":"Hong Kong","reference":"https://zhidx.com/p/421866.html","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:3ce713ec571f93ca5d91","name":"phi-3-small 7.4B","organization":"Microsoft","date":"2024-04-23","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.14219","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7ce966facd4fbbc7c2aa","name":"SI-PLM","organization":"University of Pittsburgh","date":"2024-04-23","domains":["Biology"],"country":"United States of America","reference":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11071282/","confidence":"Unknown","accessibility":""},{"id":"epoch:9cd329e091e6956e53c4","name":"Phi-3.5-mini","organization":"Microsoft","date":"2024-04-23","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.14219","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:20bb1ffc37ad54d13904","name":"Phi-3.5-vision","organization":"Microsoft","date":"2024-04-23","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.14219","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:79b2843f5a0bf2de4ca0","name":"Phi-3.5-MoE","organization":"Microsoft","date":"2024-04-23","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.14219","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8e327b138dc9b2ff3727","name":"Phoenix (\u51e4\u51f0\u5927\u6a21\u578b)","organization":"Zhixin Shuchuang (Chongqing) Technology Co., Ltd.","date":"2024-04-23","domains":[],"country":"China","reference":"PDF doc: https://www.cac.gov.cn/cms/pub/interact/downloadfile.jsp?fText=%E7%94%9F%E6%88%90%E5%BC%8F%E4%BA%BA%E5%B7%A5%E6%99%BA%E8%83%BD%E6%9C%8D%E5%8A%A1%E5%B7%B2%E5%A4%87%E6%A1%88%E5%92%8C%E5%B7%B2%E7%99%BB%E8%AE%B0%E4%BF%A1%E6%81%AF%EF%BC%882024%E5%B9%B48%E6%9C%88%EF%BC%89&filepath=NUtqEIwGiCjGm2Bhl20cvD3k2jZHxW62PpTbCodAq%2Fhvz0vHSSJ6sATU6DtmcGo1SVh2XLVTbvXpuP4KM0gfclAfqphfFn00jUfFRLdBE5I%3D&utm_source=chatgpt.com","confidence":"Unknown","accessibility":""},{"id":"epoch:483f85eaed971c291f6e","name":"Firefly Image 3","organization":"Adobe","date":"2024-04-23","domains":["Image generation"],"country":"United States of America","reference":"https://news.adobe.com/news/news-details/2024/adobe-introduces-firefly-image-3-foundation-model-to-take-creative-exploration-and-ideation-to-new-heights","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:7f1e033e5c2ee4b51265","name":"VISTA-2D","organization":"Nvidia","date":"2024-04-22","domains":["Biology","Vision"],"country":"United States of America","reference":"https://developer.nvidia.com/blog/advancing-cell-segmentation-and-morphology-analysis-with-nvidia-ai-foundation-model-vista-2d/","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:4283f3d7d3f3149c7fa0","name":"InstructPLM","organization":"Zhejiang Lab,Zhejiang University (ZJU),Nanjing University,Tsinghua University,Alibaba,Chinese University of Hong Kong (CUHK)","date":"2024-04-20","domains":["Biology"],"country":"China,China,China,China,China,Hong Kong","reference":"https://www.biorxiv.org/content/10.1101/2024.04.17.589642v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:7a594f311f73e7f622aa","name":"SaProt","organization":"Zhejiang University (ZJU),Westlake University","date":"2024-04-19","domains":["Biology"],"country":"China,China","reference":"https://www.biorxiv.org/content/10.1101/2023.10.01.560349v5.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:fc7355ed1ffd3f0790db","name":"Llama 3-70B","organization":"Meta AI","date":"2024-04-18","domains":["Language"],"country":"United States of America","reference":"https://ai.meta.com/blog/meta-llama-3/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:346c553c9a7a695b22b1","name":"Llama 3-8B","organization":"Meta AI","date":"2024-04-18","domains":["Language"],"country":"United States of America","reference":"https://ai.meta.com/blog/meta-llama-3/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:118701cac09e656a087f","name":"FRED-T5-XL","organization":"Sber","date":"2024-04-18","domains":["Language"],"country":"Russia","reference":"https://arxiv.org/abs/2309.10931","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2e4ce51ec793ae73df0a","name":"LLaMA-3-Instruct-8B","organization":"Meta AI","date":"2024-04-18","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:bdcc465a1db37e705d3b","name":"Parakeet ASR rnnt 1.1B","organization":"Nvidia","date":"2024-04-18","domains":["Speech"],"country":"United States of America","reference":"https://developer.nvidia.com/blog/pushing-the-boundaries-of-speech-recognition-with-nemo-parakeet-asr-models/#introducing_the_parakeet_asr_family","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:114756b34e117940871c","name":"Reka Edge","organization":"Reka AI","date":"2024-04-18","domains":["Multimodal","Language","Vision","Video","Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.12387","confidence":"Likely","accessibility":"API access"},{"id":"epoch:deb92279f6913f2fe1ff","name":"EVI","organization":"Hume","date":"2024-04-18","domains":["Speech"],"country":"United States of America","reference":"https://www.hume.ai/blog/introducing-hume-evi-api\n\nhttps://dev.hume.ai/docs/empathic-voice-interface-evi/configuration/evi-version","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:950765a791afd962dc7c","name":"Mixtral 8x22B","organization":"Mistral AI","date":"2024-04-17","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/mixtral-8x22b/","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9394fdd135033fb92eb7","name":"SIMA","organization":"Google DeepMind","date":"2024-04-17","domains":["Games","Video"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.10179","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:49ea8f7dd617a1bac8bb","name":"METL-Global","organization":"University of Wisconsin Madison,Morgridge Institute for Research","date":"2024-04-17","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.03.15.585128v1.abstract","confidence":"Likely","accessibility":""},{"id":"epoch:e03b9bb05e89da23b8f8","name":"GRITLM 7B","organization":"Contextual AI,The University of Hong Kong,Microsoft","date":"2024-04-17","domains":["Language"],"country":"United States of America,Hong Kong,United States of America","reference":"https://arxiv.org/abs/2402.09906","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:60d8976b4ee18f2344a4","name":"GRITLM 8x7B","organization":"Contextual AI,The University of Hong Kong,Microsoft","date":"2024-04-17","domains":["Language"],"country":"United States of America,Hong Kong,United States of America","reference":"https://arxiv.org/abs/2402.09906","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:33a4a31bdc1c7a61ecb5","name":"Tiangong 3.0 (MoE)","organization":"Kunlun Inc.","date":"2024-04-17","domains":["Language"],"country":"China","reference":"https://www.leiphone.com/category/industrynews/aKm3DiETl4rw1A6G.html","confidence":"Confident","accessibility":""},{"id":"epoch:502f160e55e0b915638f","name":"Tiangong SkyMusic","organization":"Kunlun Inc.","date":"2024-04-17","domains":["Audio"],"country":"China","reference":"https://news.futunn.com/en/post/40677784/kunlun-world-wide-300418-upgraded-tiangong-3-0-big-model?level=1&data_ticket=1733998099122708","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:c57383da057136789495","name":"abab6.5","organization":"MiniMax","date":"2024-04-17","domains":["Language"],"country":"China","reference":"https://www.minimaxi.com/en/news/abab65-series","confidence":"Confident","accessibility":"API access"},{"id":"epoch:31443c5f88444ca23ebc","name":"OLMo 1.7-7B","organization":"Allen Institute for AI","date":"2024-04-17","domains":["Language"],"country":"United States of America","reference":"https://allenai.org/blog/olmo-1-7-7b-a-24-point-improvement-on-mmlu-92b43f7d269d","confidence":"","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ba71d569be9c1552f0d5","name":"LINGO-2","organization":"Wayve","date":"2024-04-17","domains":["Driving","Vision","Language","Video"],"country":"United Kingdom","reference":"https://wayve.ai/thinking/lingo-2-driving-with-language/","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:cfb72a7e93fd82b693d9","name":"Reka Core","organization":"Reka AI","date":"2024-04-15","domains":["Multimodal","Language","Vision","Video","Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.12387","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:3e224b22526065e27291","name":"Reka Flash","organization":"Reka AI","date":"2024-04-15","domains":["Multimodal","Language","Vision","Video","Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.12387","confidence":"Likely","accessibility":"API access"},{"id":"epoch:4900665233e636b07bd7","name":"WizardLM-2 8x22B","organization":"Microsoft","date":"2024-04-15","domains":["Language"],"country":"United States of America","reference":"https://wizardlm.github.io/WizardLM2/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:85da130cfc024ec4dcbf","name":"WizardLM-2 70B","organization":"Microsoft","date":"2024-04-15","domains":["Language"],"country":"United States of America","reference":"https://wizardlm.github.io/WizardLM2/","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:d55002838db4ed3f6a73","name":"WizardLM-2 7B","organization":"Microsoft","date":"2024-04-15","domains":["Language"],"country":"United States of America","reference":"https://wizardlm.github.io/WizardLM2/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:cf0abc62c1ecb36fdda1","name":"DDPM","organization":"University Paris-Saclay,Radboud University Medical Center","date":"2024-04-13","domains":["Biology"],"country":"France,Netherlands","reference":"https://www.biorxiv.org/content/10.1101/2024.04.10.588825v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:9e59dd65c95547e30133","name":"DDIM","organization":"University Paris-Saclay,Radboud University Medical Center","date":"2024-04-13","domains":["Biology"],"country":"France,Netherlands","reference":"https://www.biorxiv.org/content/10.1101/2024.04.10.588825v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:6e461c9c93f9b90e7a25","name":"Bencao Zhiku","organization":"Chengdu University of Traditional Chinese Medicine","date":"2024-04-12","domains":["Language"],"country":"China","reference":"https://ecas.cas.cn/xxkw/kbcd/201115_145987/ml/xxhcxyyyal/202407/t20240718_5026378.html","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:21e3c2413f644c994aa7","name":"NOMI GPT","organization":"NIO","date":"2024-04-12","domains":["Multimodal","Language","Vision","Speech"],"country":"China","reference":"https://auto.gasgoo.com/news/202404/12I70388757C601.shtml\nhttps://www.nio.cn/smart-technology/20240929001","confidence":"Unknown","accessibility":""},{"id":"epoch:5210e60e6160e693aa89","name":"tsuzumi 7B upgrade 2024","organization":"NTT Communication Science Laboratories","date":"2024-04-11","domains":["Language","Multimodal","Vision"],"country":"Japan","reference":"https://ntt-research.com/2024-upgrade-reality-tsuzumi/","confidence":"Likely","accessibility":""},{"id":"epoch:52935a2a9fa3ace1e1ba","name":"HGRN2 3B","organization":"Shanghai AI Lab,Massachusetts Institute of Technology (MIT),Taptap","date":"2024-04-11","domains":["Language"],"country":"China,United States of America,Spain","reference":"https://arxiv.org/abs/2404.07904v1","confidence":"Confident","accessibility":""},{"id":"epoch:b206df62aed0a856cd50","name":"HGRN2 1B","organization":"Shanghai AI Lab,Massachusetts Institute of Technology (MIT),Taptap","date":"2024-04-11","domains":["Language"],"country":"China,United States of America,Spain","reference":"https://arxiv.org/abs/2404.07904v1","confidence":"Confident","accessibility":""},{"id":"epoch:903cfef8d7517c99fe30","name":"AF2RAVE","organization":"University of Maryland","date":"2024-04-10","domains":["Biology"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.07102v3","confidence":"Unknown","accessibility":""},{"id":"epoch:f713d98c900ae2a24640","name":"Zephyr 141B-A39B","organization":"Hugging Face,Korea Advanced Institute of Science and Technology (KAIST),Argilla","date":"2024-04-10","domains":["Language"],"country":"United States of America,South Korea,Spain","reference":"https://huggingface.co/HuggingFaceH4/zephyr-orpo-141b-A35b-v0.1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d4b0dbdedd23b31ea142","name":"DiffBindFR","organization":"Peking University,Tsinghua-Peiking Center for Life Sciences","date":"2024-04-09","domains":["Biology"],"country":"China,China","reference":"https://pubs.rsc.org/en/content/articlehtml/2024/sc/d3sc06803j","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:da1c372220f3071a39e1","name":"WeituAI 1.0","organization":"Weitu AI","date":"2024-04-09","domains":["Multimodal","Language","Vision"],"country":"Hong Kong","reference":"https://36kr.com/p/2725929816384770","confidence":"Confident","accessibility":""},{"id":"epoch:fe9cd942d7cdf6cda21f","name":"GPT-4 Turbo (Apr 2024)","organization":"OpenAI","date":"2024-04-09","domains":["Multimodal","Vision","Language","Image generation"],"country":"United States of America","reference":"https://x.com/OpenAI/status/1778574613813006610","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:ac9d7731ffa1ac2ee141","name":"Stable LM 2 12B","organization":"Stability AI","date":"2024-04-08","domains":["Language"],"country":"United Kingdom","reference":"https://stability.ai/news/introducing-stable-lm-2-12b\nhttps://huggingface.co/stabilityai/stablelm-2-12b","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:aaab24b81343bf459e26","name":"YaART","organization":"Yandex","date":"2024-04-08","domains":["Image generation"],"country":"Russia","reference":"https://arxiv.org/abs/2404.05666","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:d5fc658f7b797aeaba6a","name":"OpenThaiGPT v1.0.0 (13B)","organization":"Mahidol University,AI Entrepreneurs Association of Thailand","date":"2024-04-08","domains":["Language"],"country":"Thailand,Thailand","reference":"https://openthaigpt.aieat.or.th/openthaigpt-1.0.0-less-than-8-apr-2024-greater-than","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:d85ec97af9a0d430e360","name":"OpenThaiGPT v1.0.0 (7B)","organization":"Mahidol University,AI Entrepreneurs Association of Thailand","date":"2024-04-08","domains":["Language"],"country":"Thailand,Thailand","reference":"https://openthaigpt.aieat.or.th/openthaigpt-1.0.0-less-than-8-apr-2024-greater-than","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:2e63e4a6baa84026b036","name":"SambaLingo-Thai-Chat (7B)","organization":"SambaNova Systems, Inc","date":"2024-04-08","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.05829","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7d50637972633676ba76","name":"SambaLingo-Thai-Chat-70B","organization":"SambaNova Systems, Inc","date":"2024-04-08","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2404.05829","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:0d64bf015e8c04171839","name":"Solar Stone Mining Large Model","organization":"","date":"2024-04-07","domains":[],"country":"","reference":"https://www.stdaily.com/index/kejixinwen/202404/8e75d8cef5a648a480a79b0f2c59a128.shtml (LINK BROKEN)","confidence":"Unknown","accessibility":""},{"id":"epoch:59a900ab96c416ec3a80","name":"ESM-AA","organization":"Peking University,Nanjing University,Tsinghua University,PharMolix","date":"2024-04-05","domains":["Biology"],"country":"China,China,China,China","reference":"https://arxiv.org/abs/2403.12995","confidence":"Confident","accessibility":""},{"id":"epoch:b35f017dac59f153b0af","name":"Command R+","organization":"Cohere,Cohere Labs (formerly Cohere for AI)","date":"2024-04-04","domains":["Language"],"country":"Canada,Canada","reference":"https://txt.cohere.com/command-r-plus-microsoft-azure/","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:ba945dc03c6b9a73cafd","name":"Viking","organization":"Silo AI,University of Turku","date":"2024-04-04","domains":["Language"],"country":"Finland,Finland","reference":"https://huggingface.co/LumiOpen/Viking-33B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:cb59ff4b6a3365c656b2","name":"eFold","organization":"Harvard Medical School,Stanford University,Columbia University,University of Strasbourg","date":"2024-04-04","domains":["Biology"],"country":"United States of America,United States of America,United States of America,France","reference":"https://www.researchsquare.com/article/rs-4159627/v1","confidence":"Confident","accessibility":""},{"id":"epoch:c1ffee709d13dfeca461","name":"Sailor-7B-Chat","organization":"Sea AI Lab,Singapore University of Technology & Design","date":"2024-04-04","domains":["Language"],"country":"Singapore,Singapore","reference":"https://arxiv.org/abs/2404.03608","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:98cdecc3771119bf2148","name":"Universal-1","organization":"AssemblyAI","date":"2024-04-03","domains":["Speech"],"country":"","reference":"https://www.assemblyai.com/research/universal-1","confidence":"Confident","accessibility":"API access"},{"id":"epoch:06d8ea929d4092e1f14b","name":"Mixture-of-Depths","organization":"Google DeepMind,McGill University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms)","date":"2024-04-02","domains":["Language"],"country":"United States of America,Canada,Canada","reference":"https://arxiv.org/abs/2404.02258","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:a45530002bee632df5be","name":"POKE\u00b4LLMON","organization":"Georgia Institute of Technology","date":"2024-04-02","domains":["Games","Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2402.01118","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:af3ac04d71b7443b858a","name":"AutoDiff","organization":"Galixir Technologies,Rensselaer Polytechnic Institute,Massachusetts Institute of Technology (MIT)","date":"2024-04-02","domains":["Biology"],"country":"China,United States of America,United States of America","reference":"https://arxiv.org/abs/2404.02003","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:bc4bbd882d1c6effae8d","name":"XVERSE-MoE-A4.2B","organization":"XVERSE Technology,Shenzhen Yuanxiang Technology","date":"2024-04-02","domains":["Language"],"country":"China,China","reference":"https://huggingface.co/xverse/XVERSE-MoE-A4.2B","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:2f02e7a7ac429efbf2b0","name":"APUS-xDAN-4.0(MoE)","organization":"Qilin Hesheng Network Technology Co., Ltd. (APUS)","date":"2024-04-02","domains":["Multimodal","Language"],"country":"China","reference":"https://github.com/shootime2021/APUS-xDAN-4.0-moe\nhttps://zjb.henan.gov.cn/2024/05-27/2998639.html","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:c6d5d3205b144bc4abe9","name":"Youyuanjian (\u90ae\u8fdc\u89c1)","organization":"China Post Consumer Finance Co., Ltd.","date":"2024-04-02","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250627092016/https://www.chinapost.com.cn/html1/report/2404/1179-1.htm","confidence":"Unknown","accessibility":""},{"id":"epoch:a1f00024f09494be185c","name":"MobileCLIP-B (LT)","organization":"Apple","date":"2024-04-01","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2311.17049","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d80cf6849c2aa00c3dbd","name":"TeleChat-7B","organization":"China Telecom","date":"2024-04-01","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2401.03804","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:128fc9151537793e2ef1","name":"TeleChat-3B","organization":"China Telecom","date":"2024-04-01","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2401.03804","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:03c3187fd0d327ffe7b2","name":"Wutong 2.0","organization":"","date":"2024-04-01","domains":[],"country":"","reference":"https://news.qq.com/rain/a/20241022A077IC00","confidence":"Unknown","accessibility":""},{"id":"epoch:f70af275eead2d4c1a4c","name":"TeleChat-12B","organization":"China Telecom","date":"2024-04-01","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2401.03804","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:cf9daaea454c3f845814","name":"TW3-JRGL-v2","organization":"French Engineering School ECE,TW3 Partners","date":"2024-04-01","domains":["Language"],"country":"France,France","reference":"https://huggingface.co/paloalma/TW3-JRGL-v2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:48caa336b046b95e093d","name":"Le_Triomphant-ECE-TW3","organization":"French Engineering School ECE,TW3 Partners","date":"2024-04-01","domains":["Language"],"country":"France,France","reference":"https://huggingface.co/paloalma/Le_Triomphant-ECE-TW3","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a1dd7635c81cdf59460e","name":"ReALM","organization":"Apple","date":"2024-03-29","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2403.20329","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:cc87da4f7c9cc4f093fa","name":"Voice Engine","organization":"OpenAI","date":"2024-03-29","domains":["Audio","Speech"],"country":"United States of America","reference":"https://openai.com/blog/navigating-the-challenges-and-opportunities-of-synthetic-voices","confidence":"Unknown","accessibility":""},{"id":"epoch:a16fc094864779573173","name":"Grok-1.5","organization":"xAI","date":"2024-03-28","domains":["Language"],"country":"United States of America","reference":"https://x.ai/blog/grok-1.5","confidence":"Speculative","accessibility":"Hosted access (no API)"},{"id":"epoch:0c89c0784258892af065","name":"Grok-1.5V","organization":"xAI","date":"2024-03-28","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://x.ai/blog/grok-1.5v","confidence":"Speculative","accessibility":"Hosted access (no API)"},{"id":"epoch:57419eac6781937fca91","name":"YandexGPT 3","organization":"Yandex","date":"2024-03-28","domains":["Language"],"country":"Russia","reference":"https://ya.ru/ai/gpt-3","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:1be9bd28f6cc874e32f0","name":"Jamba","organization":"AI21 Labs","date":"2024-03-28","domains":["Language"],"country":"Israel","reference":"https://arxiv.org/abs/2403.19887","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:dfffe42cdba410c69966","name":"DBRX","organization":"Databricks","date":"2024-03-27","domains":["Language"],"country":"United States of America","reference":"https://www.databricks.com/blog/introducing-dbrx-new-state-art-open-llm","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:85bd3634557e4e1c1dbe","name":"AInno-75B 2.0","organization":"","date":"2024-03-27","domains":[],"country":"","reference":"https://www.php.cn/faq/727953.html","confidence":"Confident","accessibility":""},{"id":"epoch:539125d231543a7e88a6","name":"MultiVerse 70B","organization":"MTS","date":"2024-03-25","domains":["Language"],"country":"Russia","reference":"https://huggingface.co/MTSAIR/MultiVerse_70B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:6176d4e099235c58cc4f","name":"ProstT5","organization":"Technical University of Munich,Seoul National University,Institute for Advanced Study,TUM School of Life Sciences Weihenstephan","date":"2024-03-24","domains":["Biology"],"country":"Germany,South Korea,United States of America,Germany","reference":"https://www.biorxiv.org/content/10.1101/2023.07.23.550085v2.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:0151a26586e8adfa3064","name":"CrossBind","organization":"Shanghai AI Lab,Fudan University,Loughborough University,Chinese University of Hong Kong (CUHK),Shanghai Jiao Tong University","date":"2024-03-24","domains":["Biology"],"country":"China,China,United Kingdom,Hong Kong,China","reference":"https://ojs.aaai.org/index.php/AAAI/article/view/28044","confidence":"Confident","accessibility":""},{"id":"epoch:1c4abdb94ff712433c74","name":"BindDM","organization":"Peng Cheng Laboratory,Peking University,University of Science and Technology of China (USTC),ByteDance,Tsinghua University","date":"2024-03-24","domains":["Biology"],"country":"China,China,China,China,China","reference":"https://ojs.aaai.org/index.php/AAAI/article/view/29162","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:8f87dee6f0f0060b48b3","name":"Suno v3","organization":"Suno","date":"2024-03-21","domains":["Audio"],"country":"United States of America","reference":"https://suno.com/blog/v3","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:7e3afd2d3ce327ed422f","name":"Qiyuan 3.0","organization":"","date":"2024-03-20","domains":[],"country":"","reference":"https://www.stdaily.com/index/kejixinwen/202403/976fc5638b0c4323b19c6e4349dacfa5.shtml (LINK BROKEN)","confidence":"Unknown","accessibility":""},{"id":"epoch:83c493365cfd696e5b5f","name":"Xuanji Yuheng (\u7487\u7391\u7389\u8861)","organization":"Zhuoshi Technology","date":"2024-03-20","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250624093419/https://www.maas.com.cn/news/299.html\nhttps://web.archive.org/web/20250624094052/https://www.xhby.net/content/s67dd07c7e4b01708361b0b9e.html","confidence":"Confident","accessibility":""},{"id":"epoch:8104170c6d3df3cdf639","name":"MiniGPT4 + LRV-Instruction","organization":"University of Maryland,Microsoft","date":"2024-03-19","domains":["Language","Vision","Multimodal"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2306.14565","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3ea1d105e77ed1ad805f","name":"JetFire (GPT2-LARGE)","organization":"Tsinghua University","date":"2024-03-19","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2403.12422v1","confidence":"Confident","accessibility":""},{"id":"epoch:7b54e3d3d61e6ef54c23","name":"Stable Video 3D (SV3D)","organization":"Stability AI","date":"2024-03-18","domains":["Vision","3D modeling"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2403.12008","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:14847087a9475207e804","name":"ERNIE-RNA","organization":"Microsoft Research,Syngentech,Tsinghua University","date":"2024-03-17","domains":["Biology"],"country":"United States of America,China,China","reference":"https://www.biorxiv.org/content/10.1101/2024.03.17.585376v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:c55e1196ee8585aa6515","name":"PocketVec","organization":"Barcelona Institute of Science and Technology,Universitat de Barcelona,Instituci\u00f3 Catalana de Recerca i Estudis Avan\u00e7ats (ICREA)","date":"2024-03-16","domains":["Biology"],"country":"Spain,Spain,Spain","reference":"https://www.biorxiv.org/content/10.1101/2024.03.14.584971v1.abstract","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:6fc455e107b99685377b","name":"Xinghai (\u661f\u6d77)","organization":"Hisense","date":"2024-03-15","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://web.archive.org/web/20250627092753/https://www.sohu.com/a/764344251_115978","confidence":"Unknown","accessibility":""},{"id":"epoch:a29845109130b50f89b2","name":"MM1-30B","organization":"Apple","date":"2024-03-14","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2403.09611","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:6a4ba6e643e42f3724d2","name":"Quiet-STaR","organization":"Stanford University","date":"2024-03-14","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2403.09629","confidence":"Unknown","accessibility":""},{"id":"epoch:355191ea0218371e35a2","name":"ManiGaussian","organization":"Tsinghua University,Nanyang Technological University,Carnegie Mellon University (CMU)","date":"2024-03-13","domains":["Robotics","Vision","Video"],"country":"China,Singapore,United States of America","reference":"https://arxiv.org/abs/2403.08321\nhttps://guanxinglu.github.io/ManiGaussian/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ac9b2b9aba7013c5b6a6","name":"Recraft V2 (Recraft 20B)","organization":"Recraft","date":"2024-03-13","domains":["Image generation"],"country":"United Kingdom","reference":"https://www.recraft.ai/blog/recraft-20b","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:1aa88a065e9435027de3","name":"Command R","organization":"Cohere,Cohere Labs (formerly Cohere for AI)","date":"2024-03-11","domains":["Language"],"country":"Canada,Canada","reference":"https://cohere.com/blog/command-r","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:eab82cc564a9b063b465","name":"RFM-1","organization":"Covariant","date":"2024-03-11","domains":["Robotics","Vision","Video"],"country":"United States of America","reference":"https://covariant.ai/insights/introducing-rfm-1-giving-robots-human-like-reasoning-capabilities/","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:b3ee0e2fe248c87c1bc8","name":"Dream Home LLM (\u8d1d\u58f3\u68a6\u60f3\u5bb6\u5927\u6a21\u578b)","organization":"KE Holdings Inc. (\u201cBeike\u201d)","date":"2024-03-11","domains":["Image generation"],"country":"China","reference":"https://web.archive.org/web/20250624060002/https://www.shaqiu.cn/article/WPklLGD1VQ2z\nhttps://web.archive.org/web/20250624060209/https://finance.sina.com.cn/cj/2025-05-19/doc-inexavha2295249.shtml","confidence":"Unknown","accessibility":""},{"id":"epoch:ef0baeb61d53a1104b5f","name":"HAM-TTS","organization":"Geely Automobile Research Institute (Ningbo) Company,National Institute of Informatics,Shanghai Jiao Tong University","date":"2024-03-09","domains":["Speech"],"country":"China,Japan,China","reference":"https://arxiv.org/abs/2403.05989","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:7460cd0200b291e6fb5a","name":"Derm Foundational Model","organization":"Google Research","date":"2024-03-08","domains":["Vision","Medicine"],"country":"United States of America","reference":"https://research.google/blog/health-specific-embedding-tools-for-dermatology-and-pathology/","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:ae22cbad757cb3c6e541","name":"DeepSeek-VL-7B","organization":"DeepSeek","date":"2024-03-08","domains":["Multimodal","Vision","Language"],"country":"China","reference":"https://arxiv.org/abs/2403.05525","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:a01e9e4611785aa6de79","name":"DeepSeek-VL-1.3B","organization":"DeepSeek","date":"2024-03-08","domains":["Multimodal","Vision","Language"],"country":"China","reference":"https://arxiv.org/abs/2403.05525","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:0b2c2443e355f20f4f7d","name":"Inflection-2.5","organization":"Inflection AI","date":"2024-03-07","domains":["Language"],"country":"United States of America","reference":"https://inflection.ai/inflection-2-5","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:59887b690c3e5441d35a","name":"BaseFold","organization":"Basecamp Research","date":"2024-03-06","domains":["Biology"],"country":"United Kingdom","reference":"https://www.biorxiv.org/content/10.1101/2024.03.06.583325v1","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:b18d08e2051f0126b972","name":"GroundingGPT","organization":"ByteDance,Fudan University","date":"2024-03-05","domains":["Multimodal","Language","Video","Vision","Speech"],"country":"China,China","reference":"https://arxiv.org/abs/2401.06071v5","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:150cb3df36c1cf7008f9","name":"Claude 3 Haiku","organization":"Anthropic","date":"2024-03-04","domains":["Language"],"country":"United States of America","reference":"https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:dbe1f1d50df39b0a7e46","name":"Claude 3 Sonnet","organization":"Anthropic","date":"2024-03-04","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:b76f559cc0c7732d3844","name":"Claude 3 Opus","organization":"Anthropic","date":"2024-03-04","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:95697efde30d58472433","name":"Aramco Metabrain AI","organization":"Saudi Aramco","date":"2024-03-04","domains":["Language"],"country":"Saudi Arabia","reference":"https://www.offshore-technology.com/news/saudi-aramco-unveils-industry-first-generative-ai-model/","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:8a4dc34c2cbc616f6a7e","name":"MACE-MP-0","organization":"University of Cambridge,Federal Institute of Materials Research and Testing (BAM),NERSC, Lawrence Berkeley National Laboratory,University of British Columbia (UBC),Friedrich Schiller University Jena,University of Bayreuth,Fritz Haber Institute of the Max Planck Society,U. S. Naval Research Laboratory,Chemix,Daresbury Laboratory,BASF,University of South Carolina,University of Stuttgart,Uppsala University,Newcastle University,Technical University of Denmark,Aix-Marseille Universit\u00e9,University of Warwick,University of California Los Angeles (UCLA),InstaDeep,University of California (UC) Berkeley","date":"2024-03-01","domains":["Materials science"],"country":"United Kingdom,Germany,United States of America,Canada,Germany,Germany,Germany,United States of America,United States of America,United Kingdom,Germany,United States of America,Germany,Sweden,United Kingdom,Denmark,France,United Kingdom,United States of America,United Kingdom,United States of America","reference":"https://arxiv.org/abs/2401.00096","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:84fc7123cd43e95eff24","name":"Step-1X","organization":"StepFun","date":"2024-03-01","domains":["Image generation"],"country":"China","reference":"https://www.sohu.com/a/791101272_114877\n\nhttps://platform.stepfun.com/docs/llm/image","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:12f35dbe530034fa54a0","name":"Step-1.5V","organization":"StepFun","date":"2024-03-01","domains":["Vision"],"country":"China","reference":"https://www.sohu.com/a/791101272_114877","confidence":"Unknown","accessibility":""},{"id":"epoch:62ee67241c475de817c1","name":"Step-2","organization":"StepFun","date":"2024-03-01","domains":["Language"],"country":"China","reference":"https://www.sohu.com/a/791101272_114877\nhttps://platform.stepfun.com/#step2","confidence":"Likely","accessibility":"API access"},{"id":"epoch:91a7195fcea94a33bb55","name":"Ovis2 16B","organization":"Alibaba","date":"2024-03-01","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://huggingface.co/AIDC-AI/Ovis2-16B","confidence":"Confident","accessibility":""},{"id":"epoch:86dc77c82c1b3e41f0aa","name":"Ovis2 34B","organization":"Alibaba","date":"2024-03-01","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://huggingface.co/AIDC-AI/Ovis2-34B","confidence":"Confident","accessibility":""},{"id":"epoch:308c93994118cb2def8e","name":"StarCoder 2 15B","organization":"Hugging Face,ServiceNow,Nvidia,BigCode","date":"2024-02-29","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2402.19173","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:fa29a5f910c3e3d69948","name":"StarCoder 2 7B","organization":"Hugging Face,ServiceNow,Nvidia,BigCode","date":"2024-02-29","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2402.19173","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:2e93250b8e57225ed56d","name":"StarCoder 2 3B","organization":"Hugging Face,ServiceNow,Nvidia,BigCode","date":"2024-02-29","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2402.19173","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:16d2793addd6bbab9dad","name":"Griffin","organization":"Google DeepMind","date":"2024-02-29","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2402.19427","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:fc40dac258fadfdf8914","name":"Hawk","organization":"Google DeepMind","date":"2024-02-29","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2402.19427","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:865ec38971a6f8511964","name":"Humanoid Locomotion","organization":"University of California (UC) Berkeley","date":"2024-02-29","domains":["Robotics"],"country":"United States of America","reference":"https://arxiv.org/abs/2402.19469","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:c1f963a38523c81860a1","name":"YOLOv9-E","organization":"Academia Sinica,National Taipei University of Technology,Chung Yuan Christian University","date":"2024-02-29","domains":["Vision"],"country":"Taiwan,Taiwan,Taiwan","reference":"https://arxiv.org/abs/2402.13616","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6e259aadaa272d5d2632","name":"RiNALMo","organization":"University of Zagreb,Genome Institute of Singapore,Bioinformatics Institute","date":"2024-02-29","domains":["Biology"],"country":"Croatia,Singapore,Singapore","reference":"https://arxiv.org/abs/2403.00043","confidence":"Confident","accessibility":""},{"id":"epoch:e955c74b1bb5131cfd8d","name":"PTM-Mamba","organization":"Duke University","date":"2024-02-29","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/10.1101/2024.02.28.581983v1","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:1680d89ac81d60a5b628","name":"AI Q&A Robot","organization":"","date":"2024-02-29","domains":[],"country":"","reference":"https://www.1ai.net/en/10155.html","confidence":"Confident","accessibility":""},{"id":"epoch:ddaa1b44858d42f13792","name":"Protllm","organization":"Beijing Institute of Technology,Beihang University,Peking University,Smart Grid Research Institute,Shanghai AI Lab","date":"2024-02-28","domains":["Biology"],"country":"China,China,China,China,China","reference":"https://arxiv.org/abs/2403.07920","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:ae1e5b413db62c5d46c8","name":"Ideogram 1.0","organization":"Ideogram","date":"2024-02-28","domains":["Image generation"],"country":"Canada","reference":"https://about.ideogram.ai/1.0","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:e4f6afe9829b5831ec42","name":"Evo","organization":"Stanford University,University of California (UC) Berkeley,Together","date":"2024-02-27","domains":["Biology"],"country":"United States of America,United States of America,United States of America","reference":"https://arcinstitute.org/news/blog/evo\nhttps://www.biorxiv.org/content/10.1101/2024.02.27.582234v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:67156e3e1c7005a3a1d7","name":"BitNet b1.58","organization":"University of Chinese Academy of Sciences,Microsoft Research","date":"2024-02-27","domains":["Language"],"country":"China,United States of America","reference":"https://arxiv.org/abs/2402.17764","confidence":"Confident","accessibility":""},{"id":"epoch:f65808cf94af24b9203a","name":"Nemotron-4 15B","organization":"Nvidia","date":"2024-02-27","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2402.16819","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:7cb149540bc65767ea00","name":"Palmyra Vision","organization":"Writer","date":"2024-02-27","domains":["Vision","Video","Multimodal","Language"],"country":"United States of America","reference":"https://writer.com/blog/palmyra-vision/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:707cb31a2b6f209c5da4","name":"Playground v2.5","organization":"Playground","date":"2024-02-27","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2402.17245v1","confidence":"Unknown","accessibility":"Open weights (restricted use)"},{"id":"epoch:b04b27dcf25103bbd10c","name":"Mistral Large","organization":"Mistral AI","date":"2024-02-26","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/mistral-large/","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:e7d85b3c6b7c725c295d","name":"ProLLaMA","organization":"Peking University,Peng Cheng Laboratory","date":"2024-02-26","domains":["Biology"],"country":"China,China","reference":"https://arxiv.org/abs/2402.16445","confidence":"Confident","accessibility":""},{"id":"epoch:37dc3f4987a6aed0ea23","name":"DecompDiff","organization":"University of Illinois Urbana-Champaign (UIUC),ByteDance,University of Chinese Academy of Sciences,Chinese Academy of Sciences,Tsinghua University","date":"2024-02-26","domains":["Biology"],"country":"United States of America,China,China,China,China","reference":"https://arxiv.org/abs/2403.07902","confidence":"Confident","accessibility":""},{"id":"epoch:72831bfc1a231b5c30c9","name":"Gemma 1.1 7B Instruct","organization":"Google","date":"2024-02-24","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/google/gemma-1.1-7b-it","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:45c0c4fed93820f1f751","name":"SDXL-Lightning","organization":"ByteDance","date":"2024-02-24","domains":["Image generation"],"country":"China","reference":"https://arxiv.org/abs/2402.13929","confidence":"Unknown","accessibility":"Open weights (restricted use)"},{"id":"epoch:7263dfbb2c06bcb681f8","name":"MegaScale (175B)","organization":"ByteDance,Peking University","date":"2024-02-23","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2402.15627","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e10681a6329bcbfb2c59","name":"MegaScale (530B)","organization":"ByteDance,Peking University","date":"2024-02-23","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2402.15627","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:424234e6c06fb709aac7","name":"MegaScale (Production)","organization":"ByteDance,Peking University","date":"2024-02-23","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2402.15627","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:3135a6a49fde8a56077b","name":"Genie","organization":"Google DeepMind","date":"2024-02-23","domains":["Video","Games"],"country":"United States of America","reference":"https://arxiv.org/abs/2402.15391","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:621ac470f7eb92be08a2","name":"Stable Diffusion 3","organization":"Stability AI","date":"2024-02-22","domains":["Image generation"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2403.03206\nhttps://stability.ai/news/stable-diffusion-3","confidence":"Confident","accessibility":"API access"},{"id":"epoch:7137fa50b350ad1c0431","name":"Gemma 7B","organization":"Google DeepMind","date":"2024-02-21","domains":["Language"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/gemma/gemma-report.pdf","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:00951c50780e5067026d","name":"Re-Dock","organization":"Zhejiang University (ZJU),Westlake University,University of Washington","date":"2024-02-21","domains":["Biology"],"country":"China,China,United States of America","reference":"https://arxiv.org/abs/2402.11459","confidence":"Likely","accessibility":""},{"id":"epoch:e2260e56972d0008e402","name":"PepGLAD","organization":"Tsinghua University,Renmin University of China","date":"2024-02-21","domains":["Biology"],"country":"China,China","reference":"https://arxiv.org/abs/2402.13555","confidence":"Confident","accessibility":""},{"id":"epoch:5fb2f922d5b0bbf16b41","name":"Reinvent 4","organization":"AstraZeneca","date":"2024-02-21","domains":["Biology"],"country":"United Kingdom,Sweden","reference":"https://link.springer.com/article/10.1186/s13321-024-00812-5","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d4a5cb67acceb9350685","name":"Gemma 2B","organization":"Google DeepMind","date":"2024-02-21","domains":["Language"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/gemma/gemma-report.pdf","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:f6937647536de426f140","name":"Me Llama 70B","organization":"Yale School of Medicine,University of Florida,University of Texas Health Science Center","date":"2024-02-20","domains":["Language","Medicine"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2402.12749","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:5333a412228815691bb4","name":"Me Llama 13B","organization":"Yale School of Medicine,University of Florida,University of Texas Health Science Center","date":"2024-02-20","domains":["Language","Medicine"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2402.12749","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:6817ac8c43312fc4b568","name":"Sora","organization":"OpenAI","date":"2024-02-15","domains":["Video","Vision"],"country":"United States of America","reference":"https://openai.com/index/video-generation-models-as-world-simulators/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:487840e44a4c424bae74","name":"Gemini 1.5 Pro","organization":"Google DeepMind","date":"2024-02-15","domains":["Language","Multimodal"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/gemini/gemini_v1_5_report.pdf","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:b9eddb3414c8f2534b50","name":"ProtChatGPT","organization":"University of Technology Sydney,Zhejiang University (ZJU)","date":"2024-02-15","domains":["Biology"],"country":"Australia,China","reference":"https://arxiv.org/abs/2402.09649","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:02ef8ae02a348962fa82","name":"Gemini 1.0 Pro Vision","organization":"Google DeepMind","date":"2024-02-15","domains":["Vision","Language","Video"],"country":"United States of America","reference":"","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:e865c6a4fb05941d29cc","name":"V-JEPA","organization":"Meta AI","date":"2024-02-15","domains":["Vision"],"country":"United States of America","reference":"https://ai.meta.com/research/publications/revisiting-feature-prediction-for-learning-visual-representations-from-video/","confidence":"Confident","accessibility":""},{"id":"epoch:d374038cce7bbddf8b3a","name":"KwaiYii 175B","organization":"Kuaishou Technology","date":"2024-02-14","domains":["Language"],"country":"China","reference":"https://blog.csdn.net/kuaishoutech/article/details/140542568","confidence":"Confident","accessibility":""},{"id":"epoch:7a04d98913a1acc2adfb","name":"Aya","organization":"Cohere Labs (formerly Cohere for AI),Brown University,Cohere,Carnegie Mellon University (CMU),Massachusetts Institute of Technology (MIT)","date":"2024-02-12","domains":["Language"],"country":"Canada,United States of America,Canada,United States of America,United States of America","reference":"https://arxiv.org/abs/2402.07827","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c0958a29a22228c60bc6","name":"PLAPT","organization":"Wolfram Research,ASC27,Newport High School,Sanskriti School","date":"2024-02-12","domains":["Biology"],"country":"United States of America,Italy,United States of America,India","reference":"https://www.biorxiv.org/content/10.1101/2024.02.08.575577v3.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:aee96601b7293a27663d","name":"Stable Cascade","organization":"Stability AI","date":"2024-02-12","domains":["Image generation"],"country":"United Kingdom","reference":"https://stability.ai/news/introducing-stable-cascade","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:9a31bb1d5dc17f059050","name":"DiscDiff","organization":"Imperial College London","date":"2024-02-08","domains":["Biology"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2402.06079","confidence":"Confident","accessibility":""},{"id":"epoch:6f566d7985ea14064e7b","name":"Distilled Grandmaster","organization":"DeepMind","date":"2024-02-07","domains":["Games"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2402.04494","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e242e0b8ee81c0864d3a","name":"Structure-Informed Protein Language Model","organization":"Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Universit\u00e9 de Montr\u00e9al,IBM Research,HEC Montreal,CIFAR AI Research","date":"2024-02-07","domains":["Biology"],"country":"Canada,Canada,United States of America,Canada,Canada","reference":"https://arxiv.org/abs/2402.05856","confidence":"Likely","accessibility":""},{"id":"epoch:880f131a1249452f8a40","name":"RecGPT / Xingchen LLM (\u6dd8\u5b9d\uff08\u4e2d\u56fd\uff09\u8f6f\u4ef6\u6709\u9650\u516c\u53f8)","organization":"Alibaba","date":"2024-02-07","domains":["Multimodal","Recommendation","Language"],"country":"China","reference":"https://web.archive.org/web/20250707075732/https://finance.sina.com.cn/tech/2025-07-01/doc-infcxsmt5626420.shtml","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:2b8030ce3ab41eb4fb9b","name":"CARP","organization":"Microsoft Research","date":"2024-02-06","domains":["Biology"],"country":"United States of America","reference":"https://www.cell.com/cell-systems/abstract/S2405-4712(24)00029-2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9bf83fef077bf184d059","name":"SenseChat-DataAnalysis V4","organization":"SenseTime","date":"2024-02-06","domains":["Language"],"country":"Hong Kong","reference":"https://www.sensetime.com/en/news-detail/51167493?categoryId=1072","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:547b4a8160ae16193ea0","name":"SenseMirage V4","organization":"SenseTime","date":"2024-02-06","domains":["Image generation"],"country":"Hong Kong","reference":"https://www.sensetime.com/en/news-detail/51167493?categoryId=1072","confidence":"Confident","accessibility":"API access"},{"id":"epoch:26eb249c22bf99716b6f","name":"SenseChat-Medical V4","organization":"SenseTime","date":"2024-02-06","domains":["Language","Vision","Multimodal"],"country":"Hong Kong","reference":"https://www.sensetime.com/en/news-detail/51167493?categoryId=1072\n\nhttps://www.sensetime.com/cn/product-detail?categoryId=51134395&gioNav=1","confidence":"Confident","accessibility":"API access"},{"id":"epoch:3f14b945c897c04bf42c","name":"SenseChat-Vision V4","organization":"SenseTime","date":"2024-02-06","domains":["Vision","Multimodal","Language"],"country":"Hong Kong","reference":"https://www.sensetime.com/en/news-detail/51167493?categoryId=1072","confidence":"Confident","accessibility":"API access"},{"id":"epoch:5200c5a2211697114da1","name":"SenseChat 4.0","organization":"SenseTime","date":"2024-02-06","domains":["Language"],"country":"Hong Kong","reference":"https://www.sensetime.com/en/news-detail/51167493?categoryId=1072","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:1060c25657e5a313629d","name":"Qwen Plus","organization":"Alibaba","date":"2024-02-06","domains":["Language"],"country":"China","reference":"https://help.aliyun.com/zh/model-studio/models?scm=20140722.S_help%40%40%E6%96%87%E6%A1%A3%40%402840914.S_BB1%40bl%2BBB2%40bl%2BRQW%40ag0%2Bos0.ID_2840914-RL_plus-LOC_doc%7EUND%7Eab-OR_ser-PAR1_2102029c17437872399601371d08f3-V_4-P0_0-P1_0&spm=a2c4g.11186623.help-search.i20","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:87e482155f578e26e228","name":"Qwen-Turbo","organization":"Alibaba","date":"2024-02-06","domains":["Language"],"country":"China","reference":"https://help.aliyun.com/zh/model-studio/models?scm=20140722.S_help%40%40%E6%96%87%E6%A1%A3%40%402840914.S_BB1%40bl%2BBB2%40bl%2BRQW%40ag0%2Bos0.ID_2840914-RL_plus-LOC_doc%7EUND%7Eab-OR_ser-PAR1_2102029c17437872399601371d08f3-V_4-P0_0-P1_0&spm=a2c4g.11186623.help-search.i20","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:cb8a786c330c385f6d3a","name":"CausalLM 34B \u03b2","organization":"CausalLM","date":"2024-02-06","domains":["Language"],"country":"China","reference":"https://huggingface.co/CausalLM/34b-beta","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7f4423d89208b66b45c3","name":"Qwen1.5-32B","organization":"Alibaba","date":"2024-02-05","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen1.5/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:ba57690b8a54ddd2bd60","name":"DeepSeekMath 7B","organization":"DeepSeek,Tsinghua University,Peking University","date":"2024-02-05","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/abs/2402.03300","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:2bf6310ba51ccab63c4a","name":"Qwen1.5-72B","organization":"Alibaba","date":"2024-02-04","domains":["Language"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen1.5/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:eacf21ea341ec97e0ad0","name":"Qwen1.5-7B","organization":"Alibaba","date":"2024-02-04","domains":["Language"],"country":"China","reference":"https://huggingface.co/Qwen/Qwen1.5-7B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8f8d7f8b36998576dce8","name":"Qwen1.5-14B","organization":"Alibaba","date":"2024-02-04","domains":["Language"],"country":"China","reference":"https://huggingface.co/Qwen/Qwen1.5-14B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2d6f8146895ed06cc80d","name":"OLMo-7B","organization":"Allen Institute for AI,University of Washington","date":"2024-02-01","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2402.00838v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9cf614f3c8a76d4678d1","name":"OLMo-1B","organization":"Allen Institute for AI,University of Washington","date":"2024-02-01","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2402.00838v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d6c2c9cc2e0585775b01","name":"Hanhai (\u701a\u6d77)","organization":"Shuchi Information Technology ( Shanghai ) Co. , Ltd.","date":"2024-01-31","domains":["Multimodal","Language","Image generation","Speech"],"country":"China","reference":"https://www.payue.com/home/article/show/id/278.html","confidence":"Unknown","accessibility":""},{"id":"epoch:fd664a5d6b68b48b4fff","name":"LLaVA-NeXT-34B (LLaVA-1.6)","organization":"University of Wisconsin Madison,ByteDance,Nanyang Technological University,University of California (UC) Berkeley","date":"2024-01-30","domains":["Multimodal","Language","Vision"],"country":"United States of America,China,Singapore,United States of America","reference":"https://llava-vl.github.io/blog/2024-01-30-llava-next/, https://huggingface.co/liuhaotian/llava-v1.6-34b","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3a80f456b888a697eb51","name":"BGE-M3 Embedding","organization":"Beijing Academy of Artificial Intelligence / BAAI","date":"2024-01-30","domains":["Language"],"country":"China","reference":"https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/BGE_M3/BGE_M3.pdf","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:07a22fd054e672c8d8b0","name":"Code Llama-70B","organization":"Meta AI","date":"2024-01-29","domains":["Language"],"country":"United States of America","reference":"https://ai.meta.com/research/publications/code-llama-open-foundation-models-for-code/\nhttps://arxiv.org/abs/2308.12950","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:8035120ffb523e8fd5f8","name":"Baichuan3","organization":"Baichuan","date":"2024-01-29","domains":["Language","Medicine"],"country":"China","reference":"https://www.jiqizhixin.com/articles/2024-01-29-12","confidence":"Likely","accessibility":"API access"},{"id":"epoch:fb513609a60f9f4f2d02","name":"Karakuri LM","organization":"KARAKURI Inc.","date":"2024-01-26","domains":["Language"],"country":"Japan","reference":"https://huggingface.co/karakuri-ai/karakuri-lm-70b-v0.1","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:782fff3df82207e1a061","name":"ProteinStructureTransformer","organization":"Max Planck Institute of Biochemistry","date":"2024-01-26","domains":["Biology"],"country":"Germany","reference":"https://arxiv.org/abs/2401.14819","confidence":"Confident","accessibility":""},{"id":"epoch:a1470ba086bb47d4aeee","name":"Magic Cube","organization":"","date":"2024-01-26","domains":[],"country":"","reference":"https://en.tmtpost.com/post/6902830","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:fa5eb908c302a7a0e795","name":"Yue AI","organization":"","date":"2024-01-26","domains":[],"country":"","reference":"https://en.tmtpost.com/post/6902830","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:bf143d6715d00ef06eeb","name":"Yan","organization":"Rock AI / Shanghai Stonehill Technology","date":"2024-01-26","domains":["Language"],"country":"China","reference":"https://finance.yahoo.com/news/shanghai-stonehill-technology-unveils-first-084100444.html?guccounter=1&guce_referrer=aHR0cHM6Ly93d3cuZ29vZ2xlLmNvbS8&guce_referrer_sig=AQAAALZVOWhnGXMeBPEqVFKDfgb5yMoB2IZcvVtHDMWMIRNDKB7imuhHqZigPiw0fT76pyLSJCU-3ykJlOUUrlm_E7CjdQow3_XdDst1g0TEVvjCD2W9JPgl4Mmn5COAdqZtRzeFPX7GaHN9sp0PF7JZ85l_6Gu7LD1HnXcTsXJtQQed","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:4e059048bb942acfd9c1","name":"Qwen-VL-Max","organization":"Alibaba","date":"2024-01-25","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://qwenlm.github.io/blog/qwen-vl/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:ff5d419452f9c493093b","name":"text-embedding-3-small","organization":"OpenAI","date":"2024-01-25","domains":["Language"],"country":"United States of America","reference":"https://openai.com/index/new-embedding-models-and-api-updates/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:57822f18bcf92461fea9","name":"text-embedding-3-large","organization":"OpenAI","date":"2024-01-25","domains":["Language"],"country":"United States of America","reference":"https://openai.com/index/new-embedding-models-and-api-updates/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:2a8b1e38a978bbb52b26","name":"Fuyu-Heavy","organization":"Adept","date":"2024-01-24","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://www.adept.ai/blog/adept-fuyu-heavy","confidence":"Speculative","accessibility":"Hosted access (no API)"},{"id":"epoch:ad693605548533e2ddb0","name":"Lumiere","organization":"Google Research,Weizmann Institute of Science,Tel Aviv University,Technion - Israel Institute of Technology","date":"2024-01-23","domains":["Video","Vision"],"country":"United States of America,Israel,Israel,Israel","reference":"https://arxiv.org/abs/2401.12945","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:1815dee60460143e6ce1","name":"Yi-VL-34B","organization":"01.AI","date":"2024-01-23","domains":["Vision","Language","Multimodal"],"country":"China","reference":"https://huggingface.co/01-ai/Yi-VL-34B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:7e94bb77f2754b1d6ecb","name":"voyage-code-2","organization":"Voyage AI","date":"2024-01-23","domains":["Language"],"country":"United States of America","reference":"https://blog.voyageai.com/2024/01/23/voyage-code-2-elevate-your-code-retrieval/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:6dc92ebadbd6c2ad8f07","name":"Prothyena","organization":"Tokyo Institute of Technology","date":"2024-01-22","domains":["Biology"],"country":"Japan","reference":"https://www.biorxiv.org/content/10.1101/2024.01.18.576206v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:417b2946dc8b899b0e88","name":"Orion Star (\u730e\u6237\u661f\u7a7a\u5927\u6a21\u578b)","organization":"Beijing OrionStar Technology Co., Ltd.","date":"2024-01-21","domains":["Language"],"country":"China","reference":"https://huggingface.co/OrionStarAI/Orion-14B-Base/blob/main/README.md","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:139e17426651600c637b","name":"StableLM-2-1.6B","organization":"Stability AI","date":"2024-01-18","domains":["Language"],"country":"United Kingdom","reference":"https://huggingface.co/stabilityai/stablelm-2-1_6b","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:179b2ccdc367647367ef","name":"ESM-NBR","organization":"Hunan University","date":"2024-01-18","domains":["Biology"],"country":"China","reference":"https://ieeexplore.ieee.org/abstract/document/10385509?casa_token=3fdFlCTqsngAAAAA:e2xRjxP3NP4t39elNf9BYR-zw6AWHayrqqFsavdQYXyyVGp7Jpyu62satcyzrum7NQ-MSFK7Ug","confidence":"Confident","accessibility":""},{"id":"epoch:2d8a96d3a26a064170d9","name":"XGPT","organization":"","date":"2024-01-18","domains":[],"country":"","reference":"https://x.threatbook.com/v5/article?threatInfoID=101146","confidence":"Unknown","accessibility":""},{"id":"epoch:724db5e88df0ecf53c96","name":"AlphaGeometry","organization":"Google DeepMind,New York University (NYU)","date":"2024-01-17","domains":["Mathematics"],"country":"United States of America,United States of America","reference":"https://www.nature.com/articles/s41586-023-06747-5","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:61b2bc6ec5e15df955bb","name":"GLM-4 (0116)","organization":"Z.ai (Zhipu AI)","date":"2024-01-17","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2406.12793\nhttps://zhipuai.cn/en/devday","confidence":"Likely","accessibility":"API access"},{"id":"epoch:1fe0ffc5d71ef623ca13","name":"GLM-4","organization":"Z.ai (Zhipu AI)","date":"2024-01-17","domains":["Language","Multimodal","Image generation"],"country":"China","reference":"https://arxiv.org/abs/2406.12793","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:a60f3a44062287a7d237","name":"VideoCrafter2","organization":"Tencent","date":"2024-01-17","domains":["Video","Vision"],"country":"China","reference":"https://arxiv.org/abs/2401.09047","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:60e127f729cacecd254c","name":"Skiff LLM (\u4e00\u53f6\u8f7b\u821f\u5927\u8bed\u8a00\u6a21\u578b)","organization":"Shiyin Intelligent Technology Co., Ltd.","date":"2024-01-17","domains":["Language"],"country":"China","reference":"https://web.archive.org/web/20250624063236/https://www.bjwmb.gov.cn/wmdt/sjsq/10054023.html","confidence":"Confident","accessibility":""},{"id":"epoch:30b4d019b8b0ec184ddd","name":"Rubik (Rubik\u9b54\u65b9\u5927\u6a21\u578b)","organization":"Thunder Software Technology Co.,Ltd.","date":"2024-01-17","domains":[],"country":"China","reference":"https://jp.thundersoft.com/thundersofts-new-year-greetings-for-2024-reshaping-computation-and-enlightening-the-future/?utm_source=chatgpt.com","confidence":"Confident","accessibility":""},{"id":"epoch:402b981eb7eca89171ee","name":"ProteinINR","organization":"Kakao","date":"2024-01-16","domains":["Biology"],"country":"South Korea","reference":"https://openreview.net/forum?id=BEH4mGo7zP","confidence":"Likely","accessibility":""},{"id":"epoch:3b7613339ecc8ff81a9d","name":"DeciCoder-6B","organization":"Deci AI","date":"2024-01-15","domains":["Language"],"country":"Israel","reference":"https://huggingface.co/Deci/DeciCoder-6B","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:bd60872cf0586c1a74e1","name":"OmniNA","organization":"Tianjin Medical University","date":"2024-01-15","domains":["Biology"],"country":"China","reference":"https://www.biorxiv.org/content/10.1101/2024.01.14.575543v1.abstract","confidence":"Likely","accessibility":""},{"id":"epoch:e1a13c3df57cb5647b3c","name":"InternViT-6B","organization":"Shanghai AI Lab,Nanjing University,The University of Hong Kong,Tsinghua University,SenseTime,University of Science and Technology of China (USTC)","date":"2024-01-15","domains":["Vision","Language"],"country":"China,China,Hong Kong,China,Hong Kong,China","reference":"https://arxiv.org/abs/2312.14238","confidence":"Confident","accessibility":""},{"id":"epoch:14fbd426a1f6f58c3a96","name":"Yiye Qingzhou-0.7B","organization":"EFFYIC (\u8bc6\u56e0\u667a\u80fd)","date":"2024-01-15","domains":["Multimodal","Language","Vision","Speech","Video"],"country":"China","reference":"https://36kr.com/p/3062982772761984\nhttps://www.effyic.com/technical_core","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:80351e9110f8d33811aa","name":"Yiye Qingzhou-45B","organization":"EFFYIC (\u8bc6\u56e0\u667a\u80fd)","date":"2024-01-15","domains":["Multimodal","Language","Vision","Speech","Video"],"country":"China","reference":"https://36kr.com/p/3062982772761984\nhttps://www.effyic.com/technical_core","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:7d91b56cbc0eaf56ba20","name":"InternVL","organization":"Shanghai AI Lab,Nanjing University,The University of Hong Kong,Tsinghua University,SenseTime,University of Science and Technology of China (USTC)","date":"2024-01-15","domains":["Vision","Language"],"country":"China,China,Hong Kong,China,Hong Kong,China","reference":"https://arxiv.org/abs/2312.14238","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:006d5d33a95ade7a4de9","name":"InternLM2-20B","organization":"Shanghai AI Lab,SenseTime,Chinese University of Hong Kong (CUHK),Fudan University","date":"2024-01-12","domains":["Language"],"country":"China,Hong Kong,Hong Kong,China","reference":"https://arxiv.org/abs/2403.17297","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:a9b1f5f0b34545d7c47a","name":"DeepSeekMoE-16B","organization":"DeepSeek","date":"2024-01-11","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2401.06066","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:0686b1e0cd8349c1080f","name":"YOLOv8x","organization":"Ultralytics","date":"2024-01-10","domains":["Vision"],"country":"United States of America","reference":"https://yolov8.com/\nhttps://docs.ultralytics.com/models/yolov8\n","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:0eba14f7812284bd18a8","name":"Magic LLM (\u9b54\u6cd5\u5927\u6a21\u578b)","organization":"Shenzhen Honor Software Technology","date":"2024-01-10","domains":["Language"],"country":"China","reference":"https://www.ithome.com/0/744/405.htm","confidence":"Confident","accessibility":""},{"id":"epoch:a6dbc157a2ddf7ace88f","name":"MAGNeT","organization":"Meta AI,Hebrew University of Jerusalem,Kyutai","date":"2024-01-09","domains":["Audio"],"country":"United States of America,Israel,France","reference":"https://arxiv.org/abs/2401.04577, https://github.com/facebookresearch/audiocraft/blob/main/docs/MAGNET.md","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4a6f2b6c5796c8da502f","name":"Stable Code 3B","organization":"Stability AI","date":"2024-01-09","domains":["Language"],"country":"United Kingdom","reference":"https://huggingface.co/stabilityai/stable-code-3b","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:9154e98dbccd4c123d2b","name":"FABind","organization":"Renmin University of China,Huazhong University of Science and Technology,Microsoft Research AI for Science,University of Science and Technology of China (USTC)","date":"2024-01-09","domains":["Biology"],"country":"China,China,United States of America,China","reference":"https://arxiv.org/abs/2310.06763v5","confidence":"Likely","accessibility":""},{"id":"epoch:58e3d20f3c4ecaf31bea","name":"Improved motif-scaffolding with SE(3) flow matching","organization":"University of Oxford,Massachusetts Institute of Technology (MIT),Microsoft Research AI for Science","date":"2024-01-08","domains":["Biology"],"country":"United Kingdom,United States of America,United States of America","reference":"https://arxiv.org/abs/2401.04082","confidence":"Confident","accessibility":""},{"id":"epoch:059ad8901f5703cf5226","name":"DeepSeek LLM 67B","organization":"DeepSeek","date":"2024-01-05","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2401.02954, https://github.com/deepseek-ai/DeepSeek-LLM","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:7e8605f049f4cbeb1486","name":"DeepSeek-LLM-1.3b-base","organization":"DeepSeek","date":"2024-01-05","domains":["Language"],"country":"China","reference":"it is only mentioned in Janus 1.3B release https://huggingface.co/deepseek-ai/Janus-1.3B\n\nupd + supposedly in Janus-Pro-1B paper \"In our experiments, we utilize DeepSeek-LLM (1.5B and 7B) [3]\" while there is no 1.5B model mentioned in the linked paper","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:ea4524f43e1a68a3e7b4","name":"DeepSeek LLM 7B","organization":"DeepSeek","date":"2024-01-05","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2401.02954, https://github.com/deepseek-ai/DeepSeek-LLM","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:76c7351a8a9b12263d50","name":"Xingrui AI (\u661f\u777fAI)","organization":"Geely Automobile Research Institute (Ningbo) Company","date":"2024-01-05","domains":["Robotics","Multimodal","Language","3D modeling"],"country":"China","reference":"https://web.archive.org/web/20240111035331/https://m.tech.china.com/redian/2024/0111/012024_1469664.html","confidence":"Likely","accessibility":""},{"id":"epoch:1f66ed8b32ca1c183292","name":"RT-1 + AutoRT","organization":"Google DeepMind","date":"2024-01-04","domains":["Robotics"],"country":"United States of America","reference":"https://auto-rt.github.io/\nhttps://arxiv.org/abs/2401.12963","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:501d473abba3137c329a","name":"babbage-002","organization":"OpenAI","date":"2024-01-04","domains":["Language"],"country":"United States of America","reference":"","confidence":"Unknown","accessibility":""},{"id":"epoch:be5c099f3d7a2c16bb7f","name":"CLAPE-DB","organization":"Tsinghua University","date":"2024-01-03","domains":["Biology"],"country":"China","reference":"https://academic.oup.com/bib/article/25/1/bbad488/7505238","confidence":"Confident","accessibility":""},{"id":"epoch:a5d842fc1158d4984bba","name":"PLLaMa","organization":"University of California Santa Barbara (UCSB),University of Lincoln,Chinese Academy of Agricultural Sciences,Swedish University of Agricultural Sciences","date":"2024-01-03","domains":["Biology"],"country":"United States of America,United Kingdom,China,Sweden","reference":"https://arxiv.org/abs/2401.01600","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ed160b038850305e04d0","name":"Palmyra X 003","organization":"Writer","date":"2024-01-01","domains":["Language"],"country":"United States of America","reference":"https://writer.com/llms/palmyra-x-003/","confidence":"","accessibility":"API access"},{"id":"epoch:648348c4052cc9a7976c","name":"Kimi Explorer","organization":"Moonshot","date":"2024-01-01","domains":["Language"],"country":"China","reference":"https://www.53ai.com/news/LargeLanguageModel/2024101137012.html","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:b25ae12190a3539642e2","name":"Qarasu-14B","organization":"Lightblue","date":"2023-12-29","domains":["Language"],"country":"Japan","reference":"https://note.com/peter_lightblue/n/ne08a7c8cc47a, https://huggingface.co/lightblue/qarasu-14B-chat-plus-unleashed","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:14d719a2e29eed7ea0c0","name":"CoRe","organization":"Tsinghua University","date":"2023-12-29","domains":["Mathematics","Language"],"country":"China","reference":"https://arxiv.org/abs/2210.16257","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:794d5b04c57a86a28240","name":"Mengzi-Code-6.7B","organization":"Langboat","date":"2023-12-28","domains":["Language"],"country":"China","reference":"https://www.langboat.com/en/portal/mengzi-gpt","confidence":"Confident","accessibility":""},{"id":"epoch:bf865ae7db728e0c0e96","name":"MengziGPT-General-13B","organization":"Langboat","date":"2023-12-28","domains":["Language"],"country":"China","reference":"https://hub.baai.ac.cn/view/33895","confidence":"Confident","accessibility":""},{"id":"epoch:4111fd19615c8458eef3","name":"Elyza","organization":"Elyza","date":"2023-12-27","domains":["Language"],"country":"Japan","reference":"https://huggingface.co/elyza/ELYZA-japanese-Llama-2-13b","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:03b5232e9316c6d6799d","name":"Zhuhai-13B","organization":"Zhujian Intelligence","date":"2023-12-27","domains":[],"country":"China","reference":"https://github.com/emotibot-inc/Zhuhai-13B","confidence":"Confident","accessibility":""},{"id":"epoch:1c1a0d8e1bb1063d8279","name":"Nous-Hermes-2-Yi-34B","organization":"Nous Research","date":"2023-12-25","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/NousResearch/Nous-Hermes-2-Yi-34B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:421bb0eb1d1d7b3eca24","name":"Solar-10.7B (Solar Mini)","organization":"Upstage","date":"2023-12-23","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/abs/2312.15166","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:70129f96dc1e27d47ae1","name":"GQA-8-XXL","organization":"Google Research","date":"2023-12-23","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2305.13245","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:aeb6016a710629598e21","name":"YaYi 2.0","organization":"Yayi (Wenge)","date":"2023-12-22","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2312.14862v1","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:d87aa9983a880fd0da24","name":"Fulu Gua (\u798f\u7984\u74dc)","organization":"ByteDance","date":"2023-12-22","domains":["Image generation"],"country":"China","reference":"https://web.archive.org/web/20250708153744/https://dtzed.com/issues/2024/04/10116/","confidence":"Unknown","accessibility":""},{"id":"epoch:1ddfd8b4098b634dcc42","name":"VideoPoet","organization":"Google Research,Carnegie Mellon University (CMU),Google DeepMind","date":"2023-12-21","domains":["Video","Language","Audio"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2312.14125\nhttps://sites.research.google/videopoet/","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:7e2debc8a10324b85e27","name":"nekomata-14b","organization":"rinna","date":"2023-12-21","domains":["Language"],"country":"Japan","reference":"https://arxiv.org/abs/2404.01657","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:d8ed45375be2976dfc5a","name":"Suno Music Generation","organization":"Suno","date":"2023-12-20","domains":["Audio"],"country":"United States of America","reference":"https://suno.com/about","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:76b4b8e0c89524198761","name":"Gemini Nano-2","organization":"Google DeepMind","date":"2023-12-19","domains":["Multimodal","Language","Vision","Audio"],"country":"United States of America","reference":"https://arxiv.org/abs/2312.11805","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0996e407e6d9e17c69e4","name":"Gemini Nano-1","organization":"Google DeepMind","date":"2023-12-19","domains":["Multimodal","Language","Vision","Audio"],"country":"United States of America","reference":"https://arxiv.org/abs/2312.11805","confidence":"Confident","accessibility":"API access"},{"id":"epoch:bb3d65fe95ce1e8818cf","name":"HiFi - NN","organization":"Basecamp Research,Technical University of Munich,Molecular Institute of Biology,Microsoft Research","date":"2023-12-19","domains":["Biology"],"country":"United Kingdom,Germany,Spain,United States of America","reference":"https://developer.nvidia.com/blog/breakthrough-in-functional-annotation-with-hifi-nn/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3628bbfe658c311359bb","name":"Bird Vocalization Classifier (Perch)","organization":"Google Research,Cornell University,Naturalis Biodiversity Center,Chemnitz University of Technology","date":"2023-12-18","domains":["Audio"],"country":"United States of America,United States of America,Germany,Germany","reference":"https://www.nature.com/articles/s41598-023-49989-z","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d991138e7b937c792e7f","name":"Lyra-Fr 10B","organization":"LightOn","date":"2023-12-15","domains":["Language"],"country":"France","reference":"https://aws.amazon.com/fr/blogs/machine-learning/lighton-lyra-fr-model-is-now-available-on-amazon-sagemaker/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:0217bce51ea9131953e4","name":"Konan LLM 41B","organization":"Konan Technology","date":"2023-12-15","domains":["Language","Vision"],"country":"South Korea","reference":"https://en.konantech.com/en/llm/konanllm\nhttps://techfinch.kr/ai/konan-technology-unveils-konan-llm--its-own-ai-language-model\nhttps://www.dbpia.co.kr/journal/articleDetail?nodeId=NODE11610127 ","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:1bea31a54d1d62dfec21","name":"Poro 34B","organization":"High-Performance Language Technologies (HPLT),University of Turku","date":"2023-12-14","domains":["Language"],"country":"Multinational,Finland","reference":"https://arxiv.org/abs/2404.01856","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:73129301d76282ccf14c","name":"CogAgent","organization":"Tsinghua University,Z.ai (Zhipu AI)","date":"2023-12-14","domains":["Vision","Language"],"country":"China,China","reference":"https://arxiv.org/abs/2312.08914","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:4404f100f1e3249dd19f","name":"FunSearch","organization":"Google DeepMind","date":"2023-12-14","domains":["Language","Search"],"country":"United States of America","reference":"https://www.nature.com/articles/s41586-023-06924-6\nhttps://deepmind.google/discover/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:74ff0f4be93776ad376c","name":"Imagen 2","organization":"Google DeepMind","date":"2023-12-13","domains":["Image generation"],"country":"United States of America","reference":"https://deepmind.google/technologies/imagen-2/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:b3d1a0cf0f54f77ff4cb","name":"GigaChat Pro","organization":"Sber","date":"2023-12-13","domains":["Language"],"country":"Russia","reference":"https://habr.com/ru/companies/sberdevices/articles/780334/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:c3b4bd27f32010cb01c0","name":"MedLM","organization":"Google Cloud","date":"2023-12-13","domains":["Language","Medicine"],"country":"United States of America","reference":"https://cloud.google.com/blog/topics/healthcare-life-sciences/introducing-medlm-for-the-healthcare-industry","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:359b5085fc52b5200b9b","name":"Phi-2","organization":"Microsoft","date":"2023-12-12","domains":["Language"],"country":"United States of America","reference":"https://www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/\nhttps://huggingface.co/microsoft/phi-2 ","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:26c48a78261dca3983d2","name":"VILA-13B","organization":"Nvidia,Massachusetts Institute of Technology (MIT)","date":"2023-12-12","domains":["Multimodal","Language","Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2312.07533\nhttps://huggingface.co/Efficient-Large-Model/VILA-13b","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:312acfa943570979530d","name":"Mixtral 8x7B","organization":"Mistral AI","date":"2023-12-11","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/mixtral-of-experts/, https://arxiv.org/abs/2401.04088","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e5487fe3ff1951330fd8","name":"Mistral Medium","organization":"Mistral AI","date":"2023-12-11","domains":["Language"],"country":"France","reference":"https://mistral.ai/news/la-plateforme/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:704feff601e8379cbb35","name":"ruDalle: Kandinsky 3.0","organization":"Sber","date":"2023-12-11","domains":["Image generation"],"country":"Russia","reference":"https://arxiv.org/abs/2312.03511","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:81e383ee15b61b7cff2b","name":"ruDalle: Kandinsky 3.1","organization":"Sber","date":"2023-12-11","domains":["Image generation"],"country":"Russia","reference":"https://ai-forever.github.io/Kandinsky-3/K31/","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:182795d2d0be30e1fb97","name":"MBP","organization":"University of Science and Technology of China (USTC),Tencent,Zhejiang University (ZJU)","date":"2023-12-11","domains":["Biology"],"country":"China,China,China","reference":"https://academic.oup.com/bib/article/25/1/bbad451/7469349","confidence":"Likely","accessibility":""},{"id":"epoch:86fe950af646df44a020","name":"CRYSTALCODER","organization":"Mohamed bin Zayed University of Artificial Intelligence (MBZUAI),Petuum,University of Southern California,Carnegie Mellon University (CMU),University of Illinois Urbana-Champaign (UIUC),University of California San Diego,LLM360","date":"2023-12-11","domains":["Language"],"country":"United Arab Emirates,United States of America,United States of America,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2312.06550","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2b58239877343ea1ac4f","name":"Amber","organization":"Mohamed bin Zayed University of Artificial Intelligence (MBZUAI),Petuum,University of Southern California,Carnegie Mellon University (CMU),University of Illinois Urbana-Champaign (UIUC),University of California San Diego,LLM360","date":"2023-12-11","domains":["Language"],"country":"United Arab Emirates,United States of America,United States of America,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2312.06550","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:344e62be2952102a36d6","name":"W.A.L.T","organization":"Stanford University,Google Research,Georgia Institute of Technology","date":"2023-12-11","domains":["Video"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2312.06662","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:388d9ee302472e3a2e54","name":"XVERSE-65B-2","organization":"XVERSE Technology,Shenzhen Yuanxiang Technology","date":"2023-12-08","domains":["Language"],"country":"China,China","reference":"https://github.com/xverse-ai/XVERSE-65B/blob/main/README_EN.md","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:2a0c49ae5e2ab605c2fc","name":"SeamlessM4T","organization":"Facebook,INRIA,University of California (UC) Berkeley","date":"2023-12-08","domains":["Speech","Language"],"country":"United States of America,France,United States of America","reference":"https://arxiv.org/abs/2312.05187, https://huggingface.co/facebook/seamless-m4t-v2-large","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:abb0d0a40fd744723c8b","name":"Llama Guard","organization":"Meta AI","date":"2023-12-07","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2312.06674","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:38c6e92f48b51b819459","name":"StableLM - Zephyr 3B","organization":"Stability AI","date":"2023-12-07","domains":["Language"],"country":"United Kingdom","reference":"https://huggingface.co/stabilityai/stablelm-zephyr-3b","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:33f86dda008b131b8570","name":"Gemini 1.0 Ultra","organization":"Google DeepMind","date":"2023-12-06","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/gemini/gemini_1_report.pdf","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:81efa4888c4be3e954ab","name":"Gemini 1.0 Pro","organization":"Google DeepMind","date":"2023-12-06","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://storage.googleapis.com/deepmind-media/gemini/gemini_1_report.pdf","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:e56cf1d2793ac39fe3da","name":"OneLLM","organization":"Chinese University of Hong Kong (CUHK),Shanghai AI Lab","date":"2023-12-06","domains":["Multimodal","Language","Vision"],"country":"Hong Kong,China","reference":"https://arxiv.org/abs/2312.03700","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:6db0b9bdad06a87edd96","name":"NexusRaven-V2","organization":"Nexusflow","date":"2023-12-05","domains":["Language"],"country":"United States of America","reference":"https://nexusflow.ai/blogs/ravenv2","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:3026d97159ca087ddb9d","name":"SARA-RT-2","organization":"Google DeepMind","date":"2023-12-04","domains":["Robotics"],"country":"United States of America","reference":"https://arxiv.org/abs/2312.01990","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:8b31adb34cc7d758197a","name":"Playground v2","organization":"Playground","date":"2023-12-02","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2402.17245v1","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3d6f2ca8dcfb30712b22","name":"Baize-v2-13B (\u767d\u6cfd)","organization":"University of California San Diego,Sun Yat-sen University,Microsoft Research Asia","date":"2023-12-02","domains":["Language"],"country":"United States of America,China,China","reference":"https://arxiv.org/abs/2304.01196v4","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:72d44bcbc69ef44ebcea","name":"Mamba-24M (SC09)","organization":"Carnegie Mellon University (CMU),Princeton University","date":"2023-12-01","domains":["Speech"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2312.00752","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:fd5a00be1c6989de2f6e","name":"Mamba-2.8B","organization":"Carnegie Mellon University (CMU),Princeton University","date":"2023-12-01","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2312.00752","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fbdfd91207d693500bcc","name":"tsuzumi 7B","organization":"NTT Communication Science Laboratories","date":"2023-12-01","domains":["Language"],"country":"Japan","reference":"https://group.ntt/en/magazine/blog/tsuzumi/","confidence":"Confident","accessibility":""},{"id":"epoch:8f43ad7dcd07de997d2c","name":"NASA SMD","organization":"NASA,IBM","date":"2023-12-01","domains":["Language"],"country":"United States of America,United States of America","reference":"https://huggingface.co/nasa-impact/nasa-smd-ibm-v0.1","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e902079582a60240feb1","name":"SEA-LION V1 3B","organization":"AI Singapore","date":"2023-12-01","domains":["Language"],"country":"Singapore","reference":"https://huggingface.co/aisingapore/sea-lion-3b","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ee0bb12fafe87be0eac9","name":"SEA-LION V1 7B","organization":"AI Singapore","date":"2023-12-01","domains":["Language"],"country":"Singapore","reference":"https://huggingface.co/aisingapore/sea-lion-7b","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c3b52214d09d2def1a32","name":"SeaLLM-7B-v2","organization":"Alibaba DAMO Academy","date":"2023-12-01","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2312.00738","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:a26fd26a21e9c03bd080","name":"SeaLLM-7B-v2.5","organization":"Alibaba DAMO Academy","date":"2023-12-01","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2312.00738","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1cf5e7dea6c62ada75d9","name":"AzeroGPT","organization":"SoundAI","date":"2023-12-01","domains":["Speech","Multimodal","Language"],"country":"China","reference":"https://web.archive.org/web/20250918090911/https://zhuanlan.zhihu.com/p/636393893","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:1b443c5f537675182189","name":"Qwen-72B","organization":"Alibaba","date":"2023-11-30","domains":["Language"],"country":"China","reference":"https://huggingface.co/Qwen/Qwen-72B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:9bac24406ff6b3964f0d","name":"Granite 13B","organization":"IBM","date":"2023-11-30","domains":["Language"],"country":"United States of America","reference":"https://www.ibm.com/downloads/cas/X9W4O6BM","confidence":"Likely","accessibility":"API access"},{"id":"epoch:c6e738a4e46adbdd22b3","name":"Cohere Command Light","organization":"Cohere","date":"2023-11-30","domains":["Language"],"country":"Canada","reference":"https://txt.cohere.com/embed-command-light-fine-tuning-on-amazon-bedrock/","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:d3a8ac293c9b76fee0be","name":"GNoME for crystal discovery","organization":"Google DeepMind","date":"2023-11-29","domains":["Materials science"],"country":"United States of America","reference":"https://www.nature.com/articles/s41586-023-06735-9","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:605abdfd563356358729","name":"PPLX-70B-Online","organization":"Perplexity","date":"2023-11-29","domains":["Language"],"country":"United States of America","reference":"https://blog.perplexity.ai/blog/introducing-pplx-online-llms","confidence":"Likely","accessibility":"API access"},{"id":"epoch:1e8f29024e6aaad4f10c","name":"Amazon Titan Text Express","organization":"Amazon","date":"2023-11-29","domains":["Language"],"country":"United States of America","reference":"https://aws.amazon.com/ru/about-aws/whats-new/2023/11/amazon-titan-models-express-lite-bedrock/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:6a12f7692e0394548933","name":"Amazon Titan Text Lite","organization":"Amazon","date":"2023-11-29","domains":["Language"],"country":"United States of America","reference":"https://aws.amazon.com/ru/about-aws/whats-new/2023/11/amazon-titan-models-express-lite-bedrock/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:ea3afea64935e0de3e42","name":"Llama-3-Taiwan-70B","organization":"National Taiwan University","date":"2023-11-29","domains":["Language"],"country":"Taiwan","reference":"https://arxiv.org/abs/2311.17487","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:9cde401dc02c584e05f8","name":"SD-Turbo","organization":"Stability AI","date":"2023-11-28","domains":["Image generation"],"country":"United Kingdom","reference":"https://huggingface.co/stabilityai/sd-turbo","confidence":"Unknown","accessibility":"Open weights (non-commercial)"},{"id":"epoch:cf4b9ba6f73e7323fa70","name":"Yuan 2.0","organization":"Inspur","date":"2023-11-27","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2311.15786v1","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:9399bf9e93a6df9c64d5","name":"StripedHyena-Hessian-7B","organization":"Together,Nous Research","date":"2023-11-27","domains":["Language"],"country":"United States of America,United States of America","reference":"https://huggingface.co/togethercomputer/StripedHyena-Hessian-7B\nhttps://github.com/togethercomputer/stripedhyena\nhttps://www.together.ai/blog/stripedhyena-7b","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c1c660fd7803bc24a006","name":"Meditron-70B","organization":"Ecole Polytechnique F\u00b4ed\u00b4erale de Lausanne (EPFL)","date":"2023-11-27","domains":["Language","Medicine"],"country":"Switzerland","reference":"https://arxiv.org/abs/2311.16079","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:a008f97e9c0e44fea4fc","name":"Amazon Transcribe","organization":"Amazon","date":"2023-11-26","domains":["Speech"],"country":"United States of America","reference":"https://aws.amazon.com/ru/blogs/machine-learning/amazon-transcribe-announces-a-new-speech-foundation-model-powered-asr-system-that-expands-support-to-over-100-languages/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:ec053ba28767edafe65a","name":"ControlNet (SDv2)","organization":"Stability AI","date":"2023-11-26","domains":["Image generation"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2302.05543","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:78e126bcdf785863c4c1","name":"Starling-LM-7B-alpha","organization":"University of California (UC) Berkeley","date":"2023-11-25","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/berkeley-nest/Starling-LM-7B-alpha\nhttps://starling.cs.berkeley.edu/","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:e1f967b5a7e8048f9777","name":"Stable Video Diffusion","organization":"Stability AI","date":"2023-11-25","domains":["Image generation","Video"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2311.15127","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:32e1f22230bb9ce8f95d","name":"Belle VL","organization":"KE Holdings Inc. (\u201cBeike\u201d)","date":"2023-11-24","domains":["Language","Vision","Multimodal"],"country":"China","reference":"https://huggingface.co/BELLE-2/BELLE-VL","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ac3e08fa4ecc73e78d1e","name":"Infinity (\u65e0\u6daf)","organization":"Transwarp Technology","date":"2023-11-24","domains":["Language"],"country":"China","reference":"https://www.infoq.cn/article/gsrgwanes3bmq1ffrxb8","confidence":"Unknown","accessibility":""},{"id":"epoch:3fd98e2eeae24c30b540","name":"TAIWAN-LLM 13B","organization":"National Taiwan University","date":"2023-11-23","domains":["Language"],"country":"Taiwan","reference":"https://arxiv.org/pdf/2311.17487","confidence":"Confident","accessibility":""},{"id":"epoch:687c553ba453260d287c","name":"TAIWAN-LLM 7B","organization":"National Taiwan University","date":"2023-11-23","domains":["Language"],"country":"Taiwan","reference":"https://arxiv.org/pdf/2311.17487","confidence":"Confident","accessibility":""},{"id":"epoch:a60d37abb44d6c436248","name":"Inflection-2","organization":"Inflection AI","date":"2023-11-22","domains":["Language"],"country":"United States of America","reference":"https://inflection.ai/inflection-2","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:28f37efd0e4339b0a939","name":"OmniFusion-7B (InternViT-6B-448px V1-2)","organization":"AIRI Artificial Intelligence Research Institute,Sber,Skolkovo Institute of Science and Technology","date":"2023-11-22","domains":["Multimodal","Vision","Language"],"country":"Russia,Russia,Russia","reference":"https://arxiv.org/abs/2404.06212","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fe248e627a2df7a60624","name":"Claude 2.1","organization":"Anthropic","date":"2023-11-21","domains":["Language"],"country":"United States of America","reference":"https://www.anthropic.com/index/claude-2-1","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:4662212122979fd77916","name":"Orca 2-13B","organization":"Microsoft Research","date":"2023-11-21","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2311.11045, https://huggingface.co/microsoft/Orca-2-13b","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:5867ebdec63959a17a34","name":"Tulu V2 DPO 70B","organization":"Allen Institute for AI,University of Washington","date":"2023-11-20","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2311.10702","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:f74827252db73bd99dda","name":"StyleTTS 2","organization":"Columbia University","date":"2023-11-20","domains":["Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2306.07691","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f950d9c3e62c35007703","name":"Lyria","organization":"Google DeepMind","date":"2023-11-16","domains":["Audio"],"country":"United States of America","reference":"https://deepmind.google/discover/blog/transforming-the-future-of-music-creation/","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:f5d3ec563fd7ca8fd952","name":"Mistral 7B + OVM","organization":"Chinese University of Hong Kong (CUHK)","date":"2023-11-16","domains":["Language"],"country":"Hong Kong","reference":"https://arxiv.org/abs/2311.09724","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:82da179f7b17e6480094","name":"AndesGPT","organization":"Oppo Mobile Telecommunications","date":"2023-11-16","domains":["Language"],"country":"China","reference":"https://www.oppo.com/en/newsroom/press/2023-oppo-developers-conference-odc23/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:1e46659c9818df0bc95c","name":"Nemotron-3-8B","organization":"Nvidia","date":"2023-11-15","domains":["Language"],"country":"United States of America","reference":"https://developer.nvidia.com/blog/nvidia-ai-foundation-models-build-custom-enterprise-chatbots-and-co-pilots-with-production-ready-llms/\n\nhttps://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/nemotron-3-8b-base-4k","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:21c1f8dc49fc8aef4c6f","name":"Mi:dm 7B","organization":"KT","date":"2023-11-15","domains":["Language"],"country":"South Korea","reference":"https://huggingface.co/KT-AI/midm-bitext-S-7B-inst-v1","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:c8b32194e1ed19dd261c","name":"GraphCast","organization":"Google DeepMind","date":"2023-11-14","domains":["Earth science"],"country":"United States of America","reference":"https://www.science.org/doi/epdf/10.1126/science.adi2336\n\nhttps://arxiv.org/abs/2212.12794","confidence":"Speculative","accessibility":"Open weights (non-commercial)"},{"id":"epoch:14384b1d89b7b1edefda","name":"Qwen-Audio-Chat","organization":"Alibaba","date":"2023-11-14","domains":["Language","Speech","Audio"],"country":"China","reference":"https://arxiv.org/abs/2311.07919","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:aa9345a7d8aab9333ddf","name":"SPHINX (Llama 2 13B)","organization":"Shanghai AI Lab,Chinese University of Hong Kong (CUHK),ShanghaiTech University","date":"2023-11-13","domains":["Vision","Language","Multimodal"],"country":"China,Hong Kong,China","reference":"https://arxiv.org/abs/2311.07575","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:2f55728f4f181e6683ed","name":"Volcano 13B","organization":"Korea University,Korea Advanced Institute of Science and Technology (KAIST),LG","date":"2023-11-13","domains":["Language","Multimodal","Vision"],"country":"South Korea,South Korea,South Korea","reference":"https://arxiv.org/abs/2311.07362","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:c6508e3e65dacd262ea6","name":"LLaVA + LVIS-INSTRUCT4V","organization":"Fudan University,University of Maryland","date":"2023-11-13","domains":["Multimodal","Language","Vision"],"country":"China,United States of America","reference":"https://arxiv.org/abs/2311.07574","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0a0a3a85c6ac1f349319","name":"Intel Aurora 1T","organization":"Intel,Argonne National Laboratory","date":"2023-11-13","domains":["Language"],"country":"United States of America,United States of America","reference":"https://www.tomshardware.com/news/intel-supercomputing-2023-aurora-xeon-max-gpu-gaudi-granite-rapids\n\nhttps://newsroom.intel.com/artificial-intelligence/intel-supercomputing-2023-news","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:5ebbd510bf90f03109d2","name":"Stable Diffusion 1.6","organization":"Stability AI","date":"2023-11-10","domains":["Image generation"],"country":"United Kingdom","reference":"https://platform.stability.ai/docs/release-notes","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:e7ccad62d2c30f882cf2","name":"tts-1","organization":"OpenAI","date":"2023-11-09","domains":["Speech"],"country":"United States of America","reference":"https://platform.openai.com/docs/models/tts-1","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:94cfeea0e7b12d74e034","name":"tts-1-hd","organization":"OpenAI","date":"2023-11-09","domains":["Speech"],"country":"United States of America","reference":"https://platform.openai.com/docs/models/tts-1-hd","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:b9fc9d1447a4de6faa93","name":"MultiBand Diffusion","organization":"Meta AI,Hebrew University of Jerusalem,LORIA","date":"2023-11-08","domains":["Audio","Speech"],"country":"United States of America,Israel,France","reference":"https://arxiv.org/abs/2308.02560","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:05fdfe12d93c69b6a705","name":"Samsung Gauss Language","organization":"Samsung","date":"2023-11-08","domains":["Language"],"country":"South Korea","reference":"https://techcrunch.com/2023/11/08/samsung-unveils-chatgpt-alternative-samsung-gauss-that-can-generate-text-code-and-images/","confidence":"Unknown","accessibility":""},{"id":"epoch:2cef890e999f243be30b","name":"Samsung Gauss Code","organization":"Samsung","date":"2023-11-08","domains":["Language"],"country":"South Korea","reference":"https://techcrunch.com/2023/11/08/samsung-unveils-chatgpt-alternative-samsung-gauss-that-can-generate-text-code-and-images/","confidence":"Unknown","accessibility":""},{"id":"epoch:6795d26c9560169f31ad","name":"Samsung Gauss Image","organization":"Samsung","date":"2023-11-08","domains":["Image generation"],"country":"South Korea","reference":"https://techcrunch.com/2023/11/08/samsung-unveils-chatgpt-alternative-samsung-gauss-that-can-generate-text-code-and-images/","confidence":"Unknown","accessibility":""},{"id":"epoch:f6011b1db0687a9a3087","name":"Prithvi-100M","organization":"IBM,NASA","date":"2023-11-08","domains":["Earth science"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2310.18660","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1c6e4d69e197b1df517f","name":"HGRN 1B (WT 103)","organization":"Shanghai AI Lab,Massachusetts Institute of Technology (MIT)","date":"2023-11-08","domains":["Language"],"country":"China,United States of America","reference":"https://arxiv.org/abs/2311.04823","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:de7f1342c15df1fbe560","name":"Jais-30b (phase 1)","organization":"Cerebras Systems,Mohamed bin Zayed University of Artificial Intelligence (MBZUAI),Inception G42,G42","date":"2023-11-08","domains":["Language"],"country":"United States of America,United Arab Emirates,United Arab Emirates,United Arab Emirates","reference":"https://www.g42.ai/resources/publications/Jais-30B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:bceb9bd684413df52366","name":"RoFormer","organization":"Zhuiyi Technology","date":"2023-11-08","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2104.09864","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:d854e5061cde529c332b","name":"OmniVec","organization":"TensorTour","date":"2023-11-07","domains":["Multimodal","Vision","Speech","Language","Video","3D modeling"],"country":"United States of America","reference":"https://arxiv.org/abs/2311.05709v1","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:0bf710f6084bb79e1fbf","name":"mPLUG-Owl2","organization":"Alibaba","date":"2023-11-07","domains":["Vision","Language","Multimodal"],"country":"China","reference":"https://arxiv.org/abs/2311.04257","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:51b56cecd19ac9e601a7","name":"OtterHD-8B","organization":"Nanyang Technological University","date":"2023-11-07","domains":["Multimodal","Vision","Language"],"country":"Singapore","reference":"https://arxiv.org/abs/2311.04219","confidence":"Confident","accessibility":""},{"id":"epoch:f931e0be7f28b9aee39f","name":"XVERSE-13B-2","organization":"XVERSE Technology,Shenzhen Yuanxiang Technology","date":"2023-11-06","domains":["Language"],"country":"China,China","reference":"https://huggingface.co/xverse/XVERSE-13B","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:317b21f9cff281456b13","name":"Whisper v3","organization":"OpenAI","date":"2023-11-06","domains":["Speech"],"country":"United States of America","reference":"https://huggingface.co/openai/whisper-large-v3","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ee22c2635da62daeddf5","name":"Consistency Decoder","organization":"OpenAI","date":"2023-11-06","domains":["Image generation"],"country":"United States of America","reference":"https://github.com/openai/consistencydecoder","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1e12fa248a3807af6b2c","name":"GPT-4 Turbo (Nov 2023)","organization":"OpenAI","date":"2023-11-06","domains":["Multimodal","Vision","Language","Image generation"],"country":"United States of America","reference":"https://openai.com/blog/new-models-and-developer-products-announced-at-devday","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:b88d584a5665cf157729","name":"CogVLM-17B","organization":"Tsinghua University,Z.ai (Zhipu AI),Beihang University","date":"2023-11-06","domains":["Multimodal","Vision","Language"],"country":"China,China,China","reference":"https://arxiv.org/abs/2311.03079\nhttps://huggingface.co/THUDM/cogvlm-chat-hf\nhttps://github.com/THUDM/CogVLM\n","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:4557b0c6bb2522c51a79","name":"RNA-MSM","organization":"Peking University,Shanghai AI Lab,Griffith University,Peng Cheng Laboratory,Shenzhen Bay Laboratory","date":"2023-11-06","domains":["Biology"],"country":"China,China,Australia,China,China","reference":"https://academic.oup.com/nar/article/52/1/e3/7369930","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0452cf8c28ad40311ebb","name":"LLaVA 1.5","organization":"University of Wisconsin Madison,Microsoft Research","date":"2023-11-05","domains":["Multimodal","Language","Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2310.03744,\n","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:c78f2668c8ec706ec290","name":"Grok-1","organization":"xAI","date":"2023-11-04","domains":["Language"],"country":"United States of America","reference":"https://x.ai/model-card/, https://x.ai/blog/grok-os","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f4e7152286924de00404","name":"Grok-0","organization":"xAI","date":"2023-11-04","domains":["Language"],"country":"United States of America","reference":"https://x.ai/","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:de51036c90787f420a1f","name":"Qiyuan","organization":"","date":"2023-11-04","domains":[],"country":"","reference":"https://www.sohu.com/a/733914132_104036","confidence":"Unknown","accessibility":""},{"id":"epoch:f9c8ba25cc89f47a2cd3","name":"Sequence Monkey (\u5e8f\u5217\u7334\u5b50)","organization":"Mobvoi","date":"2023-11-04","domains":["Language"],"country":"China","reference":"https://zhuanlan.zhihu.com/p/665027266","confidence":"Unknown","accessibility":""},{"id":"epoch:1093feba8f9ccfcca331","name":"ZhihaiTu AI (\u77e5\u6d77\u56fe)","organization":"Beijing Zhizhe Tianxia Technology","date":"2023-11-04","domains":["Language"],"country":"China","reference":"https://news.qq.com/rain/a/20230413A080TY00\nhttps://www.qbitai.com/2023/04/46542.html","confidence":"Unknown","accessibility":""},{"id":"epoch:964f4f91658d4b1544b2","name":"BLUUMI","organization":"University of Turku,Hugging Face","date":"2023-11-03","domains":["Language"],"country":"Finland,United States of America","reference":"https://arxiv.org/abs/2311.05640","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:ffc86059745982be77a0","name":"RT-Trajectory","organization":"Google DeepMind,University of California San Diego,Stanford University","date":"2023-11-03","domains":["Robotics"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2311.01977","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:ead7b1dc11f69f32f684","name":"DeepSeek Coder 33B","organization":"DeepSeek,Peking University","date":"2023-11-02","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2401.14196\n\nhttps://huggingface.co/deepseek-ai/deepseek-coder-33b-base ","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:2a3898913468339977dc","name":"Yi-34B","organization":"01.AI","date":"2023-11-02","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2403.04652","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:25236b37dcc25056f00a","name":"Cohere Embed","organization":"Cohere","date":"2023-11-02","domains":["Language"],"country":"Canada","reference":"https://txt.cohere.com/introducing-embed-v3/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:dc6931eded6cfcd6e4ab","name":"DeepSA","organization":"ShanghaiTech University","date":"2023-11-02","domains":["Biology"],"country":"China","reference":"https://jcheminf.biomedcentral.com/articles/10.1186/s13321-023-00771-3","confidence":"Speculative","accessibility":"Hosted access (no API)"},{"id":"epoch:6e98ee6d9e96666867f0","name":"BlueLM 70B","organization":"vivo AI lab","date":"2023-11-02","domains":["Language"],"country":"China","reference":"https://baijiahao.baidu.com/s?id=1781445143383237948&wfr=spider&for=pc","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2bf7bb1d053bea8b4135","name":"BlueLM 130B","organization":"vivo AI lab","date":"2023-11-02","domains":["Language"],"country":"China","reference":"https://baijiahao.baidu.com/s?id=1781445143383237948&wfr=spider&for=pc","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e80c738411461f0477d1","name":"BlueLM 175B","organization":"vivo AI lab","date":"2023-11-02","domains":["Language"],"country":"China","reference":"https://baijiahao.baidu.com/s?id=1781445143383237948&wfr=spider&for=pc","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4ec6b48b16bd4ce43250","name":"Yi 6B","organization":"01.AI","date":"2023-11-02","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2403.04652","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:28a066afdf7234f4d3c9","name":"DeepSeek Coder 1.3B","organization":"DeepSeek,Peking University","date":"2023-11-02","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2401.14196\n\nhttps://huggingface.co/deepseek-ai/deepseek-coder-1.3b-base","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:601d0dc212208efcad03","name":"DeepSeek Coder 6.7B","organization":"DeepSeek,Peking University","date":"2023-11-02","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2401.14196\n\nhttps://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base ","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:723c4fa0f274d8dfbdde","name":"Nanbeige-16B","organization":"Nanbeige LLM Lab","date":"2023-11-01","domains":["Language"],"country":"China","reference":"https://github.com/Nanbeige/Nanbeige/blob/main/README_EN.md","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f00dde58c1b386a40dc3","name":"LingoWhale-8B","organization":"DeepLang AI","date":"2023-11-01","domains":["Language"],"country":"China","reference":"https://github.com/DeepLangAI/LingoWhale-8B/blob/main/README_EN.md","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:9fa545a8bf5a2e1fec41","name":"Calm2-7B","organization":"CyberAgent","date":"2023-11-01","domains":["Language"],"country":"Japan","reference":"https://huggingface.co/cyberagent/calm2-7b","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5ef27d8b707791f7dae7","name":"OpenChat 3.5-7B","organization":"Tsinghua University","date":"2023-11-01","domains":["Language"],"country":"China","reference":"https://huggingface.co/openchat/openchat_3.5","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4a3d432694fa66c7c5a5","name":"MuggleMath","organization":"University of Science and Technology of China (USTC),Alibaba","date":"2023-11-01","domains":["Language","Mathematics"],"country":"China,China","reference":"https://arxiv.org/abs/2310.05506","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e743dd6893c3b731bb0f","name":"YaRN (Llama 2 13B)","organization":"Nous Research,EleutherAI,University of Geneva","date":"2023-11-01","domains":["Language"],"country":"United States of America,United States of America,Switzerland","reference":"https://arxiv.org/abs/2309.00071","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:c9b49d00205e9e633b8c","name":"YaRN (Llama 2 70B)","organization":"Nous Research,EleutherAI,University of Geneva","date":"2023-11-01","domains":["Language"],"country":"United States of America,United States of America,Switzerland","reference":"https://arxiv.org/abs/2309.00071","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:72a3f5e53293aa4b13ce","name":"BlueLM 7B","organization":"vivo AI lab","date":"2023-10-31","domains":["Language"],"country":"China","reference":"https://github.com/vivo-ai-lab/BlueLM/blob/main/BlueLM_technical_report.pdf","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:73f0ab2f9d90c3b52c10","name":"Tongyi Qianwen 2.0","organization":"Alibaba","date":"2023-10-31","domains":["Language"],"country":"China","reference":"https://www.alibabacloud.com/blog/alibaba-cloud-launches-tongyi-qianwen-2-0-and-industry-specific-models-to-support-customers-reap-benefits-of-generative-ai_600526","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:84fb2f64eea42598ab1b","name":"Mi:dm 200B","organization":"KT","date":"2023-10-31","domains":["Language"],"country":"South Korea","reference":"https://genielabs.ai/midm/about","confidence":"Confident","accessibility":"API access"},{"id":"epoch:11f1bc8f6df4d7d55af8","name":"Skywork-13B","organization":"Kunlun Inc.","date":"2023-10-30","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2310.19341","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:e061fe1bc0aa2e58bb63","name":"Spec-Drafter","organization":"Peking University,Microsoft Research Asia","date":"2023-10-30","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2203.16487","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:a76c37eac9a67e7b3302","name":"ChatGLM3-6B","organization":"Z.ai (Zhipu AI)","date":"2023-10-27","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://www.zhipuai.cn/en/news/76\n\nhttps://huggingface.co/zai-org/chatglm3-6b","confidence":"Speculative","accessibility":"Open weights (restricted use)"},{"id":"epoch:4c403f687b44d5a26947","name":"CODEFUSION (Python)","organization":"Microsoft,Microsoft Research","date":"2023-10-26","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2310.17680 (was withdrawn)\n\nhttps://www.microsoft.com/en-us/research/wp-content/uploads/2023/11/CodeFusion-Revised-CameraReady.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0eb2f94f287a0712dcf1","name":"DiT-XL/2 + CADS","organization":"ETH Zurich,Disney Research","date":"2023-10-26","domains":["Image generation"],"country":"Switzerland,United States of America,Switzerland","reference":"https://arxiv.org/abs/2310.17347v2","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:dbd123860bd8b0aafe88","name":"QMoE: compressed SwitchTransformer","organization":"Institute of Science and Technology Austria (ISTA),Neural Magic","date":"2023-10-25","domains":["Language"],"country":"Austria,United States of America","reference":"https://arxiv.org/abs/2310.16795","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5a54d2778598bdb01f3d","name":"Xinghan Foundation Model","organization":"Dahua Technology","date":"2023-10-25","domains":["Multimodal","Video","Language","Vision"],"country":"China","reference":"https://www.prnewswire.com/news-releases/dahua-announces-think-2-0-strategy-to-accelerate-innovation-for-a-digital-intelligent-future-301967122.html","confidence":"Unknown","accessibility":""},{"id":"epoch:36887e0cf1311555c525","name":"Zephyr 7B","organization":"Hugging Face","date":"2023-10-25","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2310.16944","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f8baaccaa3cae1ce4a2e","name":"Spark 3.0","organization":"iFlytek","date":"2023-10-24","domains":["Language"],"country":"China","reference":"https://www.gizmochina.com/2023/10/26/iflytek-spark-3-0-vs-openai-gpt/","confidence":"Likely","accessibility":""},{"id":"epoch:012b757021524e936b7a","name":"Stockmark-13B","organization":"Stockmark","date":"2023-10-23","domains":["Language"],"country":"Japan","reference":"https://huggingface.co/stockmark/stockmark-13b","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a4f6dfef84b8cb079f25","name":"CausalLM 14B","organization":"CausalLM","date":"2023-10-22","domains":["Language"],"country":"China","reference":"https://huggingface.co/CausalLM/14B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:aaa0798b8da5b12d4b85","name":"SILC-S* (86M)","organization":"ETH Zurich,DeepMind,Google,Technical University of Munich","date":"2023-10-20","domains":["Vision"],"country":"Switzerland,United Kingdom,United States of America,Germany","reference":"https://arxiv.org/abs/2310.13355","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:685fb820b92c952f81ac","name":"SILC-S","organization":"ETH Zurich,DeepMind,Google,Technical University of Munich","date":"2023-10-20","domains":["Vision"],"country":"Switzerland,United Kingdom,United States of America,Germany","reference":"https://arxiv.org/abs/2310.13355","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:4cc2331ee89f96525c3b","name":"DALL\u00b7E 3","organization":"OpenAI","date":"2023-10-19","domains":["Image generation"],"country":"United States of America","reference":"https://cdn.openai.com/papers/dall-e-3.pdf","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:23788fae003e76304140","name":"KwaiYiiMath","organization":"Kuaishou Technology","date":"2023-10-19","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2310.07488","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:12a3fd797493c537bd3e","name":"Voicebox / VB-En","organization":"Facebook AI Research","date":"2023-10-19","domains":["Speech"],"country":"United States of America,France","reference":"https://arxiv.org/abs/2306.15687","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c5c5a2a1c316fda10988","name":"ERNIE 4.0","organization":"Baidu","date":"2023-10-17","domains":["Multimodal","Language","Video","Image generation"],"country":"China","reference":"https://www.prnewswire.com/news-releases/baidu-launches-ernie-4-0-foundation-model-leading-a-new-wave-of-ai-native-applications-301958681.html","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:1b20d0cb44bb69652a08","name":"Fuyu-8B","organization":"Adept","date":"2023-10-17","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://www.adept.ai/blog/fuyu-8b, www.huggingface.co/adept/fuyu-8b","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:46bbcb7773a1ff7bfba4","name":"PaLI-3","organization":"Google DeepMind,Google Research,Google Cloud","date":"2023-10-17","domains":["Multimodal","Language","Vision"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2310.09199","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f6cd47c1bd8a0371357c","name":"Llemma 34B","organization":"Princeton University,University of Toronto,Vector Institute,University of Cambridge,Carnegie Mellon University (CMU),University of Washington,EleutherAI","date":"2023-10-16","domains":["Mathematics","Language"],"country":"United States of America,Canada,Canada,United Kingdom,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2310.10631","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:04e6c15df845073ee565","name":"Llemma 7B","organization":"Princeton University,EleutherAI,University of Toronto,Vector Institute,University of Cambridge,Carnegie Mellon University (CMU),University of Washington","date":"2023-10-16","domains":["Mathematics","Language"],"country":"United States of America,United States of America,Canada,Canada,United Kingdom,United States of America,United States of America","reference":"https://arxiv.org/abs/2310.10631","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:61be6c5fcc8baa5c0b64","name":"Aquila2 34B","organization":"Beijing Academy of Artificial Intelligence / BAAI","date":"2023-10-13","domains":["Language"],"country":"China","reference":"https://github.com/FlagAI-Open/Aquila2\nhttps://huggingface.co/BAAI/Aquila2-34B","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ab543b4adaed4584d174","name":"Table-GPT","organization":"Microsoft Research","date":"2023-10-13","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2310.09263","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:ed38601ae8a7ecd39fee","name":"RT-2-X","organization":"Google DeepMind","date":"2023-10-13","domains":["Robotics"],"country":"United States of America","reference":"https://arxiv.org/abs/2310.08864","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:9ba93249af34cbfbd654","name":"Jiutian","organization":"China Mobile","date":"2023-10-12","domains":["Language"],"country":"China","reference":"https://www.globaltimes.cn/page/202310/1299716.shtml","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2ab1fa12dd9310aa60bf","name":"Ferret (13B)","organization":"Columbia University,Apple","date":"2023-10-11","domains":["Multimodal","Language","Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2310.07704","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:9967be9cdff3d866d4ae","name":"Mistral 7B","organization":"Mistral AI","date":"2023-10-10","domains":["Language"],"country":"France","reference":"https://arxiv.org/abs/2310.06825","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:073606db6a077f56510f","name":"CodeFuse-13B","organization":"Ant Group","date":"2023-10-10","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2310.06266","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c745ecfeefab95715863","name":"CELLE-2","organization":"Chan Zuckerberg Initiative,University of California San Francisco,University of California (UC) Berkeley","date":"2023-10-10","domains":["Biology"],"country":"United States of America,United States of America,United States of America","reference":"https://proceedings.neurips.cc/paper_files/paper/2023/hash/0fb7c02d420c993385c7de44c2b5bf01-Abstract-Conference.html","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b4691b32fe11ac4c0a72","name":"RoseTTAFold All-Atom (RFAA)","organization":"University of Washington,Seoul National University,University of Sheffield","date":"2023-10-09","domains":["Biology"],"country":"United States of America,South Korea,United Kingdom","reference":"https://www.biorxiv.org/content/10.1101/2023.10.09.561603v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2c8943a6654e621d47db","name":"FinGPT-13B","organization":"University of California Los Angeles (UCLA),Columbia University,New York University (NYU)","date":"2023-10-07","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2310.04793; https://github.com/AI4Finance-Foundation/FinGPT","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f75767ad03bc634a1131","name":"NAEPro","organization":"University of California Santa Barbara (UCSB),Massachusetts Institute of Technology (MIT),Carnegie Mellon University (CMU)","date":"2023-10-06","domains":["Biology"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2310.04343","confidence":"Confident","accessibility":""},{"id":"epoch:56e26914c771822b2dd3","name":"FoldFlow","organization":"McGill University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),Dreamfold,University of Montreal / Universit\u00e9 de Montr\u00e9al,University of Oxford","date":"2023-10-03","domains":["Biology"],"country":"Canada,Canada,Canada,Canada,United Kingdom","reference":"https://arxiv.org/abs/2310.02391","confidence":"Confident","accessibility":""},{"id":"epoch:0bfa3b250545c5b90ee7","name":"LLaMA-7B (protein-oriented instruction-tuned)","organization":"Zhejiang University (ZJU)","date":"2023-10-02","domains":["Language","Biology"],"country":"China","reference":"https://arxiv.org/abs/2306.08018","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:29d068ded4dca43621fb","name":"Phi-1","organization":"Microsoft Research","date":"2023-10-02","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2306.11644","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:954e28cf2a48a234b944","name":"New Intelligent Q&A","organization":"","date":"2023-10-02","domains":["Language"],"country":"","reference":"https://maimai.cn/article/detail?fid=1772490696&efid=877mstTqdcjFmkFxmte9Kg","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:b3e639dfd51543393b3d","name":"CTM (CIFAR-10)","organization":"Stanford University,Sony","date":"2023-10-01","domains":["Image generation"],"country":"United States of America,Japan","reference":"https://arxiv.org/abs/2310.02279v1","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d5f665b70ea8ac601690","name":"TinyLlama-1.1B (1T token checkpoint)","organization":"Singapore University of Technology & Design","date":"2023-10-01","domains":["Language"],"country":"Singapore","reference":"https://arxiv.org/abs/2401.02385","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5c1ab0ef9517d8577544","name":"TinyLlama-1.1B (3T token checkpoint)","organization":"Singapore University of Technology & Design","date":"2023-10-01","domains":["Language"],"country":"Singapore","reference":"https://arxiv.org/abs/2401.02385","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2cac335c8795c8590665","name":"BITTERS","organization":"LG,Shutterstock","date":"2023-10-01","domains":["Vision"],"country":"South Korea,United States of America","reference":"https://arxiv.org/abs/2211.06774","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:b7c9d8da010380ea28a2","name":"PIXART-\u03b1","organization":"Huawei Noah's Ark Lab,The University of Hong Kong,Hong Kong University of Science and Technology (HKUST)","date":"2023-09-30","domains":["Image generation"],"country":"China,Hong Kong,Hong Kong","reference":"https://arxiv.org/abs/2310.00426\nhttps://openreview.net/pdf?id=eAKmQPe3m1\nhttps://github.com/PixArt-alpha/PixArt-alpha","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:349a3a5c4e1f9950c1b1","name":"HiDream Large Model 1.0","organization":"HiDream","date":"2023-09-30","domains":["Video","Image generation","Vision","3D modeling"],"country":"China","reference":"https://blog.csdn.net/Datawhale/article/details/133366053\n\nhttps://www.sohu.com/a/722858298_490443","confidence":"Likely","accessibility":"API access"},{"id":"epoch:55c5b1d4f24afdb6efc2","name":"StableLM-3B-4E1T","organization":"Stability AI","date":"2023-09-29","domains":["Language"],"country":"United Kingdom","reference":"https://stability.wandb.io/stability-llm/stable-lm/reports/StableLM-3B-4E1T--VmlldzoyMjU4?accessToken=u3zujipenkx5g7rtcj9qojjgxpconyjktjkli2po09nffrffdhhchq045vp0wyfo","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:91191d61aea09cd4cfda","name":"Wuerstchen","organization":"Technische Hochschule Ingolstadt,University of Montreal / Universit\u00e9 de Montr\u00e9al,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),Polytechnique Montreal,Wand Technologies","date":"2023-09-29","domains":["Image generation"],"country":"Germany,Canada,Canada,Canada,United States of America","reference":"https://arxiv.org/abs/2306.00637","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:50c2a05f91e45eb95aa6","name":"GAIA-1","organization":"Wayve","date":"2023-09-29","domains":["Video","Vision","Multimodal","Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2309.17080","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:7c9c2da8f4fb08c7192b","name":"Qwen-14B","organization":"Alibaba","date":"2023-09-28","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2309.16609","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:d8acc13e15272b47ac63","name":"Amazon Titan","organization":"Amazon","date":"2023-09-28","domains":["Language","Image generation"],"country":"United States of America","reference":"https://aws.amazon.com/bedrock/titan/","confidence":"Likely","accessibility":"API access"},{"id":"epoch:694c8f1c794cba2b2688","name":"Qwen-7B","organization":"Alibaba","date":"2023-09-28","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2309.16609, https://huggingface.co/Qwen/Qwen-7B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:d08da0efdb277a8f454b","name":"PLaMo-13B","organization":"Preferred Networks Inc","date":"2023-09-28","domains":["Language"],"country":"Japan","reference":"https://huggingface.co/pfnet/plamo-13b","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:53683b04e5ff5203cf41","name":"GPT-3.5 Turbo Instruct","organization":"OpenAI","date":"2023-09-28","domains":["Language"],"country":"United States of America","reference":"https://platform.openai.com/docs/models/gpt-3.5-turbo","confidence":"Likely","accessibility":"API access"},{"id":"epoch:bdd64a7e238c2d56259c","name":"Show-1","organization":"National University of Singapore","date":"2023-09-27","domains":["Video"],"country":"Singapore","reference":"https://arxiv.org/abs/2309.15818","confidence":"Unknown","accessibility":"Open weights (non-commercial)"},{"id":"epoch:13f21326e48c7600daec","name":"Emu (Meta)","organization":"Meta AI","date":"2023-09-27","domains":["Image generation"],"country":"United States of America","reference":"https://ai.meta.com/research/publications/emu-enhancing-image-generation-models-using-photogenic-needles-in-a-haystack/","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:79b27c386797be4e5366","name":"BigRNA","organization":"DeepGenomics","date":"2023-09-27","domains":["Biology"],"country":"Canada","reference":"https://www.biorxiv.org/content/10.1101/2023.09.20.558508v1","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:d913cbea189689472df8","name":"InternLM-XComposer","organization":"Shanghai AI Lab","date":"2023-09-26","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://arxiv.org/abs/2309.15112","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:2a513cce1b024d7b8fb5","name":"GPT-4V","organization":"OpenAI","date":"2023-09-25","domains":["Multimodal","Vision","Language"],"country":"United States of America","reference":"https://cdn.openai.com/papers/GPTV_System_Card.pdf","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:f791336f65463cab8666","name":"AlphaMissense","organization":"Google DeepMind","date":"2023-09-22","domains":["Biology"],"country":"United States of America","reference":"https://www.science.org/doi/10.1126/science.adg7492","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:7a69129853ff7902a04d","name":"PengCheng Mind (\u9e4f\u57ce\u8111\u6d77)","organization":"Peng Cheng Laboratory","date":"2023-09-21","domains":["Language"],"country":"China","reference":"https://www.gitlink.org.cn/OpenI2023/PengCheng\nhttps://www.pcl.ac.cn/html/943/2023-09-21/content-4299.html","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:ef2ea70ce13bb03047e4","name":"BTLM-3B","organization":"Cerebras Systems","date":"2023-09-20","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2309.11568","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:743ea2cd909e4de01ade","name":"DreamLLM","organization":"Xi\u2019an Jiaotong University,Megvii Inc,Tsinghua University,Huazhong University of Science and Technology","date":"2023-09-20","domains":["Multimodal","Language","Vision","Image generation"],"country":"China,China,China,China","reference":"https://arxiv.org/abs/2309.11499","confidence":"Confident","accessibility":""},{"id":"epoch:286b71e46b7ee73d180e","name":"GPT-MolBERTa","organization":"Carnegie Mellon University (CMU)","date":"2023-09-20","domains":["Biology"],"country":"United States of America","reference":"https://arxiv.org/abs/2310.03030","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c66ba29fd6a724b0ce1a","name":"Baichuan 2-7B","organization":"Baichuan","date":"2023-09-20","domains":["Language"],"country":"China","reference":"https://arxiv.org/pdf/2309.10305","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:d7d79861f37b7fef9fb4","name":"OpenChat-13b","organization":"Tsinghua University,Shanghai AI Lab,01.AI","date":"2023-09-20","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/pdf/2309.11235","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:28e858696a3fc6507862","name":"Deepgram Nova-2","organization":"Deepgram","date":"2023-09-19","domains":["Speech"],"country":"United States of America","reference":"https://deepgram.com/learn/nova-2-speech-to-text-api","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:03992adc9a80bff9e945","name":"bge-reranker-large","organization":"Beijing Academy of Artificial Intelligence / BAAI,Hugging Face","date":"2023-09-14","domains":["Language"],"country":"China,United States of America","reference":"https://arxiv.org/abs/2309.07597v2\nhttps://huggingface.co/BAAI/bge-reranker-large","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d863804f279e3b15f75b","name":"DeciLM 6B","organization":"Deci AI","date":"2023-09-13","domains":["Language"],"country":"Israel","reference":"https://huggingface.co/Deci/DeciLM-6b","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:435ccab7f220bf06ab02","name":"Robot Parkour","organization":"Shanghai Qi Zhi institute,Stanford University,Carnegie Mellon University (CMU),Tsinghua University","date":"2023-09-12","domains":["Robotics"],"country":"China,United States of America,United States of America,China","reference":"https://arxiv.org/abs/2309.05665","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:70fcb3d0e457ec07c895","name":"Phi-1.5","organization":"Microsoft","date":"2023-09-11","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2309.05463, https://huggingface.co/microsoft/phi-1_5","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:277838ba0ad07aa1e2d2","name":"MADLAD-400 10B","organization":"Google DeepMind,Google Research","date":"2023-09-09","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2309.04662","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b90eb5c30e5667d01624","name":"Mobile V-MoEs","organization":"Apple","date":"2023-09-08","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2309.04354","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:bd9e4274bcf1c7871112","name":"Persimmon-8B","organization":"Adept","date":"2023-09-07","domains":["Language"],"country":"United States of America","reference":"https://www.adept.ai/blog/persimmon-8b","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:950b6110fbddb076550d","name":"FLM-101B","organization":"Chinese Academy of Sciences,Harbin Institute of Technology,Nanyang Technological University,Beijing Academy of Artificial Intelligence / BAAI","date":"2023-09-07","domains":["Language"],"country":"China,China,Singapore,China","reference":"https://arxiv.org/abs/2309.03852","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:02ae8023f63cd2d9eaef","name":"XGen-7B","organization":"Salesforce","date":"2023-09-07","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2309.03450","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a34ae52f194578cc25c4","name":"Hunyuan","organization":"Tencent","date":"2023-09-07","domains":["Language","Image generation","Multimodal"],"country":"China","reference":"https://www.tencent.com/en-us/articles/2201685.html","confidence":"Confident","accessibility":"API access"},{"id":"epoch:3caf695d611077689e58","name":"ELIXR-C","organization":"Google,Northwestern Medicine,Apollo Radiology International","date":"2023-09-07","domains":["Vision","Medicine"],"country":"United States of America,United States of America,India","reference":"https://arxiv.org/abs/2308.01317","confidence":"Unknown","accessibility":""},{"id":"epoch:3d8154c21e32afdf5e93","name":"ELIXR-B","organization":"Google,Northwestern Medicine,Apollo Radiology International","date":"2023-09-07","domains":["Vision","Medicine"],"country":"United States of America,United States of America,India","reference":"https://arxiv.org/abs/2308.01317","confidence":"Unknown","accessibility":""},{"id":"epoch:96b18990adf626435abc","name":"Falcon-180B","organization":"Technology Innovation Institute","date":"2023-09-06","domains":["Language"],"country":"United Arab Emirates","reference":"https://falconllm.tii.ae/falcon-180b.html; https://arxiv.org/abs/2311.16867","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:86c1645c25dc27d2cae0","name":"Baichuan2-13B","organization":"Baichuan","date":"2023-09-06","domains":["Language"],"country":"China","reference":"https://huggingface.co/baichuan-inc/Baichuan2-13B-Base, https://arxiv.org/abs/2309.10305","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:c0634123475f64cdb7d3","name":"TigerBot-70B","organization":"Tigerobo","date":"2023-09-06","domains":["Language"],"country":"China","reference":"https://github.com/TigerResearch/TigerBot/blob/main/README_en.md, https://arxiv.org/abs/2312.08688","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:c5793e6b9d1d0b78d457","name":"360 Smart Brain","organization":"360 Security Technology","date":"2023-09-04","domains":["Multimodal","Language","Image generation"],"country":"China","reference":"https://www.hayo.com/article/64f68f1d8578eea6c7ec663f","confidence":"Likely","accessibility":""},{"id":"epoch:4a646953d0570a4e9ae5","name":"Swift","organization":"Intel Labs","date":"2023-08-30","domains":["Robotics"],"country":"United States of America","reference":"https://www.nature.com/articles/s41586-023-06419-4","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:e488a74ebcda03fad29b","name":"ABAB","organization":"MiniMax","date":"2023-08-30","domains":["Language"],"country":"China","reference":"https://kr-asia.com/baidus-ernie-bot-among-eight-chinese-llm-products-approved-for-public-launch, https://api.minimax.chat/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:258d780bac1b4bb7d32d","name":"Jais","organization":"Cerebras Systems,Mohamed bin Zayed University of Artificial Intelligence (MBZUAI),Inception G42","date":"2023-08-29","domains":["Language"],"country":"United States of America,United Arab Emirates,United Arab Emirates","reference":"https://arxiv.org/abs/2308.16149","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:31a90462839c1744ce4f","name":"Refact-1.6B","organization":"Refact AI","date":"2023-08-29","domains":["Language"],"country":"United Kingdom","reference":"https://huggingface.co/smallcloudai/Refact-1_6B-fim","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:d0abd1a677773c8bb5e2","name":"Luca 2.0","organization":"Mianbi Intelligence","date":"2023-08-29","domains":[],"country":"China","reference":"https://www.163.com/dy/article/IDBGA8840511FQO9.html","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:047e2b73eb86e57ccd15","name":"PeptideBERT","organization":"Carnegie Mellon University (CMU)","date":"2023-08-28","domains":["Biology"],"country":"United States of America","reference":"https://arxiv.org/abs/2309.03099","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d1b125f20d10ee184192","name":"MengziGPT-General-40B","organization":"Langboat","date":"2023-08-27","domains":["Language"],"country":"China","reference":"https://hub.baai.ac.cn/view/33895","confidence":"Confident","accessibility":""},{"id":"epoch:f650b455ba1ebb1b4198","name":"Mengzi-Fin-7B","organization":"Langboat","date":"2023-08-27","domains":["Language"],"country":"China","reference":"https://www.langboat.com/en/portal/mengzi-gpt","confidence":"Confident","accessibility":""},{"id":"epoch:5c1c0b0e1b193678dfd9","name":"Mengzi-Fin-13B","organization":"Langboat","date":"2023-08-27","domains":["Language"],"country":"China","reference":"https://hub.baai.ac.cn/view/33895","confidence":"Confident","accessibility":""},{"id":"epoch:8c0d66ea3f267b022310","name":"Mengzi-Lite","organization":"Langboat","date":"2023-08-25","domains":["Language"],"country":"China","reference":"https://www.langboat.com/en/portal/mengzi-gpt","confidence":"Unknown","accessibility":""},{"id":"epoch:466557b55a209c4c6850","name":"Open-Assistant Llama2 70B SFT v10","organization":"","date":"2023-08-25","domains":["Language"],"country":"","reference":"https://huggingface.co/OpenAssistant/llama2-70b-oasst-sft-v10","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:217f00dd9617fbcbd8b7","name":"MathGPT","organization":"TAL Education Group (Xueersi)","date":"2023-08-24","domains":["Language"],"country":"China","reference":"https://www.gizmochina.com/2023/08/24/mathgpt-launch-public-beta-next-gen-math-assistant/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:cf7e993bd16067c785e9","name":"HyperCLOVA X","organization":"NAVER","date":"2023-08-24","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/abs/2404.01954\nhttps://techcrunch.com/2023/08/24/koreas-internet-giant-naver-unveils-generative-ai-services/, https://www.ncloud.com/solution/featured/hyperclovax","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:100431ba1bfe82570826","name":"Qwen-VL","organization":"Alibaba","date":"2023-08-24","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://arxiv.org/abs/2308.12966","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:913133193de50b34af2d","name":"airoboros-l2-70b-2.1","organization":"","date":"2023-08-24","domains":["Language"],"country":"","reference":"https://huggingface.co/jondurbin/airoboros-l2-70b-2.1","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:1e9e87500dfa03b54ad4","name":"PULI GPTrio","organization":"Hungarian Research Centre for Linguistics","date":"2023-08-23","domains":["Language"],"country":"Hungary","reference":"https://link.springer.com/chapter/10.1007/978-3-031-40498-6_9; https://huggingface.co/NYTK/PULI-GPTrio","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:a6951021675ad5ba6920","name":"ShapeMol","organization":"Ohio State University","date":"2023-08-23","domains":["Biology"],"country":"United States of America","reference":"https://arxiv.org/abs/2308.11890","confidence":"Confident","accessibility":""},{"id":"epoch:65665a329a370a122215","name":"IDEFICS-80B","organization":"Hugging Face","date":"2023-08-22","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://huggingface.co/blog/idefics","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:e2b42742e3bb6c0715c6","name":"IDEFICS-9B","organization":"Hugging Face","date":"2023-08-22","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://huggingface.co/blog/idefics","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:c7e21ec017f0011eedbb","name":"Eleven Multilingual v2","organization":"ElevenLabs","date":"2023-08-22","domains":["Speech"],"country":"United States of America","reference":"https://elevenlabs.io/blog/eleven-multilingual-v2","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:ec1961b4393381317d34","name":"Dou Bao","organization":"ByteDance","date":"2023-08-18","domains":["Language"],"country":"China","reference":"https://pandaily.com/bytedance-launches-its-first-large-scale-ai-conversation-product-dou-bao/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:853b710a254ab3752450","name":"KwaiYii 13B","organization":"Kuaishou Technology","date":"2023-08-16","domains":["Language"],"country":"China","reference":"https://github.com/kwai/KwaiYii","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:346fc074cd657ba310ee","name":"VARCO LLM 2.0 base","organization":"NCSOFT","date":"2023-08-16","domains":["Language"],"country":"South Korea","reference":"https://ncsoft.github.io/ncresearch/varco-llm-details/\nhttps://aws.amazon.com/marketplace/pp/prodview-d7amr4yxpibew?sr=0-3&ref_=beagle&applicationId=AWSMPContessa","confidence":"Likely","accessibility":"API access"},{"id":"epoch:93bb0b4d926aa77b9927","name":"VARCO LLM 2.0 small Finetuning","organization":"NCSOFT","date":"2023-08-16","domains":["Language"],"country":"South Korea","reference":"https://ncsoft.github.io/ncresearch/varco-llm-details/\nhttps://aws.amazon.com/marketplace/pp/prodview-or3w6j66on53s?sr=0-2&ref_=beagle&applicationId=AWSMPContessa","confidence":"Likely","accessibility":"API access"},{"id":"epoch:8db7bd47a6201f216164","name":"VARCO LLM KO/EN-13B-IST ver.1","organization":"NCSOFT","date":"2023-08-16","domains":["Language"],"country":"South Korea","reference":"https://ncsoft.github.io/ncresearch/varco-llm-details/\nhttps://aws.amazon.com/marketplace/pp/prodview-usyosolf3an3u?sr=0-1&ref_=beagle&applicationId=AWSMPContessa","confidence":"Likely","accessibility":"API access"},{"id":"epoch:29d9555c5ea3cefab6d6","name":"DeciCoder-1B","organization":"Deci AI","date":"2023-08-15","domains":["Language"],"country":"Israel","reference":"https://huggingface.co/Deci/DeciCoder-1b","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ec6cf8c451d36585a02b","name":"Konan LLM 13B","organization":"Konan Technology","date":"2023-08-15","domains":["Language","Vision"],"country":"South Korea","reference":"https://en.konantech.com/en/llm/konanllm\nhttps://techfinch.kr/ai/konan-technology-unveils-konan-llm--its-own-ai-language-model\nhttps://www.dbpia.co.kr/journal/articleDetail?nodeId=NODE11610127 ","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:60bab75fcff5f1611794","name":"A.X (Adot) 7B","organization":"SK Telecom","date":"2023-08-15","domains":["Language","Speech"],"country":"South Korea","reference":"https://a.sktelecom.com/\nhttps://m.ddaily.co.kr/page/view/2023092614585090151","confidence":"Speculative","accessibility":""},{"id":"epoch:0bea5ea8210309eacc1f","name":"Spark 2.0","organization":"iFlytek","date":"2023-08-15","domains":[],"country":"China","reference":"https://technode.com/2023/08/16/iflytek-unveils-updated-llm-sparkdesk-v2-0-and-new-product-iflycode-1-0/","confidence":"Unknown","accessibility":""},{"id":"epoch:cc5dd7ed4e59a611728c","name":"Platypus-70B","organization":"Boston University","date":"2023-08-14","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2308.07317, https://platypus-llm.github.io/","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:3ab2fe721c7e0e53e5a0","name":"Code Llama-34B","organization":"Meta AI","date":"2023-08-14","domains":["Language"],"country":"United States of America","reference":"https://ai.meta.com/research/publications/code-llama-open-foundation-models-for-code/\nhttps://arxiv.org/abs/2308.12950","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:9b6143d7f0b92bc912a6","name":"Japanese-LM-3.6B","organization":"LINE Corporation","date":"2023-08-14","domains":["Language"],"country":"Japan","reference":"https://engineering.linecorp.com/ja/blog/3.6-billion-parameter-japanese-language-model, https://huggingface.co/line-corporation/japanese-large-lm-3.6b","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1fc9c4401e4c35b0decc","name":"Code Llama-7B","organization":"Meta AI","date":"2023-08-14","domains":["Language"],"country":"United States of America","reference":"https://ai.meta.com/research/publications/code-llama-open-foundation-models-for-code/\nhttps://arxiv.org/abs/2308.12950","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:1c04442d32e715c1e093","name":"Code Llama-13B","organization":"Meta AI","date":"2023-08-14","domains":["Language"],"country":"United States of America","reference":"https://ai.meta.com/research/publications/code-llama-open-foundation-models-for-code/\nhttps://arxiv.org/abs/2308.12950","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:4a627b01f80852b17612","name":"Japanese StableLM Base Alpha 7B","organization":"Stability AI","date":"2023-08-10","domains":["Language"],"country":"United Kingdom","reference":"https://stability.ai/news/stability-ai-new-jplm-japanese-language-model-stablelm","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7fc4f002242d8da1c3a5","name":"Baichuan2-53B","organization":"Baichuan","date":"2023-08-09","domains":["Language"],"country":"China","reference":"https://technode.com/2023/08/09/chinese-ai-startup-baichuan-rolls-out-third-llm-in-four-months/","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:fced5b76dc42d5ac3175","name":"Claude Instant","organization":"Anthropic","date":"2023-08-09","domains":["Language"],"country":"United States of America","reference":"https://www.anthropic.com/news/releasing-claude-instant-1-2","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:7e365630badacb0f4ebf","name":"CALM (NVIDIA)","organization":"Nvidia,Technion - Israel Institute of Technology","date":"2023-08-06","domains":["Robotics"],"country":"United States of America,Israel","reference":"https://research.nvidia.com/labs/par/calm/\nhttps://arxiv.org/abs/2305.02195","confidence":"Unknown","accessibility":""},{"id":"epoch:c3e595fe744b1ea0ce1d","name":"SS-pLM","organization":"Nostrum Biodiscovery,Barcelona Supercomputing Center,Institucio Catalana de Recerca i Estudis Avanc\u00e7ats","date":"2023-08-06","domains":["Biology"],"country":"Spain,Spain,Spain","reference":"https://www.biorxiv.org/content/10.1101/2023.08.04.551626.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:68aeb5179a1d43d1a287","name":"StableLM-Base-Alpha-7B","organization":"Stability AI","date":"2023-08-05","domains":["Language"],"country":"United Kingdom","reference":"https://huggingface.co/stabilityai/stablelm-base-alpha-7b-v2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a70e820ec0c6bade0769","name":"GGNN","organization":"Westlake University,Tsinghua University,Toyota Technological Institute at Chicago","date":"2023-08-05","domains":["Biology"],"country":"China,China,United States of America","reference":"https://www.nature.com/articles/s42003-023-05133-1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:552f6a3409fef61fa668","name":"Weblab-10B","organization":"Matsuo Lab","date":"2023-08-04","domains":["Language"],"country":"Japan","reference":"https://huggingface.co/matsuo-lab/weblab-10b","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:553530f209d09ecbbb38","name":"YuLan-Chat-2 (13B)","organization":"Renmin University of China","date":"2023-08-02","domains":["Language"],"country":"China","reference":"https://github.com/RUC-GSAI/YuLan-Chat","confidence":"Confident","accessibility":""},{"id":"epoch:be372a234a5d6370ee95","name":"OpenFlamingo","organization":"University of Washington,Stanford University,Allen Institute for AI,Hebrew University of Jerusalem,Columbia University,Google DeepMind,University of California Santa Barbara (UCSB),Research Center Juelich","date":"2023-08-02","domains":["Image generation"],"country":"United States of America,United States of America,United States of America,Israel,United States of America,United States of America,United States of America,Germany","reference":"https://arxiv.org/abs/2308.01390","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0bad6361f04ce4b12ac1","name":"JIANG","organization":"K.D. Feddersen (KDF)","date":"2023-08-01","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2308.00624","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:eaedc6fad40c691ee51d","name":"Vicuna-7B-v1.5","organization":"Large Model Systems Organization,University of California (UC) Berkeley","date":"2023-08-01","domains":["Language"],"country":"United States of America,United States of America","reference":"https://huggingface.co/lmsys/vicuna-7b-v1.5","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:949d40bb57c906a2ac88","name":"Vicuna-13B-v1.5","organization":"Large Model Systems Organization,University of California (UC) Berkeley","date":"2023-08-01","domains":["Language"],"country":"United States of America,United States of America","reference":"https://huggingface.co/lmsys/vicuna-13b-v1.5","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:a7786c347ad75ec55bce","name":"bilingual-gpt-neox-4b","organization":"rinna","date":"2023-07-31","domains":["Language"],"country":"Japan","reference":"https://huggingface.co/rinna/bilingual-gpt-neox-4b\n\nhttps://arxiv.org/abs/2404.01657","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:28e5493ff916858af6d3","name":"RT-2","organization":"Google DeepMind","date":"2023-07-28","domains":["Robotics"],"country":"United States of America","reference":"https://arxiv.org/abs/2307.15818","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c5a657aec927ed4ec309","name":"Zi Yue","organization":"NetEase","date":"2023-07-28","domains":["Language"],"country":"China","reference":"https://www.chinadaily.com.cn/a/202307/28/WS64c3226ea31035260b8190a4.html","confidence":"Unknown","accessibility":""},{"id":"epoch:fceb8b6fdade298992da","name":"Zi Yue 2.0","organization":"NetEase","date":"2023-07-28","domains":["Language"],"country":"China","reference":"https://m.aastocks.com/en/usq/news/comment.aspx?source=AAFN&id=NOW.1317044&catg=1","confidence":"Unknown","accessibility":""},{"id":"epoch:ca07e4f45837479d7eb9","name":"BELLE-Llama2-13B-chat-0.4M","organization":"KE Holdings Inc. (\u201cBeike\u201d)","date":"2023-07-27","domains":["Language"],"country":"China","reference":"https://huggingface.co/BELLE-2/BELLE-Llama2-13B-chat-0.4M","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:f6d5ae2defd435a1d7b5","name":"Xinyi Video 1.0","organization":"","date":"2023-07-27","domains":[],"country":"","reference":"https://www.sohu.com/a/706739123_115978","confidence":"Unknown","accessibility":""},{"id":"epoch:d7eb15d3ae147aaa1f75","name":"AudioLM","organization":"Google Research","date":"2023-07-26","domains":["Audio"],"country":"United States of America","reference":"https://arxiv.org/abs/2209.03143","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:aa29017104e4331d780d","name":"CharacterGLM","organization":"Beijing Lingxin Intelligent Technology Co., Ltd.","date":"2023-07-25","domains":[],"country":"China","reference":"https://zhidx.com/p/388605.html","confidence":"Unknown","accessibility":""},{"id":"epoch:e96e61c645040e5febcf","name":"WizardLM 13B v1.2","organization":"Microsoft,Peking University","date":"2023-07-25","domains":["Language"],"country":"United States of America,China","reference":"https://huggingface.co/WizardLMTeam/WizardLM-13B-V1.2","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:ad62cf935c2fe6f941c4","name":"RFdiffusion","organization":"University of Washington,Columbia University,Ecole Normale Sup\u00e8rieure,University of Cambridge,Massachusetts Institute of Technology (MIT),Seoul National University","date":"2023-07-23","domains":["Biology"],"country":"United States of America,United States of America,France,United Kingdom,United States of America,South Korea","reference":"https://www.nature.com/articles/s41586-023-06415-8","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1284e6c2093ac971d0c8","name":"LM-Design","organization":"ByteDance,University of Wisconsin Madison","date":"2023-07-23","domains":["Biology"],"country":"China,United States of America","reference":"https://proceedings.mlr.press/v202/zheng23a.html","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:73e7ded0832e5ab225ad","name":"YAYI-13B-Llama2","organization":"Yayi (Wenge)","date":"2023-07-22","domains":["Language"],"country":"China","reference":"https://github.com/wenge-research/YAYI","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:bada4d040bc225f318a7","name":"YAYI-7B-Llama2","organization":"Yayi (Wenge)","date":"2023-07-22","domains":["Language"],"country":"China","reference":"https://github.com/wenge-research/YAYI","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:193ebaf53ee82aa3c81c","name":"Stable Beluga 1","organization":"Stability AI","date":"2023-07-21","domains":["Language"],"country":"United Kingdom","reference":"https://huggingface.co/stabilityai/StableBeluga1-Delta","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:6d39d314e168f7167129","name":"Stable Beluga 2","organization":"Stability AI","date":"2023-07-20","domains":["Language"],"country":"United Kingdom","reference":"https://stability.ai/news/stable-beluga-large-instruction-fine-tuned-models, https://huggingface.co/stabilityai/StableBeluga2","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:3d7d9fd0451e3ec24eea","name":"EXAONE 2.0","organization":"LG AI Research","date":"2023-07-19","domains":["Multimodal","Image generation","Language","Biology","Vision"],"country":"South Korea","reference":"https://aws.amazon.com/solutions/case-studies/lg-ai-research-case-study/","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:b8a094554c7fc3dc265c","name":"Llama 2-70B","organization":"Meta AI","date":"2023-07-18","domains":["Language"],"country":"United States of America","reference":"https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/\nhttps://arxiv.org/abs/2307.09288","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:cda7c117f2bf6d383e9f","name":"Llama 2-34B","organization":"Meta AI","date":"2023-07-18","domains":["Language"],"country":"United States of America","reference":"https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/\nhttps://arxiv.org/abs/2307.09288","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c12c3628066cb3ae4c37","name":"Llama 2-7B","organization":"Meta AI","date":"2023-07-18","domains":["Language"],"country":"United States of America","reference":"https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:1eb59cec44d5d07eeb79","name":"Llama 2-13B","organization":"Meta AI","date":"2023-07-18","domains":["Language"],"country":"United States of America","reference":"https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/\nhttps://arxiv.org/abs/2307.09288","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:e54c99ab9808a4739d3b","name":"GPT3-2.7B (FlashAttention-2)","organization":"Stanford University,Princeton University","date":"2023-07-18","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/pdf/2307.08691","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c3e605adf00c3f56d2c0","name":"RetNet","organization":"Microsoft Research,Tsinghua University","date":"2023-07-17","domains":["Language"],"country":"United States of America,China","reference":"https://arxiv.org/abs/2307.08621","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0c015ebf45aa8a02a4b7","name":"BaiLing","organization":"Ant Group","date":"2023-07-15","domains":["Language"],"country":"China","reference":"https://www.antgroup.com/en/technology/new-tech-Technology-Antfocuses-Tabcomnent-detail/20241028001\n\nhttps://finance.yahoo.com/news/ant-group-gets-green-light-093000023.html","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:c2dc95fd82ce10fd8f84","name":"CryoChains","organization":"University of California Santa Barbara (UCSB),Stanford University","date":"2023-07-15","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2306.07274","confidence":"Confident","accessibility":""},{"id":"epoch:363e5e9b252aa3ceec13","name":"ChatRhino","organization":"JD.com","date":"2023-07-13","domains":["Language"],"country":"China","reference":"https://jdcorporateblog.com/jd-com-introduces-chatrhino-empowering-industry-innovations-with-an-advanced-large-language-model/","confidence":"Confident","accessibility":""},{"id":"epoch:0ce782e18ab7018dcf66","name":"Uni-RNA-L8","organization":"DP Technology","date":"2023-07-12","domains":["Biology"],"country":"China","reference":"https://www.biorxiv.org/content/10.1101/2023.07.11.548588v1.abstract","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:3ca472f6e7cdda3d4918","name":"Med-PaLM","organization":"Google Research,National Library of Medicine,DeepMind","date":"2023-07-12","domains":["Medicine","Language"],"country":"United States of America,United States of America,United Kingdom","reference":"https://www.nature.com/articles/s41586-023-06291-2","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:19492f14ba2176865c63","name":"Uni-RNA-L-24","organization":"DP Technology","date":"2023-07-12","domains":["Biology"],"country":"China","reference":"https://www.biorxiv.org/content/10.1101/2023.07.11.548588v1.abstract","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:2164948363852c0083cd","name":"Uni-RNA-L12","organization":"DP Technology","date":"2023-07-12","domains":["Biology"],"country":"China","reference":"https://www.biorxiv.org/content/10.1101/2023.07.11.548588v1.abstract","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:54f3267f0a6182123c57","name":"Uni-RNA-L16","organization":"DP Technology","date":"2023-07-12","domains":["Biology"],"country":"China","reference":"https://www.biorxiv.org/content/10.1101/2023.07.11.548588v1.abstract","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:45d2e51b513a645a94b8","name":"Claude 2","organization":"Anthropic","date":"2023-07-11","domains":["Language"],"country":"United States of America","reference":"https://www.anthropic.com/index/claude-2, https://www-files.anthropic.com/production/images/Model-Card-Claude-2.pdf","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:d064f985df845d615ae8","name":"Emu1 (BAAI)","organization":"Beijing Academy of Artificial Intelligence / BAAI,Tsinghua University,Peking University","date":"2023-07-11","domains":["Vision","Multimodal","Language"],"country":"China,China,China","reference":"https://arxiv.org/abs/2307.05222","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:4cc903dc1abdcf54ec4b","name":"Baichuan 1-13B","organization":"Baichuan","date":"2023-07-11","domains":["Language"],"country":"China","reference":"https://github.com/baichuan-inc/Baichuan-13B/blob/main/README_EN.md","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:c1efb3e8d100292eafe3","name":"West Lake (\u201cX\u012bh\u00fa / \u897f\u6e56\u5927\u6a21\u578b\u201d)","organization":"West Lake Xinchen / Xinchen AI / \u897f\u6e56\u5fc3\u8fb0\uff08\u676d\u5dde\uff09\u79d1\u6280\u6709\u9650\u516c\u53f8","date":"2023-07-09","domains":["Multimodal","Language","Vision","Speech"],"country":"China","reference":"https://zhidx.com/news/38369.html\nhttps://www.sohu.com/a/660043004_112831\nhttps://mp.weixin.qq.com/s/3nYVs_jnpkhNbsA4HHTNAw\nhttps://www.xinchenai.com/product/LLM","confidence":"Unknown","accessibility":"Open weights (non-commercial)"},{"id":"epoch:1ab7631c57ad3a572d61","name":"TeleChat","organization":"China Telecom","date":"2023-07-07","domains":["Language"],"country":"China","reference":"https://m.thepaper.cn/baijiahao_23766944","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:00a0f94bdb1462c586cc","name":"Pangu 3.0","organization":"Huawei","date":"2023-07-07","domains":["Multimodal","Language","Image generation","Vision"],"country":"China","reference":"https://www.huaweicloud.com/intl/en-us/news/20230707180809498.html","confidence":"Likely","accessibility":"API access"},{"id":"epoch:46855fb2d44d42558d17","name":"InternLM","organization":"Shanghai AI Lab,SenseTime","date":"2023-07-06","domains":["Language"],"country":"China,Hong Kong","reference":"https://github.com/InternLM/InternLM-techreport/blob/main/InternLM.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:65e20aef10567b73c88c","name":"CodeGen2.5","organization":"Salesforce","date":"2023-07-06","domains":["Language"],"country":"United States of America","reference":"https://blog.salesforceairesearch.com/codegen25/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:457e7cd3c78ae743ae37","name":"xTrimoPGLM -100B","organization":"Tsinghua University,BioMap Research","date":"2023-07-06","domains":["Biology"],"country":"China,China","reference":"https://www.biorxiv.org/content/10.1101/2023.07.05.547496v4","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:31b0926f7f16aa633e2f","name":"NEC LLM 13B","organization":"NEC Laboratories","date":"2023-07-06","domains":["Language"],"country":"United States of America","reference":"https://jpn.nec.com/press/202307/20230706_02.html","confidence":"Confident","accessibility":""},{"id":"epoch:9d17c89a0906407b7bcd","name":"Pangu-Weather","organization":"Huawei","date":"2023-07-05","domains":["Earth science"],"country":"China","reference":"https://www.nature.com/articles/s41586-023-06185-3, https://www.huaweicloud.com/intl/en-us/news/20230707180809498.html,\nhttps://www.huawei.com/en/news/2023/7/pangu-ai-model-nature-publish","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:2613c3a24212862bee9d","name":"LongNet","organization":"Microsoft,Xi\u2019an Jiaotong University","date":"2023-07-05","domains":["Language"],"country":"United States of America,China","reference":"https://arxiv.org/abs/2307.02486","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:9af3bdb731ac02dbabcf","name":"Stable Diffusion XL (SDXL)","organization":"Stability AI","date":"2023-07-04","domains":["Image generation"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2307.01952","confidence":"Speculative","accessibility":"Open weights (non-commercial)"},{"id":"epoch:7044d0c6e2612b7b8e88","name":"Multilingual-E5-large","organization":"Microsoft","date":"2023-06-30","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2402.05672\nhttps://huggingface.co/intfloat/multilingual-e5-large","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:11902e71d684ff55236c","name":"Honghu Graphic","organization":"China Unicom","date":"2023-06-28","domains":["Vision","Image generation"],"country":"China","reference":"https://medium.com/@sdatokens/china-unicom-released-a-graphic-model-that-can-realize-with-text-pictures-video-clips-llm-ai-d50afd965a5b","confidence":"Likely","accessibility":""},{"id":"epoch:85854e3f2f088f93c4bd","name":"ERNIE 3.5","organization":"Baidu","date":"2023-06-27","domains":["Language"],"country":"China","reference":"http://research.baidu.com/Blog/index-view?id=185","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:7cda389f135fd31ecff1","name":"HyenaDNA","organization":"Stanford University,Harvard University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2023-06-27","domains":["Biology"],"country":"United States of America,United States of America,Canada,Canada","reference":"https://arxiv.org/abs/2306.15794","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:cde599c6d7b1fede0516","name":"Chirp","organization":"Google","date":"2023-06-27","domains":["Speech"],"country":"United States of America","reference":"https://medium.com/google-cloud/getting-started-with-chirp-the-googles-universal-speech-model-usm-on-vertex-ai-f54edaf4da93","confidence":"Confident","accessibility":"API access"},{"id":"epoch:c36e316c60361a75aa22","name":"RWKV-4 World (7B)","organization":"RWKV Foundation","date":"2023-06-26","domains":["Language"],"country":"Multinational","reference":"https://huggingface.co/BlinkDL/rwkv-4-world","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:309a766222d14a949770","name":"Kosmos-2","organization":"Microsoft","date":"2023-06-26","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://arxiv.org/abs/2306.14824","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b1dc46f914520275792b","name":"Llama-2-Chinese 13B","organization":"FlagAlpha","date":"2023-06-25","domains":["Language"],"country":"China","reference":"https://github.com/FlagAlpha/Llama2-Chinese","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:eb2f99da05ea4965be20","name":"Inflection-1","organization":"Inflection AI","date":"2023-06-23","domains":["Language"],"country":"United States of America","reference":"https://inflection.ai/assets/Inflection-1.pdf","confidence":"Speculative","accessibility":"Hosted access (no API)"},{"id":"epoch:080d7f2a20668a7bfd01","name":"MPT-30B","organization":"MosaicML","date":"2023-06-22","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/mosaicml/mpt-30b","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:845d312088523b191566","name":"Vicuna-33B-v1.3","organization":"Large Model Systems Organization,University of California (UC) Berkeley","date":"2023-06-22","domains":["Language"],"country":"United States of America,United States of America","reference":"https://huggingface.co/lmsys/vicuna-33b-v1.3","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:5a035ef8bfd2aa6e106b","name":"Vicuna-13B-v1.3","organization":"Large Model Systems Organization,University of California (UC) Berkeley","date":"2023-06-22","domains":["Language"],"country":"United States of America,United States of America","reference":"https://huggingface.co/lmsys/vicuna-13b-v1.3","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:2b425f82835908d545ab","name":"Vicuna-7B-v1.3","organization":"Large Model Systems Organization,University of California (UC) Berkeley","date":"2023-06-22","domains":["Language"],"country":"United States of America,United States of America","reference":"https://huggingface.co/lmsys/vicuna-7b-v1.3","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:280a4df48f53601c93d8","name":"RoboCat","organization":"Google DeepMind,Google","date":"2023-06-20","domains":["Robotics"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2306.11706","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:34694a28bff27e845601","name":"GigaGAN","organization":"POSTECH,Carnegie Mellon University (CMU),Adobe","date":"2023-06-19","domains":["Image generation"],"country":"South Korea,United States of America,United States of America","reference":"https://arxiv.org/abs/2303.05511","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:6ec4e0e6abdf968bed39","name":"Pix2Struct-Large","organization":"Google Research,University of Cambridge","date":"2023-06-15","domains":["Vision"],"country":"United States of America,United Kingdom","reference":"https://arxiv.org/abs/2210.03347","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:795da8095d041a593a32","name":"WizardCoder-15.5B","organization":"Microsoft","date":"2023-06-14","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2306.08568","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:ac802aeb5ad7a4796023","name":"GPT-4 (Jun 2023)","organization":"OpenAI","date":"2023-06-13","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2303.08774","confidence":"Likely","accessibility":"API access"},{"id":"epoch:6405c236a41ff0e40197","name":"GPT-3.5 Turbo","organization":"OpenAI","date":"2023-06-13","domains":["Language"],"country":"United States of America","reference":"https://platform.openai.com/docs/models/gpt-3.5-turbo","confidence":"Likely","accessibility":"API access"},{"id":"epoch:993059ba7e26ab6bbaed","name":"BELLE-LLaMA-7B-0.6M-enc","organization":"KE Holdings Inc. (\u201cBeike\u201d)","date":"2023-06-12","domains":["Language"],"country":"China","reference":"https://huggingface.co/BelleGroup/BELLE-LLaMA-7B-0.6M-enc","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:18c9419f1fa6286b599c","name":"BELLE-LLaMA-7B-2M-enc ","organization":"KE Holdings Inc. (\u201cBeike\u201d)","date":"2023-06-12","domains":["Language"],"country":"China","reference":"https://huggingface.co/BelleGroup/BELLE-LLaMA-7B-2M-enc","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:514ce9fa401a26ef2f11","name":"BELLE-LLaMA-13B-2M-enc","organization":"KE Holdings Inc. (\u201cBeike\u201d)","date":"2023-06-12","domains":["Language"],"country":"China","reference":"https://huggingface.co/BelleGroup/BELLE-LLaMA-13B-2M-enc","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:33b0326c45822a251851","name":"Wu Dao Aquila-7B","organization":"Beijing Academy of Artificial Intelligence / BAAI","date":"2023-06-10","domains":["Language"],"country":"China","reference":"https://spectrum.ieee.org/china-chatgpt-wu-dao\n\nhttps://huggingface.co/BAAI/Aquila-7B\n\nhttps://github.com/FlagAI-Open/FlagAI/blob/master/examples/Aquila/README_en.md","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:4b1cb45bd170772eaea3","name":"Wu Dao Aquila-33B","organization":"Beijing Academy of Artificial Intelligence / BAAI","date":"2023-06-10","domains":["Language"],"country":"China","reference":"https://spectrum.ieee.org/china-chatgpt-wu-dao\n\nhttps://huggingface.co/BAAI/Aquila-7B","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:1f6319039fe4a6efa599","name":"PoET","organization":"OpenProtein.ai","date":"2023-06-09","domains":["Biology"],"country":"Singapore","reference":"https://arxiv.org/abs/2306.06156","confidence":"Confident","accessibility":""},{"id":"epoch:b002942b9cc355ac8e2e","name":"SYNTERACT","organization":"University of Delaware","date":"2023-06-09","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/10.1101/2023.06.07.544109v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:93befb226a72412c68ff","name":"MusicGen","organization":"Meta AI","date":"2023-06-08","domains":["Audio"],"country":"United States of America","reference":"https://arxiv.org/abs/2306.05284","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:0546fc6fcb57a143bf00","name":"PolySphere-1","organization":"AI inside","date":"2023-06-08","domains":["Language"],"country":"Japan","reference":"https://inside.ai/en/news/2023/06/08/aiinside-xresearch/\nhttps://inside.ai/en/news/2023/06/29/polysphere-token/","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:d172c52893dc37f3cc9b","name":"RedPajama-INCITE-7B-Base","organization":"Together","date":"2023-06-06","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/togethercomputer/RedPajama-INCITE-7B-Base","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:af74b41c990b854eb5d7","name":"LTM-1","organization":"Magic","date":"2023-06-06","domains":["Language"],"country":"United States of America","reference":"https://magic.dev/blog/ltm-1","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:39749a9459edc38a1cf2","name":"GELU for CIFAR-10","organization":"University of California (UC) Berkeley,Toyota Technological Institute at Chicago","date":"2023-06-06","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1606.08415","confidence":"Speculative","accessibility":""},{"id":"epoch:75aaea891fcbcce402bb","name":"life2vec","organization":"Technical University of Denmark,University of Copenhagen","date":"2023-06-05","domains":["Medicine"],"country":"Denmark,Denmark","reference":"https://arxiv.org/abs/2306.03009","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:7ea5090576290727df23","name":"Polyglot-Ko-12.8B","organization":"EleutherAI","date":"2023-06-04","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2306.02254; https://huggingface.co/EleutherAI/polyglot-ko-12.8b","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:533ec29c9705731f6d53","name":"YaYi-7B","organization":"Yayi (Wenge)","date":"2023-06-03","domains":[],"country":"China","reference":"https://github.com/wenge-research/YAYI","confidence":"Confident","accessibility":""},{"id":"epoch:4e5d614e9491db423c59","name":"Baichuan1-7B","organization":"Baichuan","date":"2023-06-01","domains":["Language"],"country":"China","reference":"https://huggingface.co/baichuan-inc/Baichuan-7B","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:79ce2106371288d21cbc","name":"TransAct","organization":"Pinterest","date":"2023-05-31","domains":["Recommendation"],"country":"United States of America","reference":"https://arxiv.org/abs/2306.00248","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:81845d4a92216bc26d23","name":"MASSA","organization":"Guangdong-Hong Kong-Macao Joint Laboratory of Human-Machine Intelligence-Synergy Systems,Shenzhen Institute of Advanced Technology,Chinese Academy of Sciences","date":"2023-05-30","domains":["Biology"],"country":"China,China,China","reference":"https://onlinelibrary.wiley.com/doi/full/10.1002/advs.202301223","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:e4da0cc93d9b59f25edb","name":"EGNN","organization":"InstaDeep","date":"2023-05-30","domains":["Biology"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2305.19801","confidence":"Unknown","accessibility":""},{"id":"epoch:73fc904d18fa8911d25e","name":"UniDiffuser (\u591a\u6a21\u6001\u5927\u6a21\u578b)","organization":"ShengShu,Tsinghua University,Beijing Academy of Artificial Intelligence / BAAI","date":"2023-05-30","domains":["Image generation"],"country":"China,China,China","reference":"https://arxiv.org/abs/2303.06555\n\n","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5d55270f2d6dd8727827","name":"PaLI-X","organization":"Google Research","date":"2023-05-29","domains":["Multimodal","Language","Vision","Video"],"country":"United States of America","reference":"https://arxiv.org/abs/2305.18565","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:6918cc0fc23811ff31da","name":"bpRNA-align","organization":"Oregon State University","date":"2023-05-29","domains":["Biology"],"country":"United States of America","reference":"https://pmc.ncbi.nlm.nih.gov/articles/PMC10159002/","confidence":"Confident","accessibility":""},{"id":"epoch:08fa7a89259db8be0059","name":"DPO on Pythia-2.8B","organization":"Stanford University,CZ Biohub Network","date":"2023-05-29","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2305.18290","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:fab36804822bbdd56158","name":"30B-Lazarus","organization":"Caldera AI","date":"2023-05-27","domains":["Language"],"country":"","reference":"https://huggingface.co/CalderaAI/30B-Lazarus","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:d0786cf4a0cf521a3842","name":"HuatuoGPT","organization":"Shenzhen Research Institue of Big Data,Chinese University of Hong Kong (CUHK)","date":"2023-05-24","domains":["Language"],"country":"China,Hong Kong","reference":"https://arxiv.org/abs/2305.15075","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:cecd1bae001892930a2d","name":"Shanhai","organization":"Unisound","date":"2023-05-24","domains":["Language"],"country":"China","reference":"https://ai.unisound.com/doc/unigpt/Algorithm.html","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:c8eed4e4308e29e0f56d","name":"LLaMA-7B (LoRA finetuned)","organization":"NAVER","date":"2023-05-23","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/pdf/2305.14152.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:39003c512f069bb10383","name":"LLaMA-65B (LoRA finetuned)","organization":"NAVER","date":"2023-05-23","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/pdf/2305.14152.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c841d61653c80ebeb477","name":"LLaMA-13B (LoRA finetuned)","organization":"NAVER","date":"2023-05-23","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/pdf/2305.14152.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b0a6537047dbf7667e25","name":"LLaMA-33B (LoRA finetuned)","organization":"NAVER","date":"2023-05-23","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/pdf/2305.14152.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:557797e0d5571f8aba2f","name":"Goat-7B","organization":"National University of Singapore","date":"2023-05-23","domains":["Language"],"country":"Singapore","reference":"https://arxiv.org/abs/2305.14201","confidence":"Speculative","accessibility":"Open weights (non-commercial)"},{"id":"epoch:94d39a20fd3bf2169be2","name":"Guanaco-65B","organization":"University of Washington","date":"2023-05-23","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2305.14314; https://github.com/artidoro/qlora","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:aebf1dcb69e1b69764d7","name":"BiomedGPT (182M)","organization":"Lehigh University,University of Georgia,Samsung Research America,Harvard Medical School,University of Pennsylvania","date":"2023-05-23","domains":["Language","Vision","Medicine"],"country":"United States of America,United States of America,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2305.17100","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:57d57921a2486e8844da","name":"ProlificDreamer","organization":"Tsinghua University,ShengShu","date":"2023-05-23","domains":["3D modeling"],"country":"China,China","reference":"http://arxiv.org/abs/2305.16213","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:4844ba259a8307691c56","name":"RWKV-4 14B","organization":"RWKV Foundation","date":"2023-05-22","domains":["Language"],"country":"Multinational","reference":"https://arxiv.org/abs/2305.13048","confidence":"Confident","accessibility":""},{"id":"epoch:8c2cdf2bef6cf6215907","name":"MMS-1B","organization":"Meta AI","date":"2023-05-22","domains":["Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2305.13516\nhttps://huggingface.co/facebook/mms-1b-all","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:eb0f0f0e1aa56a87cbd6","name":"CodeT5+","organization":"Salesforce","date":"2023-05-20","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2305.07922","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b84f4eedf512ef0fd3fe","name":"XuanYuan 2.0","organization":"Du Xiaoman","date":"2023-05-19","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2305.12002, https://huggingface.co/Duxiaoman-DI/XuanYuan-176B","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:ae247fffb59dd13b769d","name":"ONE-PEACE","organization":"Alibaba,Huazhong University of Science and Technology","date":"2023-05-18","domains":["Multimodal","Vision","Speech","Language"],"country":"China,China","reference":"https://arxiv.org/abs/2305.11172v1","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:471ecebbfed4bedf6d43","name":"LIMA","organization":"Meta AI,Carnegie Mellon University (CMU),University of Southern California,Tel Aviv University","date":"2023-05-18","domains":["Language"],"country":"United States of America,United States of America,United States of America,Israel","reference":"https://arxiv.org/abs/2305.11206","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f3b519be376c10a3f817","name":"Tianhe Tianyuan","organization":"National Supercomputer Center in Tianjin","date":"2023-05-18","domains":[],"country":"China","reference":"https://global.chinadaily.com.cn/a/202305/19/WS646713c3a310b6054fad3f8e.html","confidence":"Confident","accessibility":""},{"id":"epoch:36be3255fab222b0ae5c","name":"CongRong (\u4ece\u5bb9\u5927\u6a21\u578b)","organization":"CloudWalk Technology","date":"2023-05-18","domains":["Language"],"country":"China","reference":"https://www.cloudwalk.com/news/show/id/182\nhttps://www.jiqizhixin.com/articles/2024-06-28-16","confidence":"Unknown","accessibility":""},{"id":"epoch:7055fca8962cd95cb995","name":"CoEdiT-xxl","organization":"University of Minnesota,Grammarly","date":"2023-05-17","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2305.09857, https://huggingface.co/grammarly/coedit-large","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:daf72083bf1eb465e31a","name":"WeLM","organization":"WeChat AI","date":"2023-05-16","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2209.10372 ","confidence":"Confident","accessibility":""},{"id":"epoch:12c974ccc3386361a9af","name":"Med-PaLM 2","organization":"Google Research,DeepMind","date":"2023-05-16","domains":["Medicine","Language"],"country":"United States of America,United Kingdom","reference":"https://arxiv.org/abs/2305.09617","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:4363583f8c6556b094ff","name":"OpenCALM","organization":"CyberAgent","date":"2023-05-15","domains":["Language"],"country":"Japan","reference":"https://huggingface.co/cyberagent/open-calm-7b","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4b9c9638495196a979ba","name":"A.X (Adot) 39B","organization":"SK Telecom","date":"2023-05-15","domains":["Language","Speech"],"country":"South Korea","reference":"https://a.sktelecom.com/\nhttps://m.ddaily.co.kr/page/view/2023092614585090151","confidence":"Speculative","accessibility":"Hosted access (no API)"},{"id":"epoch:181e5dcb10386e6586a6","name":"LMRec","organization":"NAVER,Naver AI Lab","date":"2023-05-13","domains":["Language","Recommendation"],"country":"South Korea,South Korea","reference":"https://arxiv.org/abs/2212.03760","confidence":"Likely","accessibility":""},{"id":"epoch:297046b9de3649effcde","name":"InstructBLIP","organization":"Salesforce Research,Hong Kong University of Science and Technology (HKUST),Nanyang Technological University","date":"2023-05-11","domains":["Multimodal","Language","Vision"],"country":"United States of America,Hong Kong,Singapore","reference":"https://arxiv.org/abs/2305.06500","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:c6f07dd3590cff713608","name":"ESM-GearNet","organization":"Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Universit\u00e9 de Montr\u00e9al,IBM Research,HEC Montreal,CIFAR AI Research","date":"2023-05-11","domains":["Biology"],"country":"Canada,Canada,United States of America,Canada,Canada","reference":"https://arxiv.org/abs/2303.06275","confidence":"Likely","accessibility":""},{"id":"epoch:251c8eff4b8d66ce958d","name":"PaLM 2","organization":"Google","date":"2023-05-10","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2305.10403","confidence":"Likely","accessibility":"API access"},{"id":"epoch:6e4e2087b5b2febe854b","name":"PaLM-2 Bison","organization":"Google","date":"2023-05-10","domains":["Language"],"country":"United States of America","reference":"https://ai.google.dev/palm_docs/palm#model-sizes","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:0517c97eda4e15e06d26","name":"PaLM-2 Unicorn","organization":"Google","date":"2023-05-10","domains":["Language"],"country":"United States of America","reference":"https://ai.google.dev/palm_docs/palm#model-sizes","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:ccaf93be1839431e583b","name":"PaLM-2 Gecko","organization":"Google","date":"2023-05-10","domains":["Language"],"country":"United States of America","reference":"https://ai.google.dev/palm_docs/palm#model-sizes","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:6036b91e0f5ffb7cab57","name":"PaLM-2 Otter","organization":"Google","date":"2023-05-10","domains":["Language"],"country":"United States of America","reference":"https://ai.google.dev/palm_docs/palm#model-sizes","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:6e9ecebd99573903eda9","name":"StarCoder","organization":"Hugging Face,ServiceNow,Northeastern University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),Carnegie Mellon University (CMU),Johns Hopkins University,Leipzig University,ScaDS.AI,Queen Mary University of London,Roblox,Sea AI Lab,Technion - Israel Institute of Technology,Monash University,CSIRO,Data61,McGill University,Saama,University of British Columbia (UBC),Massachusetts Institute of Technology (MIT),Technical University of Munich,IBM,University of Vermont,UnfoldML,SAP,University of Notre Dame,Columbia University,New York University (NYU),University of Allahabad,Discover Dollar,Toloka,Telefonica,Stanford University,Weizmann Institute of Science,Alan Turing Institute,Wellesley College,EleutherAI,Forschungszentrum Julich","date":"2023-05-09","domains":["Language"],"country":"United States of America,United States of America,United States of America,Canada,United States of America,United States of America,Germany,Germany,United Kingdom,United States of America,Singapore,Israel,Australia,Australia,Australia,Canada,United States of America,Canada,United States of America,Germany,United States of America,United States of America,Sweden,Germany,United States of America,United States of America,United States of America,India,India,Netherlands,Spain,United States of America,Israel,United Kingdom,United States of America,United States of America,Germany","reference":"https://arxiv.org/abs/2305.06161","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:eb0f036fa5eb3eb3dba7","name":"ImageBind","organization":"Meta AI","date":"2023-05-09","domains":["Multimodal","Vision","Audio","Language","Image generation","Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2305.05665, https://github.com/facebookresearch/ImageBind","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:99a501244a54a367912f","name":"Spark","organization":"iFlytek","date":"2023-05-06","domains":["Language"],"country":"China","reference":"https://technode.com/2023/05/08/iflytek-unveils-large-language-model-claims-it-outperforms-chatgpt/","confidence":"Unknown","accessibility":""},{"id":"epoch:6f233752fb056315e941","name":"MPT-7B","organization":"MosaicML","date":"2023-05-05","domains":["Language"],"country":"United States of America","reference":"https://www.mosaicml.com/blog/mpt-7b","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a46c8b6b9b76cff4f403","name":"Otter","organization":"Nanyang Technological University","date":"2023-05-05","domains":["Multimodal","Language","Vision"],"country":"Singapore","reference":"https://arxiv.org/abs/2305.03726","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3e63d1bc0d239c646c10","name":"CodeGen2","organization":"Salesforce","date":"2023-05-03","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2305.02309","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ea605e36e83aedab9524","name":"Perfusion","organization":"Nvidia,Tel Aviv University,Bar-Ilan University","date":"2023-05-02","domains":["Image generation"],"country":"United States of America,Israel,Israel","reference":"https://arxiv.org/abs/2305.01644","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:159380d0a4b4c19ff51c","name":"OpenLLaMA-13B","organization":"OpenLM Research","date":"2023-05-01","domains":["Language"],"country":"United States of America","reference":"https://github.com/openlm-research/open_llama","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:df28a4d95c6f7db016e8","name":"Wutong","organization":"","date":"2023-05-01","domains":[],"country":"","reference":"https://news.qq.com/rain/a/20241022A077IC00","confidence":"Unknown","accessibility":""},{"id":"epoch:8bd63a00dfab3c2fc30f","name":"MosaicML Diffusion","organization":"Databricks","date":"2023-04-28","domains":["Image generation"],"country":"United States of America","reference":"https://www.databricks.com/blog/stable-diffusion-2\nhttps://www.databricks.com/blog/diffusion\nhttps://github.com/mosaicml/diffusion","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e64dd04c81f9cc513cf6","name":"Shishuo","organization":"4Paradigm","date":"2023-04-27","domains":["Language","Multimodal","Speech"],"country":"China","reference":"https://global.chinadaily.com.cn/a/202304/27/WS644a601ea310b6054fad0390.html","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:ed82f08a529b177d2ac8","name":"Eleven Multilingual v1","organization":"ElevenLabs","date":"2023-04-27","domains":["Speech"],"country":"United States of America","reference":"https://elevenlabs.io/blog/eleven-multilingual-v1","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:f062641cbf7a300f7d53","name":"Agile Soccer Robot","organization":"Google DeepMind","date":"2023-04-26","domains":["Robotics"],"country":"United States of America","reference":"https://arxiv.org/abs/2304.13653","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:909594c5a89e6c35b0b9","name":"WizardLM-7B","organization":"Microsoft,Peking University","date":"2023-04-24","domains":["Language"],"country":"United States of America,China","reference":"https://arxiv.org/abs/2304.12244","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:f032db05c85e9754ed64","name":"Falcon-7B","organization":"Technology Innovation Institute","date":"2023-04-24","domains":["Language"],"country":"United Arab Emirates","reference":"https://huggingface.co/tiiuae/falcon-7b","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fc386fa0b1080eb8d9b8","name":"WizardLM 70B","organization":"Microsoft,Peking University","date":"2023-04-24","domains":["Language"],"country":"United States of America,China","reference":"https://huggingface.co/WizardLM/WizardLM-70B-V1.0\nhttps://arxiv.org/abs/2304.12244","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:dc7fe349a757649ad00c","name":"ruGPT-3.5 13B","organization":"Sber","date":"2023-04-24","domains":["Language"],"country":"Russia","reference":"https://huggingface.co/ai-forever/ruGPT-3.5-13B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:03c452b04842681f96d0","name":"GigaChat","organization":"Sber","date":"2023-04-24","domains":["Language"],"country":"Russia","reference":"https://www.sberbank.ru/ru/sberpress/vazhnoe/article?newsID=8d5ab003-3d9d-4342-abdb-693a6d99a47d&blockID=8a5ea25e-318c-4d17-a60d-e806c4b0bc07&regionID=77&lang=ru&type=NEWS","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:54e8bf4eab44dc56bb6d","name":"MiniGPT4 (Vicuna finetune)","organization":"King Abdullah University of Science and Technology (KAUST)","date":"2023-04-20","domains":["Language","Vision","Multimodal"],"country":"Saudi Arabia","reference":"https://arxiv.org/abs/2304.10592","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ba2de316913873bb4aca","name":"AlphaLink","organization":"Technische Universitat Berlin,Research Cluster of Excellence,University of Edinburgh","date":"2023-04-20","domains":["Biology"],"country":"Germany,Germany,United Kingdom","reference":"https://www.nature.com/articles/s41587-023-01704-z","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ef68eb44b14b4c53d2b1","name":"MOSS-Moon-003","organization":"Fudan University","date":"2023-04-19","domains":["Language"],"country":"China","reference":"https://huggingface.co/fnlp/moss-moon-003-base","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a12f93b73197873f28eb","name":"Claude 1.3","organization":"Anthropic","date":"2023-04-18","domains":["Language"],"country":"United States of America","reference":"https://twitter.com/AnthropicAI/status/1648353600350060545?lang=en","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:0d21f533ee9f25f00753","name":"AiLMe-100B v3","organization":"Qilin Hesheng Network Technology Co., Ltd. (APUS)","date":"2023-04-18","domains":["Multimodal","Language","Video","Image generation","Audio"],"country":"China","reference":"https://cevalbenchmark.com/static/model.html?method=AiLMe-100B%20v3\nhttps://tech.chinadaily.com.cn/a/202304/20/WS6440a111a310537989370a76.html","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f41df359fc5926e7d0d1","name":"LLaVA","organization":"University of Wisconsin Madison,Microsoft Research,Columbia University","date":"2023-04-17","domains":["Multimodal","Vision","Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2304.08485","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:76f58e84095f9f49df61","name":"BELLE-LLaMA-EXT-7B","organization":"KE Holdings Inc. (\u201cBeike\u201d)","date":"2023-04-16","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2304.07854","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:3b0f266295cfe5883146","name":"OpenCLIP ViT-H-14-378-quickgelu","organization":"Massachusetts Institute of Technology (MIT)","date":"2023-04-16","domains":["Vision"],"country":"United States of America","reference":"https://github.com/mlfoundations/open_clip","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:481763bcdc18fb17afbc","name":"Suno Bark Model","organization":"Suno","date":"2023-04-15","domains":["Audio"],"country":"United States of America","reference":"https://github.com/suno-ai/bark\nhttps://huggingface.co/suno/bark","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:b10d1d2a2f5c7e731ed1","name":"DINOv2","organization":"Facebook AI Research,INRIA","date":"2023-04-14","domains":["Vision"],"country":"United States of America,France,France","reference":"https://arxiv.org/abs/2304.07193","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9e227fb55b0800d654ba","name":"HuaTuo","organization":"Harbin Institute of Technology","date":"2023-04-14","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2304.06975","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:9ac3b61b50d48c4bf1ad","name":"Dolly 2.0-12b","organization":"Databricks","date":"2023-04-12","domains":["Language"],"country":"United States of America","reference":"https://www.databricks.com/blog/2023/04/12/dolly-first-open-commercially-viable-instruction-tuned-llm","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f885124a631412a7b3e7","name":"SenseChat","organization":"SenseTime","date":"2023-04-10","domains":["Language"],"country":"Hong Kong","reference":"https://www.sensetime.com/en/news-detail/51166397?categoryId=1072","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:b7a1195c0d0c53cc6faf","name":"Incoder-6.7B","organization":"Facebook AI Research,University of Washington,University of California (UC) Berkeley,Carnegie Mellon University (CMU),Toyota Technological Institute at Chicago","date":"2023-04-09","domains":["Language"],"country":"United States of America,France,United States of America,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2204.05999","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:965029a6b0095cc7591d","name":"gLM","organization":"Harvard University","date":"2023-04-08","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/10.1101/2023.04.07.536042v1.full\nhttps://github.com/y-hwang/gLM","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:108ba06a9ec8067baa6e","name":"DiffDock-PP","organization":"Technical University of Munich,Massachusetts Institute of Technology (MIT)","date":"2023-04-08","domains":["Biology"],"country":"Germany,United States of America","reference":"https://arxiv.org/abs/2304.03889","confidence":"Likely","accessibility":""},{"id":"epoch:98d2e3e3fb75e8ab92a1","name":"EvoMIL","organization":"University of Glasgow,Cancer Research UK Beatson Institute","date":"2023-04-08","domains":["Biology"],"country":"United Kingdom,United Kingdom","reference":"https://www.biorxiv.org/content/10.1101/2023.04.07.536023v1","confidence":"Unknown","accessibility":""},{"id":"epoch:6ef1bbeda267a3e4d4aa","name":"Cerebras-GPT-13B","organization":"Cerebras Systems","date":"2023-04-06","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2304.03208","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:64463ad2cf03fe3bf125","name":"Segment Anything Model","organization":"Meta AI","date":"2023-04-05","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2304.02643","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2970c8b576752dabbb71","name":"BELLE-7B-0.2M","organization":"KE Holdings Inc. (\u201cBeike\u201d)","date":"2023-04-05","domains":["Language"],"country":"China","reference":"https://huggingface.co/BelleGroup/BELLE-7B-0.2M","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:23966325d4d5906bc75d","name":"BELLE-7B-1M","organization":"KE Holdings Inc. (\u201cBeike\u201d)","date":"2023-04-05","domains":["Language"],"country":"China","reference":"https://huggingface.co/BelleGroup/BELLE-7B-1M","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:34f27e37b67ed3580b75","name":"BELLE-7B-2M","organization":"KE Holdings Inc. (\u201cBeike\u201d)","date":"2023-04-05","domains":["Language"],"country":"China","reference":"https://huggingface.co/BelleGroup/BELLE-7B-2M","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0d268cb69e8b64c08d28","name":"EigenFold","organization":"Massachusetts Institute of Technology (MIT)","date":"2023-04-05","domains":["Biology"],"country":"United States of America","reference":"https://arxiv.org/abs/2304.02198","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:dc92e28ea4f016b4563e","name":"Pythia-12b","organization":"EleutherAI,Booz Allen Hamilton, McLean,University of Cambridge,Indraprastha Institute of Information Technology\nDelhi,Stability AI,datasaur.ai,University of Amsterdam","date":"2023-04-03","domains":["Language"],"country":"United States of America,United States of America,United Kingdom,India,United Kingdom,United States of America,Netherlands","reference":"https://arxiv.org/abs/2304.01373","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e8a54c69920007826808","name":"Pythia-2.8b","organization":"EleutherAI,Booz Allen Hamilton, McLean,University of Cambridge,Indraprastha Institute of Information Technology\nDelhi,Stability AI,datasaur.ai,University of Amsterdam","date":"2023-04-03","domains":["Language"],"country":"United States of America,United States of America,United Kingdom,India,United Kingdom,United States of America,Netherlands","reference":"https://arxiv.org/abs/2304.01373","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e3da5b6c3151e4f865e9","name":"Pythia-6.9b","organization":"EleutherAI,Booz Allen Hamilton, McLean,University of Cambridge,Indraprastha Institute of Information Technology\nDelhi,Stability AI,datasaur.ai,University of Amsterdam","date":"2023-04-03","domains":["Language"],"country":"United States of America,United States of America,United Kingdom,India,United Kingdom,United States of America,Netherlands","reference":"https://arxiv.org/abs/2304.01373","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:98674d464b207339d7ad","name":"Pythia-160m","organization":"EleutherAI,Booz Allen Hamilton, McLean,University of Cambridge,Indraprastha Institute of Information Technology\nDelhi,Stability AI,datasaur.ai,University of Amsterdam","date":"2023-04-03","domains":["Language"],"country":"United States of America,United States of America,United Kingdom,India,United Kingdom,United States of America,Netherlands","reference":"https://arxiv.org/abs/2304.01373","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a697305a198bb67918e2","name":"Pythia-1b","organization":"EleutherAI,Booz Allen Hamilton, McLean,University of Cambridge,Indraprastha Institute of Information Technology\nDelhi,Stability AI,datasaur.ai,University of Amsterdam","date":"2023-04-03","domains":["Language"],"country":"United States of America,United States of America,United Kingdom,India,United Kingdom,United States of America,Netherlands","reference":"https://arxiv.org/abs/2304.01373","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e72d9a787831945df2aa","name":"Pythia-1.4b","organization":"EleutherAI,Booz Allen Hamilton, McLean,University of Cambridge,Indraprastha Institute of Information Technology\nDelhi,Stability AI,datasaur.ai,University of Amsterdam","date":"2023-04-03","domains":["Language"],"country":"United States of America,United States of America,United Kingdom,India,United Kingdom,United States of America,Netherlands","reference":"https://arxiv.org/abs/2304.01373","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:73b0c8819712f0a267cf","name":"Pythia-70m","organization":"EleutherAI,Booz Allen Hamilton, McLean,University of Cambridge,Indraprastha Institute of Information Technology\nDelhi,Stability AI,datasaur.ai,University of Amsterdam","date":"2023-04-03","domains":["Language"],"country":"United States of America,United States of America,United Kingdom,India,United Kingdom,United States of America,Netherlands","reference":"https://arxiv.org/abs/2304.01373","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:61395f76ad68777c5694","name":"Pythia-410m","organization":"EleutherAI,Booz Allen Hamilton, McLean,University of Cambridge,Indraprastha Institute of Information Technology\nDelhi,Stability AI,datasaur.ai,University of Amsterdam","date":"2023-04-03","domains":["Language"],"country":"United States of America,United States of America,United Kingdom,India,United Kingdom,United States of America,Netherlands","reference":"https://arxiv.org/abs/2304.01373","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ff1e30178df2ed7d6893","name":"MahLool","organization":"University of Rochester","date":"2023-04-03","domains":["Biology"],"country":"United States of America","reference":"https://pubs.acs.org/doi/full/10.1021/acs.jcim.2c01317","confidence":"Confident","accessibility":""},{"id":"epoch:952c6ce863b8efb2ab06","name":"STEPS","organization":"McGill University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),Baidu Research - Silicon Valley AI Lab,Baidu,Boston Consulting Group X","date":"2023-04-01","domains":["Biology"],"country":"Canada,Canada,United States of America,China,United States of America","reference":"https://academic.oup.com/bioinformatics/article/39/4/btad189/7117544","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:dd9a0cb6d57eb767710e","name":"Yiye Qingzhou","organization":"EFFYIC (\u8bc6\u56e0\u667a\u80fd)","date":"2023-04-01","domains":["Multimodal","Language","Vision","Speech","Video"],"country":"China","reference":"https://36kr.com/p/3062982772761984\nhttps://www.effyic.com/technical_core","confidence":"Confident","accessibility":""},{"id":"epoch:5b7f27f94b7d4f03bf07","name":"Vicuna-13B v0","organization":"Large Model Systems Organization,University of California (UC) Berkeley","date":"2023-03-30","domains":["Language"],"country":"United States of America,United States of America","reference":"https://lmsys.org/blog/2023-03-30-vicuna/\n\nhttps://github.com/lm-sys/FastChat/blob/main/docs/vicuna_weights_version.md","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:6f4bb0809b911b9aa2d9","name":"BloombergGPT","organization":"Bloomberg,Johns Hopkins University","date":"2023-03-30","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2303.17564","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b9f75c435165582123bd","name":"Vicuna-7B v0","organization":"Large Model Systems Organization,University of California (UC) Berkeley","date":"2023-03-30","domains":["Language"],"country":"United States of America,United States of America","reference":"https://lmsys.org/blog/2023-03-30-vicuna/\n\nhttps://github.com/lm-sys/FastChat/blob/main/docs/vicuna_weights_version.md","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:a66bf542a216ebfa807c","name":"VideoMAE V2","organization":"Nanjing University,Shenzhen Institute of Advanced Technology,Shanghai AI Lab","date":"2023-03-29","domains":["Video"],"country":"China,China,China","reference":"https://arxiv.org/abs/2303.16727v2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9630873618c12054abda","name":"ERNIE-ViLG 2.0","organization":"Baidu,Wuhan University of Science and Technology","date":"2023-03-28","domains":["Image generation"],"country":"China,China","reference":"https://arxiv.org/abs/2210.15257","confidence":"Confident","accessibility":"API access"},{"id":"epoch:8161ddbfe6e91e42d715","name":"SigLIP 400M","organization":"Google DeepMind","date":"2023-03-27","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2303.15343","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7d4d54aaf991ed3c474f","name":"SigLiT","organization":"Google DeepMind","date":"2023-03-27","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2303.15343","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:746231bb9faf79ba0dd9","name":"EVA-CLIP (EVA-02-CLIP-E/14+)","organization":"Beijing Academy of Artificial Intelligence / BAAI,Huazhong University of Science and Technology","date":"2023-03-27","domains":["Vision"],"country":"China,China","reference":"https://arxiv.org/abs/2303.15389\t","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:35383cdc9f8a9f547258","name":"BELLE-7B-0.6M","organization":"KE Holdings Inc. (\u201cBeike\u201d)","date":"2023-03-26","domains":["Language"],"country":"China","reference":"https://huggingface.co/BelleGroup/BELLE-7B-0.6M","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8f8c9e248fdbf4cc2b85","name":"CPM-Bee","organization":"Tsinghua University,Beijing Academy of Artificial Intelligence / BAAI,ModelBest,OpenBMB (Open Lab for Big Model Base)","date":"2023-03-24","domains":["Language"],"country":"China,China,China,China","reference":"https://github.com/OpenBMB/CPM-Bee","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:411b2232061f7d67abd1","name":"minChatGPT","organization":"Stanford University","date":"2023-03-24","domains":["Language"],"country":"United States of America","reference":"https://github.com/ethanyanjiali/minChatGPT/blob/main/report.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a560066e1f42f44d6e1d","name":"Lightweight Fine-tuning a Pretrained Protein Language Model for Protein Secondary","organization":"Henan University","date":"2023-03-23","domains":["Biology"],"country":"China","reference":"https://www.biorxiv.org/content/10.1101/2023.03.22.530066v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:5924b44d16e65e2f319c","name":"Sparse Wide GPT-3 Small","organization":"Cerebras Systems","date":"2023-03-21","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2303.11525","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:8d928e0d434a0341ad57","name":"LightOn Mini","organization":"LightOn","date":"2023-03-21","domains":["Language"],"country":"France","reference":"https://www.lighton.ai/blog/lighton-s-blog-4/lighton-s-large-language-model-of-40-billion-parameters-mini-19","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:aee6beb10e49501a0e45","name":"Firefly","organization":"Adobe","date":"2023-03-21","domains":["Image generation"],"country":"United States of America","reference":"https://news.adobe.com/news/news-details/2023/Adobe-Unveils-Firefly-a-Family-of-new-Creative-Generative-AI/default.aspx","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:687801b001d275773c15","name":"Gen-2","organization":"Runway","date":"2023-03-20","domains":["Video"],"country":"United States of America","reference":"https://research.runwayml.com/gen2","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:82cc3c6fb3336885302a","name":"PanGu-\u03a3","organization":"Huawei Noah's Ark Lab","date":"2023-03-20","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2303.10845","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2958d3f9a0a1ddf77b22","name":"GPT-4 (Mar 2023)","organization":"OpenAI","date":"2023-03-15","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2303.08774","confidence":"Likely","accessibility":"API access"},{"id":"epoch:3440ed3662690e00b7f9","name":"Falcon-40B","organization":"Technology Innovation Institute","date":"2023-03-15","domains":["Language"],"country":"United Arab Emirates","reference":"https://arxiv.org/abs/2311.16867; https://www.tii.ae/news/abu-dhabi-based-technology-innovation-institute-introduces-falcon-llm-foundational-large","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:89914dc9452e4d858d5c","name":"LEP-AD","organization":"King Abdullah University of Science and Technology (KAUST),Karolinska Institute","date":"2023-03-15","domains":["Biology"],"country":"Saudi Arabia,Sweden","reference":"https://www.biorxiv.org/content/10.1101/2023.03.14.532563v1.full.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b39e5c28f4d57584fbe1","name":"RIO","organization":"West Lake Xinchen / Xinchen AI / \u897f\u6e56\u5fc3\u8fb0\uff08\u676d\u5dde\uff09\u79d1\u6280\u6709\u9650\u516c\u53f8","date":"2023-03-15","domains":["Language"],"country":"China","reference":"https://www.xinchenai.com/product/LLM\n\nhttps://36kr.com/p/2179385107198212","confidence":"Likely","accessibility":""},{"id":"epoch:491239f2ec2285213a52","name":"Claude","organization":"Anthropic","date":"2023-03-14","domains":["Language"],"country":"United States of America","reference":"https://www.anthropic.com/index/introducing-claude","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:041b3cc333b13bed8982","name":"Alpaca","organization":"Stanford University","date":"2023-03-13","domains":["Language"],"country":"United States of America","reference":"https://crfm.stanford.edu/2023/03/13/alpaca.html","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:219a12f95aad057ce31a","name":"Pythia-Chat-Base-7B-v0.16","organization":"Together","date":"2023-03-08","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/togethercomputer/Pythia-Chat-Base-7B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1389bdfa981fca92afb6","name":"VALL-E X","organization":"Microsoft","date":"2023-03-07","domains":["Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2303.03926","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:f7dcf4a9e6c8bddd3ca7","name":"PaLM-E","organization":"Google,TU Berlin","date":"2023-03-06","domains":["Robotics","Vision","Language"],"country":"United States of America,Germany","reference":"https://arxiv.org/abs/2303.03378","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:48b1e7eed75731e1c140","name":"Uni-Mol Molecular Model","organization":"Renmin University of China,DP Technology,AI for Science Institute, Beijing (AISI)","date":"2023-03-06","domains":["Biology"],"country":"China,China,China","reference":"https://chemrxiv.org/engage/chemrxiv/article-details/6402990d37e01856dc1d1581","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e5bf9f623029c9078dd4","name":"AudioGen","organization":"Meta AI,Hebrew University of Jerusalem","date":"2023-03-05","domains":["Audio"],"country":"United States of America,Israel","reference":"https://arxiv.org/abs/2209.15352","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:bfb7978289133e6994d5","name":"Flan UL2","organization":"Google Brain","date":"2023-03-03","domains":["Language"],"country":"United States of America","reference":"https://www.yitay.net/blog/flan-ul2-20b, https://arxiv.org/abs/2205.05131","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e9599513dc001a6539f9","name":"DiT-XL/2","organization":"New York University (NYU),University of California (UC) Berkeley","date":"2023-03-02","domains":["Image generation"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2212.09748","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:cc16a36b20ea42a2a06f","name":"Consistency Model (CIFAR-10)","organization":"OpenAI","date":"2023-03-02","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2303.01469","confidence":"Unknown","accessibility":""},{"id":"epoch:13a39470ef3db60b2765","name":"Cohere Command","organization":"Cohere","date":"2023-03-01","domains":["Language"],"country":"Canada","reference":"https://cohere.com/models/command","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:03a19848367ea80b3f37","name":"Palmyra Large 20B","organization":"Writer","date":"2023-03-01","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/Writer/palmyra-large","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:12b3d2dd8afbd0c82967","name":"Jurassic-2 Jumbo","organization":"AI21 Labs","date":"2023-03-01","domains":["Language"],"country":"Israel","reference":"https://www.ai21.com/blog/introducing-j2","confidence":"","accessibility":"API access"},{"id":"epoch:2578ab7061a08c1dcb72","name":"gpt-sw3-40b","organization":"AI Sweden","date":"2023-03-01","domains":["Language"],"country":"Sweden","reference":"https://huggingface.co/AI-Sweden-Models/gpt-sw3-40b","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:feaab2441ebeef4aa5cd","name":"MengziGPT-General-7B","organization":"Langboat","date":"2023-03-01","domains":["Language"],"country":"China","reference":"https://hub.baai.ac.cn/view/33895","confidence":"Confident","accessibility":""},{"id":"epoch:ba48a2bbe0d23d6e6f32","name":"Kosmos-1","organization":"Microsoft","date":"2023-03-01","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2302.14045","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:60cf6ec2813fedc7c2ee","name":"CodeGen-Mono 16.1B","organization":"Salesforce","date":"2023-02-27","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2203.13474","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ebaf735a206d666db461","name":"LLaMA-13B","organization":"Meta AI","date":"2023-02-27","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2302.13971","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:d48b540948ec8568c817","name":"LLaMA-33B","organization":"Meta AI","date":"2023-02-27","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2302.13971","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:544a08f3fbd3f6167199","name":"MsPBRsP","organization":"Zhengzhou University","date":"2023-02-27","domains":["Biology"],"country":"China","reference":"https://www.biorxiv.org/content/10.1101/2023.02.26.528265v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:7e9f8b7b78339843ee41","name":"LLaMA-65B","organization":"Meta AI","date":"2023-02-24","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2302.13971","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:27caf645f7347afb9a6e","name":"LLaMA-7B","organization":"Meta AI","date":"2023-02-24","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2302.13971","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:7a3fb47a7713a652c324","name":"Hyena-3-slim","organization":"Stanford University,University of Montreal / Universit\u00e9 de Montr\u00e9al,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms)","date":"2023-02-21","domains":["Language"],"country":"United States of America,Canada,Canada","reference":"https://arxiv.org/abs/2302.10866","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:73fe53d51d569ea73f2d","name":"Hyena 1.3B","organization":"Stanford University,University of Montreal / Universit\u00e9 de Montr\u00e9al,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms)","date":"2023-02-21","domains":["Language"],"country":"United States of America,Canada,Canada","reference":"https://arxiv.org/abs/2302.10866","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:0a035418354462d8b26f","name":"Hyena-2 355M","organization":"Stanford University,University of Montreal / Universit\u00e9 de Montr\u00e9al,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms)","date":"2023-02-21","domains":["Language"],"country":"United States of America,Canada,Canada","reference":"https://arxiv.org/abs/2302.10866","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:7638b26b4298914eab67","name":"Hyena-2 153M","organization":"Stanford University,University of Montreal / Universit\u00e9 de Montr\u00e9al,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms)","date":"2023-02-21","domains":["Language"],"country":"United States of America,Canada,Canada","reference":"https://arxiv.org/abs/2302.10866","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:86c2da44dd3d6d056fc8","name":"E-SPA","organization":"University of Oxford,DeepMind","date":"2023-02-20","domains":["Language"],"country":"United Kingdom,United Kingdom","reference":"https://arxiv.org/abs/2302.10322","confidence":"","accessibility":"Unreleased"},{"id":"epoch:6a8e71de486a08dd83b7","name":"Anthropic LM 175B","organization":"Anthropic","date":"2023-02-15","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2302.07459","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:24ced54a8a474da336cf","name":"BASIC-L + Lion","organization":"Google,University of California Los Angeles (UCLA)","date":"2023-02-13","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2302.06675v4","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:bdc82c61fc3a79e259d7","name":"ViT-22B","organization":"Google","date":"2023-02-10","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2302.05442v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:3842161a2930ba676adc","name":"ControlNet","organization":"Stanford University","date":"2023-02-10","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2302.05543","confidence":"Unknown","accessibility":""},{"id":"epoch:85b4541d73b354263e9c","name":"ProteinDT","organization":"University of California (UC) Berkeley,California Institute of Technology,University of Toronto,University of Wisconsin Madison,Texas A&M,Nvidia,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms)","date":"2023-02-09","domains":["Biology","Language"],"country":"United States of America,United States of America,Canada,United States of America,United States of America,United States of America,Canada","reference":"https://arxiv.org/abs/2302.04611","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0c984351118aa63b9e5b","name":"ToolFormer","organization":"Meta AI","date":"2023-02-09","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2302.04761","confidence":"Confident","accessibility":""},{"id":"epoch:d8737192abdff1944cdd","name":"Gen-1","organization":"Runway","date":"2023-02-06","domains":["Video"],"country":"United States of America","reference":"https://arxiv.org/abs/2302.03011","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:c98910d886196ae6299b","name":"ProteinSGM","organization":"University of Toronto","date":"2023-02-04","domains":["Biology"],"country":"Canada","reference":"https://www.biorxiv.org/content/10.1101/2022.07.13.499967v2.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:d8ef69a1b4fdfdd51b97","name":"CD-GraB (WT103)","organization":"Cornell University","date":"2023-02-02","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2302.00845","confidence":"","accessibility":"Unreleased"},{"id":"epoch:3d6bbeb11702296b8504","name":"CD-GraB (WT2)","organization":"Cornell University","date":"2023-02-02","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2302.00845","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b695f21dcd217c0b2422","name":"Chemistry42","organization":"Insilico Medicine AI","date":"2023-02-02","domains":["Biology"],"country":"Hong Kong","reference":"https://pubs.acs.org/doi/full/10.1021/acs.jcim.2c01191","confidence":"Unknown","accessibility":""},{"id":"epoch:25ecdfeb539e283dbe31","name":"UniPi","organization":"Google DeepMind,Massachusetts Institute of Technology (MIT),University of California (UC) Berkeley,Georgia Institute of Technology,University of Alberta","date":"2023-01-31","domains":["Video","Robotics","Vision"],"country":"United States of America,United States of America,United States of America,United States of America,Canada","reference":"https://arxiv.org/abs/2302.00111","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:39b287947934302e3f4d","name":"Flan T5-XXL + BLIP-2","organization":"Salesforce Research","date":"2023-01-30","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2301.12597, https://huggingface.co/Salesforce/blip2-flan-t5-xl","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:000a247c9b6e909358ee","name":"BLIP-2 (Q-Former)","organization":"Salesforce Research","date":"2023-01-30","domains":["Vision","Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2301.12597","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b408a7aba63207114d6e","name":"KeAP","organization":"The University of Hong Kong,ByteDance,JancsiTech,OPPO HealthLab","date":"2023-01-30","domains":["Biology"],"country":"Hong Kong,China,China,China","reference":"https://arxiv.org/abs/2301.13154","confidence":"Likely","accessibility":""},{"id":"epoch:19a9e602185199a4d016","name":"Genie-SCOPe (bio)","organization":"Columbia University","date":"2023-01-29","domains":["Biology"],"country":"United States of America","reference":"https://arxiv.org/abs/2301.12485","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:34a5e26707ed3e440f05","name":"RESP AI","organization":"University of California San Diego","date":"2023-01-28","domains":["Biology"],"country":"United States of America","reference":"https://www.nature.com/articles/s41467-023-36028-8#Sec28","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d2181542376161c1406c","name":"Protst","organization":"Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Universit\u00e9 de Montr\u00e9al,Intel Labs,HEC Montreal,CIFAR AI Research","date":"2023-01-28","domains":["Biology"],"country":"Canada,Canada,United States of America,Canada,Canada","reference":"https://proceedings.mlr.press/v202/xu23t/xu23t.pdf","confidence":"Likely","accessibility":""},{"id":"epoch:f1f3807bf013c264ab7b","name":"DDPM-IP (CelebA)","organization":"Utrecht University","date":"2023-01-27","domains":["Image generation"],"country":"Netherlands","reference":"https://arxiv.org/abs/2301.11706v3","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fd4cf9ad929269fe9ac9","name":"MusicLM","organization":"Google","date":"2023-01-26","domains":["Audio"],"country":"United States of America","reference":"https://arxiv.org/abs/2301.11325","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2821dc129665df6cf790","name":"MoLFormer-XL","organization":"IBM","date":"2023-01-25","domains":["Biology"],"country":"United States of America","reference":"https://research.ibm.com/blog/molecular-transformer-discovery\nhttps://www.nature.com/articles/s42256-022-00580-7","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6a8bf224f64d5dd3d06a","name":"Moxiaoxian","organization":"","date":"2023-01-25","domains":[],"country":"","reference":"https://stcn.com/article/detail/1106222.html","confidence":"Likely","accessibility":""},{"id":"epoch:49c6681f5642ff20a932","name":"Pinshang","organization":"","date":"2023-01-25","domains":[],"country":"","reference":"https://stcn.com/article/detail/1106222.html","confidence":"Unknown","accessibility":""},{"id":"epoch:9750f2d555aa2e6bca54","name":"GPT-2+Active-SGD (WT2)","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2023-01-24","domains":["Language"],"country":"Canada","reference":"https://arxiv.org/abs/2301.10133","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:0f1ef23766ddd0ff496c","name":"mini-GPT-2+Active-AdamW","organization":"HEC Montreal","date":"2023-01-24","domains":["Language"],"country":"Canada","reference":"https://arxiv.org/abs/2301.10133","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:00849cf2bdcfc51493b6","name":"Operator","organization":"OpenAI","date":"2023-01-23","domains":["Language","Vision"],"country":"United States of America","reference":"https://openai.com/index/introducing-operator/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:a80292738f2f404e72d5","name":"Adaptive Agent","organization":"DeepMind","date":"2023-01-18","domains":["Games"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2301.07608","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:6125d25fb6f2ca2fe42e","name":"Ankh_large","organization":"Technical University of Munich,Columbia University","date":"2023-01-16","domains":["Biology"],"country":"Germany,United States of America","reference":"https://arxiv.org/abs/2301.06568","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:1b3eee4e8c54f35b8a59","name":"Ankh_base","organization":"Technical University of Munich,Columbia University","date":"2023-01-16","domains":["Biology"],"country":"Germany,United States of America","reference":"https://arxiv.org/abs/2301.06568","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:16406f4c79e749e4e3ee","name":"Eleven Monolingual v1","organization":"ElevenLabs","date":"2023-01-16","domains":["Speech"],"country":"United States of America","reference":"https://elevenlabs.io/docs/models#older-models","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:81dbf71ff2f79eb7f6de","name":"Nucleotide Transformer","organization":"Nvidia,Technical University of Munich,InstaDeep","date":"2023-01-15","domains":["Biology"],"country":"United States of America,Germany,United Kingdom","reference":"https://www.biorxiv.org/content/10.1101/2023.01.11.523679v1.full.pdf","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:69bfe4f9130be6b88680","name":"DreamerV3","organization":"DeepMind,University of Toronto","date":"2023-01-10","domains":["Games"],"country":"United Kingdom,Canada","reference":"https://arxiv.org/abs/2301.04104v1","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:7023650db951ad7309a2","name":"SantaCoder","organization":"Hugging Face,ServiceNow,Massachusetts Institute of Technology (MIT),Wellesley College,Saama,EleutherAI,Huawei Noah's Ark Lab,Carnegie Mellon University (CMU)","date":"2023-01-09","domains":["Language"],"country":"United States of America,United States of America,United States of America,United States of America,United States of America,United States of America,China,United States of America","reference":"https://arxiv.org/abs/2301.03988","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:5e3c6187e82022d188ff","name":"VALL-E","organization":"Microsoft","date":"2023-01-05","domains":["Audio","Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2301.02111","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:715fd5655e87dde49d04","name":"SparseOPT-175B","organization":"Institute of Science and Technology Austria (ISTA),Neural Magic","date":"2023-01-02","domains":["Language"],"country":"Austria,United States of America","reference":"https://arxiv.org/abs/2301.00774","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:546645a51bbce9def3cc","name":"SparseOPT-66B","organization":"Institute of Science and Technology Austria (ISTA)","date":"2023-01-02","domains":["Language"],"country":"Austria","reference":"https://arxiv.org/abs/2301.00774","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:451d71ab816b0f929229","name":"SparseOPT-13B","organization":"Institute of Science and Technology Austria (ISTA)","date":"2023-01-02","domains":["Language"],"country":"Austria","reference":"https://arxiv.org/abs/2301.00774","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:8e7f8acf4190e92f1eeb","name":"SparseOPT-30B","organization":"Institute of Science and Technology Austria (ISTA)","date":"2023-01-02","domains":["Language"],"country":"Austria","reference":"https://arxiv.org/abs/2301.00774","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:02d17a24d4f3bfff11d5","name":"DNA Fine-Tuned Language Model (DFLM)","organization":"Tongji University","date":"2023-01-02","domains":["Biology"],"country":"China","reference":"https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9751352","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f4e76a3449b066682e5e","name":"AI Resume","organization":"","date":"2023-01-01","domains":[],"country":"","reference":"https://www.163.com/news/article/IPUPCUV000019UD6.html\n(LINK NOW BROKEN as of 8/14/25)","confidence":"Unknown","accessibility":""},{"id":"epoch:ad204d69c1db5ac9c4c4","name":"Nanbei Pavilion","organization":"","date":"2023-01-01","domains":[],"country":"","reference":"https://www.careersciencelab.com/ser-detail?id=56","confidence":"Unknown","accessibility":""},{"id":"epoch:c3b3d7a6874142247d42","name":"Hybrid H3-2.7B","organization":"Stanford University,University at Buffalo","date":"2022-12-28","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2212.14052","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:23d821c7e55701e3ebd2","name":"Hybrid H3-125M","organization":"Stanford University,University at Buffalo","date":"2022-12-28","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2212.14052","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b20fac4e5673b80208e8","name":"Hybrid H3-355M","organization":"Stanford University,University at Buffalo","date":"2022-12-28","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2212.14052","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7f43478679e5e76e07ea","name":"Hybrid H3-1.3B","organization":"Stanford University,University at Buffalo","date":"2022-12-28","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2212.14052","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:063ffdecb28b89c3f1ba","name":"OPT-IML (175B)","organization":"Meta AI","date":"2022-12-22","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2212.12017","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:2db664e4bb357a7148c5","name":"CaLM","organization":"University of Oxford","date":"2022-12-19","domains":["Biology"],"country":"United Kingdom","reference":"https://www.biorxiv.org/content/10.1101/2022.12.15.519894v1.full.pdf","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:39329bc008c20191dfcb","name":"text-embedding-ada-002","organization":"OpenAI","date":"2022-12-15","domains":["Language"],"country":"United States of America","reference":"https://openai.com/index/new-and-improved-embedding-model/","confidence":"Unknown","accessibility":""},{"id":"epoch:9103dae4a4df04ebabd2","name":"RT-1","organization":"Google","date":"2022-12-13","domains":["Robotics"],"country":"United States of America","reference":"https://arxiv.org/abs/2212.06817","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6d26a4879fed8e147f82","name":"XiaoAI 6.0","organization":"Xiaomi Corp","date":"2022-12-11","domains":[],"country":"China","reference":"https://www.ithome.com/0/660/360.htm","confidence":"Unknown","accessibility":""},{"id":"epoch:96b2b65e5911ba212fb6","name":"TranceptEve","organization":"University of Oxford,Harvard Medical School","date":"2022-12-10","domains":["Biology"],"country":"United Kingdom,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2022.12.07.519495v1","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:eaa8a96e87efc6ad58da","name":"Stable Diffusion 2.1","organization":"Stability AI","date":"2022-12-07","domains":["Image generation"],"country":"United Kingdom","reference":"https://stability.ai/news/stablediffusion2-1-release7-dec-2022","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f491046ed9555f9d13d8","name":"Transformer-XL + FWL","organization":"Google Research","date":"2022-12-05","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2212.02475","confidence":"","accessibility":"Unreleased"},{"id":"epoch:25dae3f8de98cdc2cb1b","name":"Whisper v2","organization":"OpenAI","date":"2022-12-05","domains":["Speech"],"country":"United States of America","reference":"https://huggingface.co/openai/whisper-large-v2\n\nhttps://arxiv.org/abs/2212.04356","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c703bcf164b8a9a21c95","name":"Vega v2","organization":"Wuhan University,JD Explore Academy,Shanghai AI Lab,Nanyang Technological University,Washington University in St Louis,Chongqing University of Posts and Telecommunications,University of Sydney","date":"2022-12-04","domains":["Language"],"country":"China,China,Singapore,United States of America,China,Australia","reference":"https://arxiv.org/abs/2212.01853","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:06ae871037a94514f103","name":"Transformer + GFM","organization":"Nanjing University","date":"2022-12-01","domains":["Language"],"country":"China","reference":"https://cs.nju.edu.cn/wujx/paper/AAAI2023_AFM.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e1b318ad26dd9b9a4360","name":"ZymCTRL","organization":"Basecamp Research,Friedrich-Alexander-Universit\u00e4t,University of Girona","date":"2022-12-01","domains":["Biology"],"country":"United Kingdom,Germany,Spain","reference":"https://www.mlsb.io/papers_2022/ZymCTRL_a_conditional_language_model_for_the_controllable_generation_of_artificial_enzymes.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:09392d62b6bcb80a0630","name":"DeepNash","organization":"DeepMind","date":"2022-12-01","domains":["Games"],"country":"United Kingdom","reference":"https://www.science.org/stoken/author-tokens/ST-887/full","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:59066b70355c90f989c1","name":"ALM 1.0","organization":"Beijing Academy of Artificial Intelligence / BAAI","date":"2022-11-28","domains":["Language"],"country":"China","reference":"https://github.com/FlagAI-Open/FlagAI/blob/master/examples/ALM/README.md","confidence":"Speculative","accessibility":"Hosted access (no API)"},{"id":"epoch:13a9cf1ae7cc16ac5cf8","name":"DiT-XL/2 + Discriminator Guidance","organization":"Korea Advanced Institute of Science and Technology (KAIST),NAVER","date":"2022-11-28","domains":["Image generation"],"country":"South Korea,South Korea","reference":"https://arxiv.org/abs/2211.17091v4","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:296adaffe3e5361d0a4e","name":"Discriminator Guidance","organization":"Korea Advanced Institute of Science and Technology (KAIST),NAVER","date":"2022-11-28","domains":["Image generation"],"country":"South Korea,South Korea","reference":"https://arxiv.org/abs/2211.17091v4","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:6f433a1a3c1f536ed1d2","name":"GPT-3.5 (davinci-003)","organization":"OpenAI","date":"2022-11-28","domains":["Language"],"country":"United States of America","reference":"","confidence":"Likely","accessibility":""},{"id":"epoch:1a4ebaef0d9714e3e36f","name":"CICERO","organization":"Meta AI","date":"2022-11-22","domains":["Games"],"country":"United States of America","reference":"https://www.science.org/doi/10.1126/science.ade9097","confidence":"Unknown","accessibility":"Open weights (non-commercial)"},{"id":"epoch:6c23db050f8b67f5c18a","name":"CLUE","organization":"Naver Clova,Naver AI Lab","date":"2022-11-22","domains":["Recommendation"],"country":"South Korea,South Korea","reference":"https://arxiv.org/abs/2111.11294","confidence":"Likely","accessibility":""},{"id":"epoch:ee0db01cc2d045893907","name":"scFormer","organization":"University of Toronto,Vector Institute,University Health Network,Microsoft Research","date":"2022-11-22","domains":["Biology"],"country":"Canada,Canada,Canada,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2022.11.20.517285v1.abstract","confidence":"Speculative","accessibility":""},{"id":"epoch:a7ca0f8bb69515d0c1b3","name":"AR-LDM","organization":"Alibaba,University of Waterloo,Vector Institute","date":"2022-11-20","domains":["Image generation"],"country":"China,Canada,Canada","reference":"https://arxiv.org/abs/2211.10950","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:a1e534abaef6b095c719","name":"Fusion in Encoder","organization":"Samsung","date":"2022-11-18","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/abs/2211.10147","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:d2e891942726bcd9119b","name":"Galactica","organization":"Meta AI","date":"2022-11-16","domains":["Language","Biology"],"country":"United States of America","reference":"https://arxiv.org/abs/2211.09085","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:6dccc88ff0b3c056f0d8","name":"Luminous Sparse","organization":"Aleph Alpha,Graphcore","date":"2022-11-16","domains":["Language"],"country":"Germany,United Kingdom","reference":"https://www.graphcore.ai/posts/graphcore-and-aleph-alpha-demonstrate-80-sparsified-ai-model","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:bea16d135a7e4ee85876","name":"EVA-01","organization":"Beijing Academy of Artificial Intelligence / BAAI,Huazhong University of Science and Technology,Zhejiang University (ZJU),Beijing Institute of Technology","date":"2022-11-14","domains":["Vision"],"country":"China,China,China,China","reference":"https://arxiv.org/abs/2211.07636","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3eb18cf9d752016dcf4c","name":"AltCLIP_M9","organization":"Beijing Academy of Artificial Intelligence / BAAI","date":"2022-11-12","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://arxiv.org/abs/2211.06679","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9c0f118dc147a6951b17","name":"Innovative Drug-like Molecule Generation from","organization":"University of Pittsburgh,Carnegie Mellon University (CMU)","date":"2022-11-12","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2211.06566","confidence":"Unknown","accessibility":""},{"id":"epoch:2fa9f21e8deaeac0e651","name":"InternImage","organization":"Shanghai AI Lab,Tsinghua University,Nanjing University,SenseTime,Chinese University of Hong Kong (CUHK)","date":"2022-11-10","domains":["Vision"],"country":"China,China,China,Hong Kong,Hong Kong","reference":"https://arxiv.org/abs/2211.05778","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f712ad6b3ffc8a9c7bba","name":"Mogrifier RLSTM (WT2)","organization":"DeepMind","date":"2022-11-03","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2211.01848","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:1617617569438323317c","name":"Mogrifier RLSTM (PTB)","organization":"DeepMind","date":"2022-11-03","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2211.01848","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:142a651e532613e8a003","name":"BLOOMZ-176B","organization":"Hugging Face","date":"2022-11-03","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2211.01786, https://huggingface.co/bigscience/bloomz","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:012d489cd5477f24279c","name":"mT0-13B","organization":"Hugging Face,BigScience","date":"2022-11-03","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/2211.01786","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:638b20a4ff6166bb4601","name":"eDiff-I","organization":"Nvidia","date":"2022-11-02","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2211.01324","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:6484c2900c8e56204d08","name":"GearNet","organization":"Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Universit\u00e9 de Montr\u00e9al,University of Cambridge,IBM Research,HEC Montreal,CIFAR AI Research","date":"2022-11-01","domains":["Biology"],"country":"Canada,Canada,United Kingdom,United States of America,Canada,Canada","reference":"https://arxiv.org/abs/2203.06125","confidence":"Confident","accessibility":""},{"id":"epoch:3dd21ae3a20b8096cf11","name":"Taiyi-Stable Diffusion","organization":"IDEA CCNL","date":"2022-10-31","domains":["Image generation"],"country":"China","reference":"https://arxiv.org/abs/2209.02970","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:4e5c01cec4cee8e9630e","name":"Transformer-XL + PowerSGD + L-Greco","organization":"Institute of Science and Technology Austria (ISTA),Neural Magic","date":"2022-10-31","domains":["Language"],"country":"Austria,United States of America","reference":"https://arxiv.org/abs/2210.17357","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2cf592f30732a1f9e9ca","name":"retrieval-quality-kNN-LMs","organization":"University of Massachusetts Amherst","date":"2022-10-28","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2210.15859","confidence":"","accessibility":"Unreleased"},{"id":"epoch:59f831ffd6061c6dc9de","name":"ADP-FAIRSEQ + NGRAMRES","organization":"Nara Institute of Science and Technology,Chinese University of Hong Kong (CUHK),Tsinghua University","date":"2022-10-26","domains":["Language"],"country":"Japan,Hong Kong,China","reference":"https://arxiv.org/abs/2210.14431","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:ce4e6fe05661e41a9b57","name":"MIF-ST","organization":"Microsoft Research,OpenBioML,University of Chicago","date":"2022-10-26","domains":["Biology"],"country":"United States of America,Multinational,United States of America","reference":"https://academic.oup.com/peds/article-abstract/doi/10.1093/protein/gzad015/7330543?redirectedFrom=fulltext#no-access-message","confidence":"Confident","accessibility":""},{"id":"epoch:bd022d01c3024df46b8f","name":"XY-LENTXL","organization":"Microsoft","date":"2022-10-26","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2210.14867","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2ddacff548313e0c20d8","name":"Verbatim Memory Transformer (182M)","organization":"Johns Hopkins University,New York University (NYU)","date":"2022-10-24","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2210.13569","confidence":"","accessibility":"Unreleased"},{"id":"epoch:1b24d7e68f9edb3e7080","name":"Verbatim Memory Transformer (117M)","organization":"Johns Hopkins University,New York University (NYU)","date":"2022-10-24","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2210.13569","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:1f36bdc5ca6c9001dfc8","name":"Verbatim Memory Transformer (108M)","organization":"Johns Hopkins University,New York University (NYU)","date":"2022-10-24","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2210.13569","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:017d7016d30fa34a1f8b","name":"EnCodec","organization":"Meta AI","date":"2022-10-24","domains":["Audio"],"country":"United States of America","reference":"https://arxiv.org/abs/2210.13438, ","confidence":"Unknown","accessibility":"Open weights (non-commercial)"},{"id":"epoch:1f33408dac67362321a9","name":"Tk-Instruct","organization":"University of Washington,Arizona State University,Allen Institute for AI","date":"2022-10-24","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2204.07705","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c8b8d2188d9df20488cc","name":"DiffSBDD (CrossDocked)","organization":"Ecole Polytechnique F\u00b4ed\u00b4erale de Lausanne (EPFL),University of Cambridge,Cornell University,Chinese Academy of Mathematics and System Science,University of Rome,Microsoft Research,University of Oxford,AITHYRA Institute","date":"2022-10-24","domains":["Biology"],"country":"Switzerland,United Kingdom,United States of America,China,Italy,United States of America,United Kingdom,Austria","reference":"https://arxiv.org/abs/2210.13695","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a923052564d296e096bc","name":"U-PaLM (540B)","organization":"Google","date":"2022-10-20","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2210.11399","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b4d84038ceba962e36ab","name":"LMSI-Palm","organization":"Google,University of Illinois Urbana-Champaign (UIUC)","date":"2022-10-20","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2210.11610","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:914896617440aaa755f6","name":"Flan-PaLM 540B","organization":"Google","date":"2022-10-20","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2210.11416","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:64ea93a0805dbe3f82b5","name":"Flan-T5 11B","organization":"Google","date":"2022-10-20","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2210.11416, https://huggingface.co/google/flan-t5-xxl","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d958c2cbd6bbf77705b3","name":"GPT-2 + Progressive LRD","organization":"Huawei,Huawei Noah's Ark Lab","date":"2022-10-12","domains":["Language"],"country":"China,China","reference":"https://neurips2022-enlsp.github.io/papers/paper_33.pdf","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:da048b33ea4b147e4544","name":"GenSLM","organization":"University of Chicago,Nvidia,Harvard University,Cerebras Systems,Technical University of Munich,California Institute of Technology","date":"2022-10-11","domains":["Biology"],"country":"United States of America,United States of America,United States of America,United States of America,Germany,United States of America","reference":"https://www.biorxiv.org/content/biorxiv/early/2022/10/11/2022.10.10.511571.full.pdf","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:071aa4d7c3e8ea0cc7b2","name":"Instruct-GPT + Mind's Eye","organization":"Google,Dartmouth College","date":"2022-10-11","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2210.05359","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:523bf8d43606896d1df0","name":"Diplodocus","organization":"Meta AI,Massachusetts Institute of Technology (MIT)","date":"2022-10-11","domains":["Games"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2210.05492","confidence":"Unknown","accessibility":"Open weights (non-commercial)"},{"id":"epoch:e94d847b96aeb67123f3","name":"TPM-LVD","organization":"University of California Los Angeles (UCLA)","date":"2022-10-10","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2210.04398","confidence":"","accessibility":"Unreleased"},{"id":"epoch:5326f27b63885ef623a2","name":"Decaying Fast Weights Transformer (WT-103)","organization":"Jenni","date":"2022-10-09","domains":["Language"],"country":"Singapore","reference":"https://arxiv.org/abs/2210.04243","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:be520cca5fb3b1dc297e","name":"Phenaki","organization":"University College London (UCL),University of Michigan,Google Brain","date":"2022-10-05","domains":["Video"],"country":"United Kingdom,United States of America,United States of America","reference":"https://arxiv.org/abs/2210.02399","confidence":"","accessibility":"Unreleased"},{"id":"epoch:982db4ad746bbeb9908e","name":"Imagen Video","organization":"Google Brain","date":"2022-10-05","domains":["Video"],"country":"United States of America","reference":"https://arxiv.org/abs/2210.02303","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4b2943647fb6beffe18d","name":"AlphaTensor","organization":"DeepMind","date":"2022-10-05","domains":["Other","Games","Mathematics"],"country":"United Kingdom","reference":"https://www.nature.com/articles/s41586-022-05172-4","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:247a7fd6ae7b1ad465f4","name":"VRNS-RNN-3-3-5","organization":"University of Notre Dame","date":"2022-10-04","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2210.01343","confidence":"","accessibility":"Unreleased"},{"id":"epoch:e248a8397a01373a7dad","name":"DiffDock","organization":"Massachusetts Institute of Technology (MIT)","date":"2022-10-04","domains":["Biology"],"country":"United States of America","reference":"https://arxiv.org/abs/2210.01776, https://docs.nvidia.com/bionemo-framework/latest/models/diffdock.html","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d919944edbaaaa98f145","name":"NMST+GPT-2","organization":"New York University (NYU)","date":"2022-10-03","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2210.00660","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:af07c746992fb4130662","name":"AminoBert","organization":"Harvard Medical School,Nabla Bio,Columbia University","date":"2022-10-03","domains":["Biology"],"country":"United States of America,United States of America,United States of America","reference":"https://www.nature.com/articles/s41587-022-01432-w","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:9c47673da5474ea33d97","name":"GemNet-OC","organization":"Technical University of Munich,Carnegie Mellon University (CMU),Facebook AI Research","date":"2022-09-30","domains":["Materials science"],"country":"Germany,United States of America,United States of America,France","reference":"https://arxiv.org/abs/2204.02782","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e7e2880763b0258f6ad4","name":"Make-A-Video","organization":"Meta AI","date":"2022-09-29","domains":["Video"],"country":"United States of America","reference":"https://arxiv.org/abs/2209.14792","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:1023fb967191b3443ce5","name":"Sparrow","organization":"DeepMind","date":"2022-09-28","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2209.14375","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:41afa5636229ce6c1665","name":"CPM-Ant","organization":"Tsinghua University,Beijing Academy of Artificial Intelligence / BAAI,ModelBest,OpenBMB (Open Lab for Big Model Base)","date":"2022-09-22","domains":["Language"],"country":"China,China,China,China","reference":"https://github.com/OpenBMB/CPM-Live/tree/cpm-ant/cpm-live#model-checkpoints","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:27482b52c5711ccbc3c9","name":"Whisper","organization":"OpenAI","date":"2022-09-21","domains":["Speech"],"country":"United States of America","reference":"https://cdn.openai.com/papers/whisper.pdf\n\nhttps://arxiv.org/abs/2212.04356","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7599287795c918b70b05","name":"DistilProtBert","organization":"Bar-Ilan University","date":"2022-09-18","domains":["Biology"],"country":"Israel","reference":"https://academic.oup.com/bioinformatics/article/38/Supplement_2/ii95/6701995","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2ccd11d053c1535776d5","name":"CPAC","organization":"Texas A&M","date":"2022-09-18","domains":["Biology"],"country":"United States of America","reference":"https://academic.oup.com/bioinformatics/article/38/Supplement_2/ii68/6702011","confidence":"Speculative","accessibility":""},{"id":"epoch:4725c5978bb6730e0602","name":"CLIP ViT-H/14 - LAION-2B","organization":"LAION","date":"2022-09-15","domains":["Vision"],"country":"Germany","reference":"https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f44e043900dc06b35238","name":"NeMO Megatron GPT 20B","organization":"Nvidia","date":"2022-09-15","domains":["Language"],"country":"United States of America","reference":"https://huggingface.co/nvidia/nemo-megatron-gpt-20B","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:89e3edf875d5ad36ea53","name":"ProteinMPNN","organization":"University of Washington,Wageningen University and Research,NERSC, Lawrence Berkeley National Laboratory","date":"2022-09-15","domains":["Biology"],"country":"United States of America,Netherlands,United States of America","reference":"https://www.science.org/doi/full/10.1126/science.add2187","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fc8a0df87b5dc8a2bbff","name":"PaLI","organization":"Google","date":"2022-09-14","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://arxiv.org/abs/2209.06794v4","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:9f2d46ca003f40447235","name":"Deep-LDA","organization":"Fudan University","date":"2022-09-14","domains":["Biology"],"country":"China","reference":"https://www.biorxiv.org/content/10.1101/2022.09.12.507562v1.full","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:5f73a70d93e2e54dd394","name":"SauTech","organization":"Saudi Data and Artificial Intelligence Authority,Saudi Company for Artificial Intelligence","date":"2022-09-14","domains":["Speech","Audio"],"country":"Saudi Arabia,Saudi Arabia","reference":"https://globalaisummit.org/en/News/Pages/NewsDetails.aspx?NewsId=73\nhttps://saudigazette.com.sa/article/624984/SAUDI-ARABIA/SDAIA-SCAI-unveil-SauTech-speech-to-text-software","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:ec9cad24cc69b6b93166","name":"ACT-1","organization":"","date":"2022-09-14","domains":["Language"],"country":"","reference":"https://www.adept.ai/act","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:24e0fbe3822fc0801e66","name":"BEIT-3","organization":"Microsoft","date":"2022-08-22","domains":["Multimodal","Vision","Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2208.10442","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:edd18ba046e1f097442f","name":"Stable Diffusion 1.5","organization":"Runway","date":"2022-08-22","domains":["Image generation"],"country":"United States of America","reference":"https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:6bc1eed009c75eb1f537","name":"Stable Diffusion 1.2","organization":"Ludwig Maximilian University of Munich","date":"2022-08-22","domains":["Image generation"],"country":"Germany","reference":"https://huggingface.co/CompVis/stable-diffusion-v1-2","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:a4e69d9483133b88f7ed","name":"Stable Diffusion 1.4","organization":"Ludwig Maximilian University of Munich","date":"2022-08-22","domains":["Image generation"],"country":"Germany","reference":"https://huggingface.co/CompVis/stable-diffusion-v1-4","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:e55c06e5963da75c8691","name":"Stable Diffusion 1.1","organization":"Ludwig Maximilian University of Munich","date":"2022-08-22","domains":["Image generation"],"country":"Germany","reference":"https://huggingface.co/CompVis/stable-diffusion-v1-1","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:abb469a0e09322ae7b15","name":"PaLM-SayCan","organization":"Google","date":"2022-08-16","domains":["Robotics"],"country":"United States of America","reference":"https://arxiv.org/abs/2204.01691","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:67bfd5e56b37747cf53a","name":"Luminous-supreme","organization":"Aleph Alpha","date":"2022-08-15","domains":["Language"],"country":"Germany","reference":"https://docs.aleph-alpha.com/docs/introduction/model-card/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:6a6359c298852b46197f","name":"Luminous-extended","organization":"Aleph Alpha","date":"2022-08-15","domains":["Language"],"country":"Germany","reference":"https://docs.aleph-alpha.com/docs/Deprecated%20Luminous/Deprecated-Luminous/model-card/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:8ac70764a8eaaa9c5e40","name":"Luminous-base","organization":"Aleph Alpha","date":"2022-08-15","domains":["Language"],"country":"Germany","reference":"https://docs.aleph-alpha.com/docs/Deprecated%20Luminous/Deprecated-Luminous/model-card/","confidence":"Confident","accessibility":"API access"},{"id":"epoch:89e4b362492a3351e625","name":"Dream Diary (\u9020\u68a6\u65e5\u8bb0)","organization":"West Lake Xinchen / Xinchen AI / \u897f\u6e56\u5fc3\u8fb0\uff08\u676d\u5dde\uff09\u79d1\u6280\u6709\u9650\u516c\u53f8","date":"2022-08-15","domains":["Image generation"],"country":"China","reference":"https://zmrj.art/\nhttps://zhuanlan.zhihu.com/p/581520482","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:f442682820d5e93b10b3","name":"PeTriBERT","organization":"University of Montpellier,BionomeeX","date":"2022-08-13","domains":["Biology"],"country":"France,France","reference":"https://www.biorxiv.org/content/10.1101/2022.08.10.503344v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:f6147d9b5b49ed24233c","name":"M3GNet","organization":"University of California San Diego","date":"2022-08-11","domains":["Materials science"],"country":"United States of America","reference":"https://arxiv.org/abs/2202.02450","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a70a6e122ef72b60101e","name":"BlenderBot 3","organization":"McGill University,Meta AI,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms)","date":"2022-08-10","domains":["Language"],"country":"Canada,United States of America,Canada","reference":"https://arxiv.org/abs/2208.03188, https://github.com/facebookresearch/ParlAI/blob/main/parlai/zoo/bb3/model_card.md\n\ntraining code: https://parl.ai/projects/bb3/ ","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:5c9d1bbdb2a55c1165c9","name":"RNA-FM","organization":"Chinese University of Hong Kong (CUHK),Fudan University,Shanghai AI Lab,Harbin Institute of Technology,University of Electronic Science and Technology of China,Massachusetts Institute of Technology (MIT),Harvard University,Shanghai Zelixir Biotech,CUHK Shenzhen Research Institute","date":"2022-08-08","domains":["Biology"],"country":"Hong Kong,China,China,China,China,United States of America,United States of America,China,China","reference":"https://arxiv.org/abs/2204.00300","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3c54ee6d024b87967ffd","name":"FastSpeech 2","organization":"Zhejiang University (ZJU),Microsoft Research Asia","date":"2022-08-08","domains":["Speech"],"country":"China,China","reference":"https://arxiv.org/abs/2006.04558","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:16c34f9da8627ca131b7","name":"SGPT BE 5.8B","organization":"Peking University","date":"2022-08-05","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2202.08904","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:01a3f8d13fbc150dc6d0","name":"GLM-130B","organization":"Tsinghua University","date":"2022-08-04","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2210.02414","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:6d8e7837926bcb0ee5e0","name":"AlexaTM 20B","organization":"Amazon","date":"2022-08-02","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2208.01448","confidence":"Confident","accessibility":"API access"},{"id":"epoch:7a3fa2d7e2588672e88f","name":"ProtGPT2","organization":"University of Bayreuth","date":"2022-07-27","domains":["Biology"],"country":"Germany","reference":"https://www.nature.com/articles/s41467-022-32007-7","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:678fc4ffa5c953ebbe13","name":"GPT-NeoX-Japanese","organization":"Abeja","date":"2022-07-27","domains":["Language"],"country":"Japan","reference":"https://huggingface.co/abeja/gpt-neox-japanese-2.7b, https://tech-blog.abeja.asia/entry/abeja-gpt-project-202207","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d61f8cccc10584535ca7","name":"OmegaPLM","organization":"Massachusetts Institute of Technology (MIT),Westlake University","date":"2022-07-22","domains":["Biology"],"country":"United States of America,China","reference":"https://www.biorxiv.org/content/10.1101/2022.07.21.500999v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b136b42f7f24f8f74999","name":"ESM2-15B","organization":"Meta AI,New York University (NYU),Stanford University,Massachusetts Institute of Technology (MIT)","date":"2022-07-21","domains":["Biology"],"country":"United States of America,United States of America,United States of America,United States of America","reference":"https://www.science.org/doi/abs/10.1126/science.ade2574\nhttps://www.biorxiv.org/content/10.1101/2022.07.20.500902v2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1d90515f16d445f99ed1","name":"ESM2-3B","organization":"Meta AI,New York University (NYU),Stanford University,Massachusetts Institute of Technology (MIT)","date":"2022-07-21","domains":["Biology"],"country":"United States of America,United States of America,United States of America,United States of America","reference":"https://www.science.org/doi/abs/10.1126/science.ade2574","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:67634548355656e5c40b","name":"ESM2-650M","organization":"Meta AI,New York University (NYU),Stanford University,Massachusetts Institute of Technology (MIT)","date":"2022-07-21","domains":["Biology"],"country":"United States of America,United States of America,United States of America,United States of America","reference":"https://www.science.org/doi/abs/10.1126/science.ade2574","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d5c26b8185959f392627","name":"ESM2-150M","organization":"Meta AI,New York University (NYU),Stanford University,Massachusetts Institute of Technology (MIT)","date":"2022-07-21","domains":["Biology"],"country":"United States of America,United States of America,United States of America,United States of America","reference":"https://www.science.org/doi/abs/10.1126/science.ade2574","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:bf24aab3ee8e0193351b","name":"ESM2-35M","organization":"Meta AI,New York University (NYU),Stanford University,Massachusetts Institute of Technology (MIT)","date":"2022-07-21","domains":["Biology"],"country":"United States of America,United States of America,United States of America,United States of America","reference":"https://www.science.org/doi/abs/10.1126/science.ade2574","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5cb729103d3c98f8031d","name":"ESM2-8M","organization":"Meta AI,New York University (NYU),Stanford University,Massachusetts Institute of Technology (MIT)","date":"2022-07-21","domains":["Biology"],"country":"United States of America,United States of America,United States of America,United States of America","reference":"https://www.science.org/doi/abs/10.1126/science.ade2574","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f3517279d4b3272b73fa","name":"YuYan 11B","organization":"Hong Kong Baptist University,NetEase","date":"2022-07-15","domains":["Language"],"country":"Hong Kong,China","reference":"https://arxiv.org/abs/2104.12470","confidence":"Confident","accessibility":""},{"id":"epoch:74623e36e431f799c5de","name":"Transformer-XL + RMT","organization":"Moscow Institute of Physics and Technology,AIRI Artificial Intelligence Research Institute","date":"2022-07-14","domains":["Language"],"country":"Russia,Russia","reference":"https://arxiv.org/abs/2207.06881","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:8684289d4c0ef11e2765","name":"Rita-XLarge","organization":"LightOn,Harvard University,University of Oxford","date":"2022-07-14","domains":["Biology"],"country":"France,United States of America,United Kingdom","reference":"https://arxiv.org/abs/2205.05789","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:aaf478f87269a2ba3df8","name":"Delphi","organization":"Allen Institute for AI,University of Washington","date":"2022-07-12","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2110.07574","confidence":"Speculative","accessibility":"Hosted access (no API)"},{"id":"epoch:be9b8a581be359066ea1","name":"BLOOM-176B","organization":"Hugging Face,BigScience","date":"2022-07-11","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/2211.05100","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:85d1fa8b33f176cc60f5","name":"NLLB","organization":"Meta AI","date":"2022-07-06","domains":["Language"],"country":"United States of America","reference":"https://research.facebook.com/publications/no-language-left-behind/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5bfd56b50468e2dba11c","name":"BLOOM-7.1B","organization":"Hugging Face,BigScience","date":"2022-07-05","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/2211.05100","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:da614b9af1862f2b97d3","name":"CodeT5-large","organization":"Salesforce","date":"2022-07-05","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2207.01780","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c154f7d8f1aba24c90e3","name":"BLOOM-1.7B","organization":"Hugging Face,BigScience","date":"2022-07-05","domains":["Language"],"country":"United States of America,France","reference":"https://huggingface.co/bigscience/bloom-3b","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:a3e4037cfca4bb0542e6","name":"BLOOM-560M","organization":"Hugging Face,BigScience","date":"2022-07-05","domains":["Language"],"country":"United States of America,France","reference":"https://huggingface.co/bigscience/bloom-3b","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:ebc6049045f95b3b511d","name":"BLOOM-1B","organization":"Hugging Face,BigScience","date":"2022-07-05","domains":["Language"],"country":"United States of America,France","reference":"https://huggingface.co/bigscience/bloom-3b","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:c68e6076e9873974d5c0","name":"BLOOM-3B","organization":"Hugging Face,BigScience","date":"2022-07-05","domains":["Language"],"country":"United States of America,France","reference":"https://huggingface.co/bigscience/bloom-3b","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:23f1fc5ff64792d215b7","name":"WebGPT","organization":"OpenAI","date":"2022-07-01","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2112.09332","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:267061eeb43dac4b1320","name":"Drahim PFM AI","organization":"Drahim","date":"2022-06-30","domains":["Language"],"country":"Saudi Arabia","reference":"","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:5f91e7d88eb86ecf7c2c","name":"Minerva (540B)","organization":"Google","date":"2022-06-29","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2206.14858","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:3e4908b91c159f2e5019","name":"DALL-E mega","organization":"Craiyon","date":"2022-06-28","domains":["Image generation"],"country":"United States of America","reference":"https://huggingface.co/dalle-mini/dalle-mega\nhttps://github.com/borisdayma/dalle-mini","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0de8afd45ab5664cf3ac","name":"ProGen2-xlarge","organization":"Salesforce Research,Columbia University,Johns Hopkins University","date":"2022-06-27","domains":["Biology"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2206.13517","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5462d2a7b3583b35b590","name":"ProGen2-base","organization":"Salesforce Research,Columbia University,Johns Hopkins University","date":"2022-06-27","domains":["Biology"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2206.13517","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c323f4dcb307bb217564","name":"GPT-SW3","organization":"AI Sweden,RISE","date":"2022-06-25","domains":["Language"],"country":"Sweden,Sweden","reference":"http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf\n","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:ce36b430aad1983de1c5","name":"CodeWhisperer","organization":"Amazon","date":"2022-06-24","domains":["Language"],"country":"United States of America","reference":"https://aws.amazon.com/blogs/machine-learning/introducing-amazon-codewhisperer-the-ml-powered-coding-companion/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:f16fb17630ce9d321658","name":"YaLM","organization":"Yandex","date":"2022-06-23","domains":["Language"],"country":"Russia","reference":"https://medium.com/yandex/yandex-publishes-yalm-100b-its-the-largest-gpt-like-neural-network-in-open-source-d1df53d0e9a6","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a0c8320d7538133f17f7","name":"Parti","organization":"Google Research","date":"2022-06-22","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2206.10789v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:30bcbbcf121ab1168e12","name":"CodeGeeX","organization":"Z.ai (Zhipu AI),Tsinghua University","date":"2022-06-22","domains":["Language"],"country":"China,China","reference":"https://github.com/THUDM/CodeGeeX","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:1bc806f6fca763eb18c8","name":"OPT-2.7B (finetuned on PTB)","organization":"Meta AI","date":"2022-06-21","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.01068","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f17ec7fd1c3d2f3ffc5b","name":"OPT-1.3B","organization":"Meta AI","date":"2022-06-21","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.01068","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:dbaab1977f0f32ba0de3","name":"OPT-1.3B (finetuned on PTB)","organization":"Meta AI","date":"2022-06-21","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.01068","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:82527e608af0b64acdbe","name":"OPT-2.7B (finetuned on WT2)","organization":"Meta AI","date":"2022-06-21","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.01068","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8d3df3c534965b013c48","name":"OPT-125M (finetuned)","organization":"Meta AI","date":"2022-06-21","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.01068","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3dafb903bb49a6034c51","name":"OPT-6.7B","organization":"Meta AI","date":"2022-06-21","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.01068","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:00eda65e5d1285b030d2","name":"OPT-66B","organization":"Meta AI","date":"2022-06-21","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.01068","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:be5019511dfd72a295ea","name":"OPT-350M","organization":"Meta AI","date":"2022-06-21","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.01068","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:1791311955d86aee7b72","name":"OPT-2.7B","organization":"Meta AI","date":"2022-06-21","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.01068","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:053bf3d6cabaec22d0d1","name":"OPT-125M (finetuned on PTB)","organization":"Meta AI","date":"2022-06-21","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.01068","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d69cdf3effda6ded5664","name":"OPT-30B","organization":"Meta AI","date":"2022-06-21","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.01068","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:207ebbda5a2f74981cf0","name":"OPT-1.3B (finetuned)","organization":"Meta AI","date":"2022-06-21","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.01068","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0d9cd359da6a4494b07c","name":"Unified-IO (XL)","organization":"Allen Institute for AI,University of Washington","date":"2022-06-17","domains":["Multimodal","Vision","Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2206.08916","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:cd4726333c93d4902cc1","name":"CoCa","organization":"Google Research","date":"2022-06-14","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.01917v2","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:52dcd5387855ed5dd226","name":"MetaLM","organization":"Microsoft Research","date":"2022-06-13","domains":["Multimodal","Language","Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2206.06336v1","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:4665707e50f970a601bd","name":"EGRU (WT2)","organization":"Ruhr University Bochum,Technische Universit\u00e4t Dresden,University of London","date":"2022-06-13","domains":["Language"],"country":"Germany,Germany,United Kingdom","reference":"https://arxiv.org/abs/2206.06178v3","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:fb38a29c2ee0fd140456","name":"EGRU (PTB)","organization":"Ruhr University Bochum,Technische Universit\u00e4t Dresden,University of London","date":"2022-06-13","domains":["Language"],"country":"Germany,Germany,United Kingdom","reference":"https://arxiv.org/abs/2206.06178v3","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c1c86d4b1d5403910796","name":"BIG-G 137B","organization":"Google","date":"2022-06-09","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2206.04615","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b81b0a6172e55cb0e601","name":"LIMoE-H/14","organization":"Google","date":"2022-06-06","domains":["Multimodal","Vision","Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2206.02770","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0d30cb418d18f6e9c93e","name":"DITTO","organization":"Tsinghua University,Apple,Westlake University,Chinese University of Hong Kong (CUHK)","date":"2022-06-06","domains":["Language"],"country":"China,United States of America,China,Hong Kong","reference":"https://arxiv.org/abs/2206.02369","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f791f414695848078930","name":"Diffusion-GAN","organization":"UT Austin,Microsoft","date":"2022-06-05","domains":["Image generation"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2206.02262v4","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:849c434615f20f1e8ef0","name":"B2T connection (16L)","organization":"LINE Corporation,Tohoku University","date":"2022-06-01","domains":["Language"],"country":"Japan,Japan","reference":"https://arxiv.org/abs/2206.00330","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5c4d45cefd5f053ec6eb","name":"CRL","organization":"Ulm University","date":"2022-05-31","domains":["Biology"],"country":"Germany","reference":"https://arxiv.org/abs/2205.15675","confidence":"Confident","accessibility":""},{"id":"epoch:9fd6175cf43ed21a0643","name":"PFP","organization":"Preferred Networks Inc","date":"2022-05-30","domains":["Materials science"],"country":"Japan","reference":"https://www.nature.com/articles/s41467-022-30687-9","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:9443ebb361c485a88e2b","name":"CogVideo","organization":"Tsinghua University,Beijing Academy of Artificial Intelligence / BAAI","date":"2022-05-29","domains":["Video"],"country":"China,China","reference":"https://arxiv.org/abs/2205.15868","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:aca0e4863f30a42cdff0","name":"Tranception","organization":"University of Oxford,Harvard Medical School,Cohere","date":"2022-05-27","domains":["Biology"],"country":"United Kingdom,United States of America,Canada","reference":"https://arxiv.org/abs/2205.13760","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:455f19bbf6eeba9098e0","name":"GPT-2 Medium (FlashAttention)","organization":"Stanford University,University at Buffalo","date":"2022-05-27","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2205.14135","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6e60f949cb3221bbeff5","name":"TRIMELMext (247M)","organization":"Princeton University","date":"2022-05-25","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.12674","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:5712906bbaf999ecdfaa","name":"TRIMELMext (7M)","organization":"Princeton University","date":"2022-05-25","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.12674","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:391ff1f0e0910a82008b","name":"TRIMELMlong (150M)","organization":"Princeton University","date":"2022-05-25","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.12674","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:a604f87cfe5165970181","name":"Imagen","organization":"Google Brain","date":"2022-05-23","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.11487","confidence":"Likely","accessibility":"API access"},{"id":"epoch:bbeedf5ce69b2ae3d1fa","name":"improved U-Net for chest X-ray images segmentation","organization":"Henan University of Technology,Nanyang Central Hospital","date":"2022-05-23","domains":["Medicine","Vision"],"country":"China,China","reference":"https://www.nature.com/articles/s41598-022-12743-y","confidence":"Unknown","accessibility":""},{"id":"epoch:6a96a961c5c08f82d1f0","name":"LSTM+GraB","organization":"Cornell University","date":"2022-05-22","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.10733","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:62af22297f9cc02b6c62","name":"SimCSE","organization":"Princeton University,Tsinghua University","date":"2022-05-18","domains":["Language"],"country":"United States of America,China","reference":"https://arxiv.org/abs/2104.08821","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:bc3e63ceb2c4e755b188","name":"Gato","organization":"DeepMind","date":"2022-05-12","domains":["Multimodal","Robotics","Games","Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2205.06175","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:ee5e1bf9814e2f3bbf97","name":"UL2","organization":"Google Research,Google Brain","date":"2022-05-10","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2205.05131v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7df68d8ef5e5dd3aaa68","name":"ASE","organization":"Nvidia,University of California (UC) Berkeley","date":"2022-05-05","domains":["Robotics"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2205.01906","confidence":"Likely","accessibility":""},{"id":"epoch:04c42849544903e122e3","name":"StyleGAN-XL","organization":"Max Planck Institute for Intelligent Systems,University of T\u00fcbingen","date":"2022-05-05","domains":["Image generation"],"country":"Germany,Germany","reference":"https://arxiv.org/abs/2202.00273","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:249426959eae9bc3c55b","name":"DeBERTaV3large + KEAR","organization":"Microsoft","date":"2022-05-04","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2112.03254v3","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0c0b5e91383cf9f57910","name":"Jurassic-X","organization":"AI21 Labs","date":"2022-05-03","domains":["Language"],"country":"Israel","reference":"https://www.ai21.com/blog/jurassic-x-crossing-the-neuro-symbolic-chasm-with-the-mrkl-system","confidence":"","accessibility":"API access"},{"id":"epoch:2930a517af00d1acab2b","name":"OPT-175B","organization":"Meta AI","date":"2022-05-02","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2205.01068","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:6f2177b03db29ef2b9c7","name":"OPT-13B","organization":"Meta AI","date":"2022-05-02","domains":[],"country":"United States of America","reference":"https://arxiv.org/abs/2205.01068","confidence":"Likely","accessibility":""},{"id":"epoch:2e834940b727713e5259","name":"Flamingo","organization":"DeepMind","date":"2022-04-29","domains":["Multimodal","Vision","Language","Video"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2204.14198","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:22a73e38ce0900f66388","name":"CogView2","organization":"Tsinghua University,Beijing Academy of Artificial Intelligence / BAAI","date":"2022-04-28","domains":["Image generation"],"country":"China,China","reference":"https://arxiv.org/abs/2204.14217","confidence":"Confident","accessibility":""},{"id":"epoch:daa306d9ad5445f68ee9","name":"GraphBP","organization":"Texas A&M,Fujitsu","date":"2022-04-19","domains":["Biology"],"country":"United States of America,Japan","reference":"https://arxiv.org/abs/2204.09410","confidence":"Speculative","accessibility":""},{"id":"epoch:005cad56b62fa71d2efe","name":"LaMemo","organization":"Tsinghua University,NetEase,Oppo Mobile Telecommunications","date":"2022-04-15","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/abs/2204.07341","confidence":"","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b80051bcf677becfeb1d","name":"Sparse all-MLP","organization":"Meta AI","date":"2022-04-14","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2203.06850","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:987cfba88e2d84d589c7","name":"XMC-GAN","organization":"Google Research","date":"2022-04-14","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2101.04702","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:8f11ee0234bae2fbb839","name":"Stable Diffusion (LDM-KL-8-G)","organization":"Runway,Ludwig Maximilian University of Munich,Heidelberg University","date":"2022-04-13","domains":["Image generation"],"country":"United States of America,Germany,Germany","reference":"https://arxiv.org/abs/2112.10752","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:dddec47b69d8904eac84","name":"Anthropic LM 52B","organization":"Anthropic","date":"2022-04-12","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2204.05862","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2f105a2ec7a24f81aa41","name":"STT Conformer-Transducer XL","organization":"Nvidia","date":"2022-04-12","domains":["Speech"],"country":"United States of America","reference":"https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/stt_en_conformer_transducer_xlarge","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1f599ba7f4833bb9a7fc","name":"VLM-4","organization":"LightOn","date":"2022-04-12","domains":["Language"],"country":"France","reference":"https://web.archive.org/web/20220506095925/https://lighton.ai/blog/lighton-publicly-launches-muse-an-api-to-vlm-4-large-language-models-trained-natively-in-five-european-languages/","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:2ab0e32abfeeed92b5e0","name":"BERT-RBP","organization":"Waseda University","date":"2022-04-07","domains":["Biology"],"country":"Japan","reference":"https://academic.oup.com/bioinformaticsadvances/article/2/1/vbac023/6564689","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:fa4f8c70f3a596441ad1","name":"DALL\u00b7E 2","organization":"OpenAI","date":"2022-04-06","domains":["Image generation"],"country":"United States of America","reference":"https://cdn.openai.com/papers/dall-e-2.pdf","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:44d3df064dc368229df4","name":"PaLM (540B)","organization":"Google Research","date":"2022-04-04","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2204.02311","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:cee1b982275b6624c509","name":"Monarch-GPT-2-Medium","organization":"Stanford University,University at Buffalo,University of Michigan","date":"2022-04-01","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2204.00595","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0f68ec7229da0e3a7cb8","name":"Monarch-GPT-2-Small","organization":"Stanford University,University at Buffalo,University of Michigan","date":"2022-04-01","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2204.00595","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:167f9b008e46d9932287","name":"NoPos","organization":"Tel Aviv University,University of Washington,Intel Labs,Meta AI","date":"2022-03-30","domains":["Language"],"country":"Israel,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2203.16634","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e2e066fffe636ddf8ae0","name":"Chinchilla","organization":"DeepMind","date":"2022-03-29","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2203.15556","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:5252c7291a038325731e","name":"BaGuaLu","organization":"Tsinghua University,Zhejiang Lab,Beijing Academy of Artificial Intelligence / BAAI,Alibaba","date":"2022-03-28","domains":["Multimodal","Language","Vision"],"country":"China,China,China,China","reference":"https://dl.acm.org/doi/abs/10.1145/3503221.3508417","confidence":"","accessibility":""},{"id":"epoch:6ad06a972ade9738d44b","name":"GraSR","organization":"Shanghai Jiao Tong University,Ministry of Education of China","date":"2022-03-24","domains":["Biology"],"country":"China,China","reference":"https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1009986","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a31c72bc0122c09aed2c","name":"Make-A-Scene","organization":"Meta AI","date":"2022-03-24","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2203.13131","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:1eea970277a5941d7c68","name":"MemSizer (language modeling)","organization":"Meta AI,Chinese University of Hong Kong (CUHK)","date":"2022-03-23","domains":["Language"],"country":"United States of America,Hong Kong","reference":"https://arxiv.org/abs/2203.12644","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:cdd96070786b49b27917","name":"Segatron-XL large, M=384 + HCP","organization":"Microsoft Research,University of Waterloo","date":"2022-03-21","domains":["Language"],"country":"United States of America,Canada","reference":"https://arxiv.org/abs/2203.10692","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:065f162a533c15e5eb3e","name":"Transformer Large + HCP","organization":"University of Waterloo,Microsoft Research","date":"2022-03-21","domains":["Language"],"country":"Canada,United States of America","reference":"https://arxiv.org/abs/2203.10692","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:bf837f572e35e63ba6e1","name":"Segatron -XL base, M=150 + HCP","organization":"Microsoft Research,University of Waterloo","date":"2022-03-21","domains":["Language"],"country":"United States of America,Canada","reference":"https://arxiv.org/abs/2203.10692","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:d27a26a0a0b4d919a5d5","name":"GPT-3.5 (davinci-002)","organization":"OpenAI","date":"2022-03-15","domains":["Language"],"country":"United States of America","reference":"https://platform.openai.com/docs/models/gpt-3-5","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:cee9ff95fb281a458fcb","name":"ViT-G (model soup)","organization":"University of Washington,Columbia University,Google,Meta AI,Tel Aviv University","date":"2022-03-10","domains":["Vision"],"country":"United States of America,United States of America,United States of America,United States of America,Israel","reference":"https://arxiv.org/abs/2203.05482v3","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:69861de1b89254a54c8f","name":"GPT3-6.7B + muP","organization":"Microsoft,OpenAI","date":"2022-03-07","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2203.03466","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:fd671b629c0130d4eef4","name":"MegaSyn","organization":"Collaborations Pharmaceuticals","date":"2022-03-07","domains":["Medicine"],"country":"United States of America","reference":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9544280/","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:98d19503c689865e03cf","name":"RQ-Transformer (LSUN-cat dataset)","organization":"Kakao,POSTECH","date":"2022-03-03","domains":["Image generation"],"country":"South Korea,South Korea","reference":"https://arxiv.org/abs/2203.01941","confidence":"Likely","accessibility":""},{"id":"epoch:183d5ecfc64912f5e214","name":"RQ-Transformer (1.4B params ImageNet dataset)","organization":"Kakao,POSTECH","date":"2022-03-03","domains":["Vision","Image generation"],"country":"South Korea,South Korea","reference":"https://arxiv.org/abs/2203.01941","confidence":"Likely","accessibility":""},{"id":"epoch:2a351c15436bb988420e","name":"RQ-Transformer (3.8B params ImageNet dataset)","organization":"Kakao,POSTECH","date":"2022-03-03","domains":["Vision","Image generation"],"country":"South Korea,South Korea","reference":"https://arxiv.org/abs/2203.01941","confidence":"Likely","accessibility":""},{"id":"epoch:6381a806bc6ad87251bd","name":"Statement Curriculum Learning","organization":"OpenAI","date":"2022-03-02","domains":["Language","Mathematics"],"country":"United States of America","reference":"https://arxiv.org/abs/2202.01344","confidence":"Likely","accessibility":"API access"},{"id":"epoch:a301d4f0e43569ce4b3a","name":"DeepNet","organization":"Microsoft Research","date":"2022-03-01","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2203.00555","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c426b79b481bb7234308","name":"PolyCoder","organization":"Carnegie Mellon University (CMU)","date":"2022-02-26","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2202.13169","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:52dca517bccb5609e3b0","name":"FourCastNet","organization":"Nvidia,NERSC, Lawrence Berkeley National Laboratory,University of Michigan,Rice University,California Institute of Technology,Purdue University","date":"2022-02-22","domains":["Earth science"],"country":"United States of America,United States of America,United States of America,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2202.11214","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2b182e56fc8efcd84711","name":"ST-MoE","organization":"Google,Google Brain,Google Research","date":"2022-02-17","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2202.08906v2","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:4d2ce74b15df0a26ec6e","name":"Midjourney V1","organization":"Midjourney","date":"2022-02-15","domains":["Image generation"],"country":"United States of America","reference":"","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:17491ba6b850c190fc3f","name":"MuZero VP9","organization":"DeepMind","date":"2022-02-14","domains":["Video"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2202.06626","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:fa47ad087eed8383ecae","name":"LaMDA","organization":"Google","date":"2022-02-10","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2201.08239","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4024dee3f1dc8f18692b","name":"ProteinBERT","organization":"Hebrew University of Jerusalem,Ben-Gurion University of the Negev,Deep Trading","date":"2022-02-10","domains":["Biology"],"country":"Israel,Israel,United States of America","reference":"https://academic.oup.com/bioinformatics/article/38/8/2102/6502274","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3128319c95db6bbf0ba9","name":"GPT-NeoX-20B","organization":"EleutherAI","date":"2022-02-09","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2204.06745","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2b3c52e5c476d5ca06a0","name":"MaskGIT (ImageNet)","organization":"Google Research","date":"2022-02-08","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2202.04200","confidence":"Confident","accessibility":""},{"id":"epoch:b5e0e05c8766b3dec100","name":"RETRO-7B","organization":"DeepMind","date":"2022-02-07","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2112.04426","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:8931581341801d11fc40","name":"AlphaCode","organization":"DeepMind","date":"2022-02-02","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2203.07814","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:697d23c9b75c0de1e801","name":"DARK","organization":"University College London (UCL)","date":"2022-01-28","domains":["Biology"],"country":"United Kingdom","reference":"https://www.biorxiv.org/content/10.1101/2022.01.27.478087v1.full","confidence":"Speculative","accessibility":""},{"id":"epoch:97015afc08b3a5e2c7a2","name":"InstructGPT 175B","organization":"OpenAI","date":"2022-01-27","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/pdf/2203.02155","confidence":"Confident","accessibility":"API access"},{"id":"epoch:ac5f744ef181c8cf0f07","name":"InstructGPT 6B","organization":"OpenAI","date":"2022-01-27","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2203.02155","confidence":"Confident","accessibility":"API access"},{"id":"epoch:8b0fb16cf2c660ccd28f","name":"InstructGPT 1.3B","organization":"OpenAI","date":"2022-01-27","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2203.02155","confidence":"Confident","accessibility":"API access"},{"id":"epoch:dd5b38c94b4b17959252","name":"InstructGPT 350M","organization":"OpenAI","date":"2022-01-27","domains":["Language"],"country":"United States of America","reference":"","confidence":"Confident","accessibility":""},{"id":"epoch:be51ae923c756e0ab8c8","name":"Primer (GPT-3 XL-like 1.9B)","organization":"Google Brain","date":"2022-01-24","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2109.08668","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:dd163314859b4b95f098","name":"SPALM + RelationLM","organization":"DeepMind,University of Oxford","date":"2022-01-24","domains":["Language"],"country":"United Kingdom,United Kingdom","reference":"https://arxiv.org/abs/2201.09680","confidence":"","accessibility":"Unreleased"},{"id":"epoch:de77cfb8c78d4c6de26a","name":"Transformer-XL + RelationLM","organization":"DeepMind,University of Oxford","date":"2022-01-24","domains":["Language"],"country":"United Kingdom,United Kingdom","reference":"https://arxiv.org/abs/2201.09680","confidence":"","accessibility":"Unreleased"},{"id":"epoch:2888d08a807d8f7978e4","name":"OntoProtein","organization":"Zhejiang University (ZJU)","date":"2022-01-23","domains":["Biology","Language"],"country":"China","reference":"https://openreview.net/pdf?id=yfe1VMYAXa4","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7d236d34ea4114066666","name":"AbLang (heavy sequences)","organization":"University of Oxford","date":"2022-01-22","domains":["Biology"],"country":"United Kingdom","reference":"https://academic.oup.com/bioinformaticsadvances/article/2/1/vbac046/6609807","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a7f06fe4c67be675df86","name":"data2vec (language)","organization":"Meta AI","date":"2022-01-20","domains":["Language"],"country":"United States of America","reference":"https://ai.facebook.com/research/data2vec-a-general-framework-for-self-supervised-learning-in-speech-vision-and-language/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c9f7656505e1b5e22c54","name":"data2vec (speech)","organization":"Meta AI","date":"2022-01-20","domains":["Speech"],"country":"United States of America","reference":"https://ai.facebook.com/research/data2vec-a-general-framework-for-self-supervised-learning-in-speech-vision-and-language/","confidence":"","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4772c5a9ae66bbaf943c","name":"data2vec (vision)","organization":"Meta AI","date":"2022-01-20","domains":["Vision"],"country":"United States of America","reference":"https://ai.facebook.com/research/data2vec-a-general-framework-for-self-supervised-learning-in-speech-vision-and-language/","confidence":"","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0e527a48bb964b132bc6","name":"Japanese-GPT-1B","organization":"rinna","date":"2022-01-19","domains":["Language"],"country":"Japan","reference":"https://huggingface.co/rinna/japanese-gpt-1b","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6628a22fed5542419710","name":"Detic","organization":"Meta AI,University of Texas at Austin","date":"2022-01-07","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2201.02605","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5668f9bdbe8e177d05d8","name":"SignalP 6.0","organization":"Technical University of Denmark,ETH Zurich,University of Copenhagen,Stanford University,Stockholm University,European Bioinformatics Institute","date":"2022-01-03","domains":["Biology"],"country":"Denmark,Switzerland,Denmark,United States of America,Sweden,Multinational,United Kingdom","reference":"https://www.nature.com/articles/s41587-021-01156-3","confidence":"Unknown","accessibility":""},{"id":"epoch:a721b5ba289e95b23ef0","name":"ERNIE-ViLG","organization":"Baidu","date":"2021-12-31","domains":["Multimodal","Image generation","Vision","Language"],"country":"China","reference":"https://arxiv.org/abs/2112.15283","confidence":"","accessibility":"Unreleased"},{"id":"epoch:caedede42c67f75c5ba4","name":"Vespa","organization":"Technical University of Munich","date":"2021-12-30","domains":["Biology"],"country":"Germany","reference":"https://link.springer.com/article/10.1007/s00439-021-02411-y","confidence":"Confident","accessibility":""},{"id":"epoch:cfb87857dcbf935f11c0","name":"ERNIE 3.0 Titan","organization":"Baidu,Peng Cheng Laboratory","date":"2021-12-23","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2112.12731","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:56bd6acc890973b24fbc","name":"GLIDE","organization":"OpenAI","date":"2021-12-20","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2112.10741","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8e59bebd3ceda47fe92c","name":"MoE-1.1T","organization":"Meta AI","date":"2021-12-20","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2112.10684","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:0115f8e7ba3b95c17311","name":"Fairseq-dense 13B","organization":"Meta AI","date":"2021-12-20","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2112.10684","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2671f346833a68f5ada4","name":"LDM-1.45B","organization":"Heidelberg University,Runway","date":"2021-12-20","domains":["Image generation"],"country":"Germany,United States of America","reference":"https://arxiv.org/abs/2112.10752","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d793c0dd2f8c261f0ec6","name":"XGLM-7.5B","organization":"Meta AI,Facebook AI Research","date":"2021-12-20","domains":["Language"],"country":"United States of America,United States of America,France","reference":"https://arxiv.org/abs/2112.10668","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:beb182e5bc7ed7c68761","name":"XGLM","organization":"Meta AI","date":"2021-12-20","domains":[],"country":"United States of America","reference":"https://arxiv.org/abs/2112.10668","confidence":"Confident","accessibility":""},{"id":"epoch:320c089f77f836fb4445","name":"HSO","organization":"Toyota Technological Institute at Chicago","date":"2021-12-16","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2112.08653","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:6329641b16d5efaec129","name":"Contriever","organization":"Meta AI,University College London (UCL),PSL University,Universit\u00e9 Grenoble Alpes","date":"2021-12-16","domains":["Language"],"country":"United States of America,United Kingdom,France,France","reference":"https://arxiv.org/abs/2112.09118","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:140234ae5841ac7979c9","name":"LongT5","organization":"Google Research","date":"2021-12-15","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2112.07916","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2a92be43df61340268c1","name":"EXAONE 1.0","organization":"LG","date":"2021-12-14","domains":["Multimodal","Language","Vision"],"country":"South Korea","reference":"https://www.lgcorp.com/media/release/27387#:~:text=LG%20AI%20Research%20proposes%20EXAONE,performance%20while%20learning%20fewer%20parameters.","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:bd223dfb34de9762ac6e","name":"GLaM","organization":"Google","date":"2021-12-13","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2112.06905","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:8e450db30c34904a5793","name":"Engine-XL(NE)","organization":"Boston University","date":"2021-12-11","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://arxiv.org/abs/2112.05917","confidence":"","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6c9d23f0486e472dd520","name":"Engine-Base (NE)","organization":"Boston University","date":"2021-12-11","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://arxiv.org/abs/2112.05917","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1b0eea1678c996b8021e","name":"Engine-Medium(NE)","organization":"Boston University","date":"2021-12-11","domains":["Language","Vision","Multimodal"],"country":"United States of America","reference":"https://arxiv.org/abs/2112.05917","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:515c0929331389cdec2e","name":"Gopher (280B)","organization":"DeepMind","date":"2021-12-08","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2112.11446","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:85d1de1dc89b931508bf","name":"Gopher (7.1B)","organization":"DeepMind","date":"2021-12-08","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2112.11446","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:61c407934c94c1183789","name":"Student of Games","organization":"DeepMind","date":"2021-12-06","domains":["Games"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2112.03178","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:60885aa7a850a2676c57","name":"CTR-BERT","organization":"Amazon","date":"2021-12-06","domains":["Recommendation"],"country":"United States of America","reference":"https://neurips2021-nlp.github.io/papers/20/CameraReady/camera_ready_final.pdf","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:a7f43517b886b3de85c1","name":"T-NLRv5 XXL","organization":"Microsoft","date":"2021-12-03","domains":["Language"],"country":"United States of America","reference":"https://www.marketscreener.com/quote/stock/MICROSOFT-CORPORATION-4835/news/Microsoft-Turing-NLRv5-achieves-new-performance-milestones-37207301/\nhttps://www.microsoft.com/en-us/research/blog/efficiently-and-effectively-scaling-up-language-model-pretraining-for-best-language-representation-model-on-glue-and-superglue/","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:87f3ff5cb16d17592f58","name":"GPT-2-Medium+Pixelfly","organization":"Stanford University,SambaNova Systems, Inc,Peking University,Adobe,University at Buffalo","date":"2021-11-30","domains":["Language"],"country":"United States of America,United States of America,China,United States of America,United States of America","reference":"https://arxiv.org/abs/2112.00029","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c9f3aeaaa1a89eb198ba","name":"GPT-2-Small+Pixelfly","organization":"Stanford University,SambaNova Systems, Inc,Peking University,Adobe,University at Buffalo","date":"2021-11-30","domains":["Language"],"country":"United States of America,United States of America,China,United States of America,United States of America","reference":"https://arxiv.org/abs/2112.00029","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:3c3cab57cc5a0f642c5a","name":"Quantized ADMM","organization":"Chinese University of Hong Kong (CUHK),Microsoft","date":"2021-11-29","domains":["Language"],"country":"Hong Kong,United States of America","reference":"https://arxiv.org/abs/2111.14836","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:9ffb3bb07a530318d9b5","name":"Transformer LM + MinSen","organization":"Chinese University of Hong Kong (CUHK)","date":"2021-11-29","domains":["Language"],"country":"Hong Kong","reference":"https://arxiv.org/abs/2112.11540","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:85c1d937e040842a5313","name":"N\u00dcWA","organization":"Microsoft Research,Peking University","date":"2021-11-24","domains":["Multimodal","Vision","Image generation","Video","Language"],"country":"United States of America,China","reference":"https://arxiv.org/abs/2111.12417","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:d63b8b887bb4d8b349a4","name":"Persia","organization":"ETH Zurich,Kuaishou Technology","date":"2021-11-23","domains":["Recommendation"],"country":"Switzerland,China","reference":"https://arxiv.org/abs/2111.05897","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:11771256e13f96a2501e","name":"Florence","organization":"Microsoft","date":"2021-11-22","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2111.11432v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:ea50311c5d0059338afd","name":"BASIC-L","organization":"Google","date":"2021-11-19","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2111.10050","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:4566f07e9630d55539a8","name":"Swin Transformer V2 (SwinV2-G)","organization":"Microsoft Research Asia","date":"2021-11-18","domains":["Vision","Video"],"country":"China","reference":"https://arxiv.org/abs/2111.09883v2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:60fe5f005aa474313243","name":"DeBERTaV3large","organization":"Microsoft Research","date":"2021-11-18","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2111.09543","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fb754077fa66617af9b4","name":"ESM1v","organization":"Facebook AI Research,New York University (NYU),University of California (UC) Berkeley","date":"2021-11-17","domains":["Biology"],"country":"United States of America,France,United States of America,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2021.07.09.450648v2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:cdbd307027b86338cca0","name":"AliceMind-MMU","organization":"Alibaba","date":"2021-11-17","domains":[],"country":"China","reference":"https://arxiv.org/abs/2111.08896","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:3e2d274295b81b142837","name":"ViT-G/14 (LiT)","organization":"Google Research","date":"2021-11-15","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2111.07991v3","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:817f16f4590f3813e1e8","name":"EquiDock","organization":"Massachusetts Institute of Technology (MIT),ETH Zurich,Tencent","date":"2021-11-15","domains":["Biology"],"country":"United States of America,Switzerland,China","reference":"https://arxiv.org/abs/2111.07786","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2a8aa1e72d21e2af5f83","name":"A.X (Adot) 18B","organization":"SK Telecom","date":"2021-11-15","domains":["Language","Speech"],"country":"South Korea","reference":"https://aclanthology.org/2023.acl-industry.68.pdf","confidence":"Speculative","accessibility":""},{"id":"epoch:7c23b0d05c4eabe594e0","name":"KoGPT","organization":"Kakao","date":"2021-11-12","domains":["Language"],"country":"South Korea","reference":"https://github.com/kakaobrain/kogpt","confidence":"Speculative","accessibility":"Open weights (non-commercial)"},{"id":"epoch:19cc7b29825237669f4c","name":"Masked Autoencoders ViT-H","organization":"Facebook AI Research","date":"2021-11-11","domains":["Vision"],"country":"United States of America,France","reference":"https://arxiv.org/abs/2111.06377","confidence":"Speculative","accessibility":"Open weights (non-commercial)"},{"id":"epoch:2cfda97128a73ab49f98","name":"Japanese dialog transformers","organization":"NTT Communication Science Laboratories","date":"2021-11-09","domains":["Language"],"country":"Japan","reference":"https://arxiv.org/abs/2109.05217","confidence":"","accessibility":"Open weights (non-commercial)"},{"id":"epoch:2eb6b4d1f634e227ddec","name":"GPT-2 (AMPS)","organization":"University of California (UC) Berkeley","date":"2021-11-08","domains":["Mathematics","Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2103.03874","confidence":"Speculative","accessibility":""},{"id":"epoch:aee74d5245c133b106ed","name":"GPT2+CoreLM+Fine-Tuning","organization":"Aristotle University of Thessaloniki","date":"2021-11-04","domains":["Language"],"country":"Greece","reference":"https://arxiv.org/abs/2111.02687","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:23bc3dfb42f6b62c46f1","name":"NCP-VAE (CIFAR 10)","organization":"University of Illinois Urbana-Champaign (UIUC),Nvidia","date":"2021-11-03","domains":["Image generation"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2010.02917","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:2a32d6852fa0ac50be09","name":"NCP-VAE (Celeba  HQ)","organization":"University of Illinois Urbana-Champaign (UIUC),Nvidia","date":"2021-11-03","domains":["Image generation"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2010.02917","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:d0b60d028c55ed1ae3c2","name":"CodeT5-base","organization":"Salesforce,Nanyang Technological University","date":"2021-11-01","domains":["Language"],"country":"United States of America,Singapore","reference":"https://aclanthology.org/2021.emnlp-main.685/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:89d4de2a1f92c5755865","name":"Projected GAN","organization":"Heidelberg University","date":"2021-11-01","domains":["Image generation"],"country":"Germany","reference":"https://proceedings.neurips.cc/paper/2021/hash/9219adc5c42107c4911e249155320648-Abstract.html","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:64b54c8117efcba0bbc9","name":"S4","organization":"Stanford University","date":"2021-10-31","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2111.00396","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ea8a236e8205fb8b7cb7","name":"EfficientZero","organization":"Tsinghua University,University of California (UC) Berkeley,Shanghai Qi Zhi institute","date":"2021-10-30","domains":["Games"],"country":"China,United States of America,China","reference":"https://arxiv.org/abs/2111.00210","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:7225f7e3d22c7d9c5b68","name":"Scatterbrain","organization":"Stanford University,Adobe,University at Buffalo","date":"2021-10-28","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2110.15343","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:7da457eb136a649ef501","name":"Eve","organization":"Harvard Medical School,University of Oxford","date":"2021-10-27","domains":["Biology"],"country":"United States of America,United Kingdom","reference":"https://www.nature.com/articles/s41586-021-04043-8#change-history","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:8abed99d01857ad51632","name":"DALL-E mini","organization":"Craiyon","date":"2021-10-26","domains":["Vision"],"country":"United States of America","reference":"https://huggingface.co/dalle-mini/dalle-mini","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a60e5d1a9c5e8bb38417","name":"Cloob","organization":"Johannes Kepler University,HERE Technologies,Institute of Advanced Research in Artificial Intelligence","date":"2021-10-21","domains":["Multimodal","Language","Vision"],"country":"Austria,Switzerland,Austria","reference":"https://arxiv.org/abs/2110.11316","confidence":"","accessibility":""},{"id":"epoch:d5ea441072877bb1b78f","name":"PMLM-large","organization":"Microsoft Research Asia,Nanyang Technological University,Xi\u2019an Jiaotong University,Sun Yat-sen University","date":"2021-10-21","domains":["Biology"],"country":"China,Singapore,China,China","reference":"https://arxiv.org/abs/2110.15527","confidence":"Confident","accessibility":""},{"id":"epoch:bffa056c0e1d4bb106c8","name":"WD+LR+M","organization":"University of Cambridge,Alan Turing Institute","date":"2021-10-20","domains":["Language"],"country":"United Kingdom,United Kingdom","reference":"https://arxiv.org/abs/2110.10461","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:c01a1f719e7bbaa0d627","name":"base LM+GNN+kNN","organization":"Shannon.AI,Nanjing University,Nanyang Technological University,Zhejiang University (ZJU)","date":"2021-10-17","domains":["Language"],"country":"China,China,Singapore,China","reference":"https://arxiv.org/abs/2110.08743","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c033d251743851866a4f","name":"base LM+GNN (WT103)","organization":"Shannon.AI,Nanjing University,Nanyang Technological University,Zhejiang University (ZJU)","date":"2021-10-17","domains":["Language"],"country":"China,China,Singapore,China","reference":"https://arxiv.org/abs/2110.08743","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:d8a64d3084399cd08381","name":"PAGnol-XL","organization":"LightOn,Laboratoire de Physique de l'Ecole Normale (LPENS),INRIA","date":"2021-10-16","domains":["Language"],"country":"France,France,France","reference":"https://arxiv.org/abs/2110.08554","confidence":"Confident","accessibility":"API access"},{"id":"epoch:a5a0810561bf849fdd31","name":"GPT-2 (fine-tuned with HYDRA)","organization":"University of California San Diego","date":"2021-10-16","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2110.08633","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c73a3c075ce826961c52","name":"MGK 4 heads (medium)","organization":"FPT Software AI Center,University of California Los Angeles (UCLA),VinUniversity,Deezer Research,Rice University,University of Texas at Austin","date":"2021-10-16","domains":["Language"],"country":"Vietnam,United States of America,Vietnam,France,United States of America,United States of America","reference":"https://arxiv.org/abs/2110.08678","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:d0f35dc7a808154c46c4","name":"MGK 8 heads (small)","organization":"FPT Software AI Center,University of California Los Angeles (UCLA),VinUniversity,Deezer Research,Rice University,University of Texas at Austin","date":"2021-10-16","domains":["Language"],"country":"Vietnam,United States of America,Vietnam,France,United States of America,United States of America","reference":"https://arxiv.org/abs/2110.08678","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:82f553cd1bf7b788c369","name":"T0-XXL","organization":"Hugging Face,Brown University","date":"2021-10-15","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2110.08207","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3c57dfc352259cdd2009","name":"KnGPT2","organization":"Huawei Noah's Ark Lab,McGill University","date":"2021-10-15","domains":["Language"],"country":"China,Canada","reference":"https://arxiv.org/abs/2110.08152","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:6caacf9af13cbdb4e060","name":"Yuan 1.0","organization":"Inspur","date":"2021-10-12","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2110.04725","confidence":"Confident","accessibility":"API access"},{"id":"epoch:75f189aeb9d468cef418","name":"TOME","organization":"University of Southern California,Google","date":"2021-10-12","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2110.06176v2","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:b2214e00967e177f9eae","name":"Megatron-Turing NLG 530B","organization":"Microsoft,Nvidia","date":"2021-10-11","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2201.11990","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f0c16afd16b27e65c967","name":"M6-10T","organization":"Alibaba","date":"2021-10-08","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://arxiv.org/abs/2110.03888","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:addc8c68109e41e4938c","name":"AlphaFold-Multimer","organization":"Google DeepMind,DeepMind","date":"2021-10-04","domains":["Biology"],"country":"United States of America,United Kingdom","reference":"https://www.biorxiv.org/content/10.1101/2021.10.04.463034v1","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4f8d3833d653a6244ed1","name":"Turing ULRv5","organization":"Microsoft","date":"2021-09-28","domains":["Language"],"country":"United States of America","reference":"https://www.microsoft.com/en-us/research/blog/microsoft-turing-universal-language-representation-model-t-ulrv5-tops-xtreme-leaderboard-and-trains-100x-faster/","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:ca296bb0707445511a76","name":"TrOCR","organization":"Beihang University,Microsoft Research Asia","date":"2021-09-21","domains":["Vision"],"country":"China,China","reference":"https://arxiv.org/abs/2109.10282","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:68b0c1abfbc53bfb6f5a","name":"LM-GVP","organization":"Amazon Machine Learning Solutions Lab,Johnson & Johnson","date":"2021-09-21","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2021.09.21.460852v1.abstract","confidence":"Unknown","accessibility":"Open weights (non-commercial)"},{"id":"epoch:8db4ad012a486ed01d2a","name":"PLATO-XL","organization":"Baidu","date":"2021-09-20","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2109.09519","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:851525f91a090b541ac0","name":"DLRM-2022","organization":"Facebook","date":"2021-09-15","domains":["Recommendation"],"country":"United States of America","reference":"https://arxiv.org/abs/2104.05158","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:8354e838bd1dd537bdcf","name":"MegaMolBART","organization":"Nvidia","date":"2021-09-14","domains":["Biology"],"country":"United States of America","reference":"https://docs.nvidia.com/bionemo-framework/0.4.0/models/megamolbart.html, https://github.com/NVIDIA/MegaMolBART","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:84d9d55c65d28b197888","name":"HyperCLOVA 82B","organization":"NAVER,Search Solutions","date":"2021-09-10","domains":["Language"],"country":"South Korea,South Korea","reference":"https://arxiv.org/abs/2109.04650","confidence":"Confident","accessibility":"API access"},{"id":"epoch:e76fa04e66c09d5dd0a3","name":"HyperCLOVA 204B","organization":"NAVER","date":"2021-09-10","domains":["Language"],"country":"South Korea","reference":"https://aibusiness.com/nlp/south-korea-s-naver-unveils-hyperscale-ai-platform-language-model-with-more-parameters-than-gpt-3","confidence":"Speculative","accessibility":"Hosted access (no API)"},{"id":"epoch:1a3d8704df88a6902c38","name":"NLM","organization":"Carnegie Mellon University (CMU),University of California San Diego","date":"2021-09-09","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2109.04212","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:6b2e27078dd1f379e487","name":"Speechmatics Enhanced","organization":"Speechmatics","date":"2021-09-07","domains":["Speech"],"country":"United Kingdom","reference":"https://docs-beta.internal.speechmatics.com/release-notes/batch-saas-release-notes#2022-03-16","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:66d22337766b6a0ab71e","name":"PermuteFormer","organization":"Peking University","date":"2021-09-06","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2109.02377","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:a2fceb709dcdf656bc96","name":"RNS-RNN","organization":"University of Notre Dame","date":"2021-09-05","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2109.01982","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:27f4a5ad2c65d99b49e6","name":"MEB","organization":"Microsoft","date":"2021-09-04","domains":["Search","Language"],"country":"United States of America","reference":"https://www.microsoft.com/en-us/research/blog/make-every-feature-binary-a-135b-parameter-sparse-neural-network-for-massively-improved-search-relevance/","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:ccdcb314c386c3128618","name":"FLAN 137B","organization":"Google Research","date":"2021-09-03","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2109.01652","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:a5b70bdb63125042d532","name":"PLUS-RNN","organization":"Seoul National University,LG AI Research,NAVER,Kangwon National University","date":"2021-09-03","domains":["Biology"],"country":"South Korea,South Korea,South Korea,South Korea","reference":"https://ieeexplore.ieee.org/abstract/document/9529198","confidence":"Confident","accessibility":""},{"id":"epoch:a5eb49dcc0654a5c4447","name":"$\\infty$-former (SM)","organization":"Universidade de Lisboa (ULisboa),DeepMind","date":"2021-09-01","domains":["Language"],"country":"Portugal,United Kingdom","reference":"https://arxiv.org/abs/2109.00301","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:2be7e2bbebe4e75eca7c","name":"HJRSS","organization":"University of Washington,Microsoft","date":"2021-09-01","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2021.09.01.458592v1.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:f9a70bb0e0604ee2e868","name":"ALiBi (L=3072, Lvalid = 3072)","organization":"University of Washington,Facebook AI Research,Allen Institute for AI","date":"2021-08-27","domains":["Language"],"country":"United States of America,United States of America,France,United States of America","reference":"https://arxiv.org/abs/2108.12409","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0b95ee432a8bb4684f8b","name":"Speechmatics Standard","organization":"Speechmatics","date":"2021-08-23","domains":["Speech"],"country":"United Kingdom","reference":"https://docs-beta.internal.speechmatics.com/release-notes/batch-saas-release-notes#2022-03-16","confidence":"Unknown","accessibility":"API access"},{"id":"epoch:743ab3cdb2ee7d5c54ce","name":"XLMR-XXL","organization":"Facebook AI Research","date":"2021-08-17","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/2105.00572","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:dfbcd134d88f116da911","name":"ProteinLM","organization":"Tsinghua University,Beijing Academy of Artificial Intelligence / BAAI,Tencent","date":"2021-08-17","domains":["Biology"],"country":"China,China,China","reference":"https://arxiv.org/abs/2108.07435c","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b4683206ac90f1484cb9","name":"DNABERT","organization":"Northeastern University","date":"2021-08-15","domains":["Biology"],"country":"United States of America","reference":"https://academic.oup.com/bioinformatics/article/37/15/2112/6128680","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a9e98b70a6b89103035e","name":"GPT-2 (1.5B, Curriculum Learning 45K)","organization":"Microsoft","date":"2021-08-13","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2108.06084","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:4d022a4e5ac97cc67688","name":"GPT-2 (117M, SLW 110K)","organization":"Microsoft","date":"2021-08-13","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2108.06084","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:56e982341ba13a392306","name":"Jurassic-1-Jumbo","organization":"AI21 Labs","date":"2021-08-11","domains":["Language"],"country":"Israel","reference":"https://uploads-ssl.webflow.com/60fd4503684b466578c0d307/61138924626a6981ee09caf6_jurassic_tech_paper.pdf","confidence":"Confident","accessibility":"API access"},{"id":"epoch:f32494dccafb2050bffa","name":"Zidong Taichu","organization":"Chinese Academy of Sciences,Wuhan AI Computing Center","date":"2021-08-11","domains":["Multimodal","Speech","Vision","Language"],"country":"China,China","reference":"https://gitee.com/zidongtaichu/multi-modal-models","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0d5f2a39c435c79a8e9c","name":"W2v-BERT","organization":"Google Brain,Massachusetts Institute of Technology (MIT)","date":"2021-08-07","domains":["Speech"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2108.06209v2","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:98c4bdc81e92e61b2a04","name":"YOLOX-X","organization":"Megvii Inc","date":"2021-08-06","domains":["Vision"],"country":"China","reference":"https://arxiv.org/abs/2107.08430","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0834d70926a110c977fa","name":"FMMformer (2-kernel fast weight + Band20)","organization":"University of California Los Angeles (UCLA),University of Utah","date":"2021-08-05","domains":["Language"],"country":"United States of America,United States of America","reference":"https://proceedings.neurips.cc/paper/2021/file/f621585df244e9596dc70a39b579efb1-Paper.pdf","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:44257df680180ad7b835","name":"6-Act Tether","organization":"Facebook AI Research,Georgia Institute of Technology","date":"2021-08-03","domains":["Robotics"],"country":"United States of America,France,United States of America","reference":"https://openaccess.thecvf.com/content/ICCV2021/html/Ye_Auxiliary_Tasks_and_Exploration_Enable_ObjectGoal_Navigation_ICCV_2021_paper.html\nhttps://arxiv.org/abs/2104.04112","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:512f70100fb382c241a1","name":"SEER","organization":"Facebook AI Research,INRIA","date":"2021-07-29","domains":["Vision"],"country":"United States of America,France,France","reference":"https://arxiv.org/abs/2103.01988","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:ec03c5b5f407a9e69471","name":"GOAT","organization":"DeepMind","date":"2021-07-27","domains":["Games"],"country":"United Kingdom","reference":"https://deepmind.com/blog/article/generally-capable-agents-emerge-from-open-ended-play\n\nhttps://arxiv.org/abs/2107.12808","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:6e55f27849b551578a1a","name":"HuBERT","organization":"Facebook AI Research","date":"2021-07-27","domains":["Speech"],"country":"United States of America,France","reference":"https://arxiv.org/abs/2106.07447","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2d856e6e6841d3ec7319","name":"FNetAR Medium","organization":"X-Mechanics","date":"2021-07-22","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2107.10932","confidence":"","accessibility":"Unreleased"},{"id":"epoch:6290efc9bd438bd48718","name":"Codex","organization":"OpenAI","date":"2021-07-07","domains":["Language"],"country":"United States of America","reference":"https://openai.com/blog/openai-codex/\nhttps://arxiv.org/abs/2107.03374","confidence":"Likely","accessibility":"API access"},{"id":"epoch:d84bda0c90183e2afb86","name":"ERNIE 3.0","organization":"Baidu","date":"2021-07-05","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2107.02137","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a5619895277a09af5b75","name":"GemNet-T (OC20)","organization":"Technical University of Munich","date":"2021-07-01","domains":["Materials science"],"country":"Germany","reference":"https://arxiv.org/abs/2106.08903","confidence":"Likely","accessibility":"Open weights (restricted use)"},{"id":"epoch:7ade6ffc79fd1e29c47c","name":"DEQ-Transformer (Post-LN) + Jacobian Regularisation","organization":"Carnegie Mellon University (CMU),Intel Labs","date":"2021-06-28","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2106.14342","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:37481dae21574ea44232","name":"Adaptive Input Transformer + RD","organization":"Microsoft Research Asia,Soochow University","date":"2021-06-28","domains":["Language"],"country":"China,Taiwan","reference":"https://arxiv.org/abs/2106.14448","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:5ea191bb5477a0d5189a","name":"CPM-2","organization":"Tsinghua University,Beijing Academy of Artificial Intelligence / BAAI","date":"2021-06-24","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2106.10715","confidence":"Likely","accessibility":""},{"id":"epoch:90a2bdd34335acc87ee8","name":"Fold2Seq","organization":"IBM,Texas A&M","date":"2021-06-24","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2106.13058","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:ba9ad11cbfd3c19ad8cb","name":"EfficientNetV2-XL","organization":"Google,Google Brain","date":"2021-06-23","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2104.00298","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:64fcb33f785c97ca3cec","name":"StyleGAN3-T","organization":"Nvidia,Aalto University","date":"2021-06-21","domains":["Image generation"],"country":"United States of America,Finland","reference":"https://arxiv.org/abs/2106.12423","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:a57665a7177289132dc8","name":"StyleGAN3-R","organization":"Nvidia,Aalto University","date":"2021-06-21","domains":["Image generation"],"country":"United States of America,Finland","reference":"https://arxiv.org/pdf/2106.12423","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:16402d0b97d2a93b19b3","name":"ALIGN","organization":"Google Research","date":"2021-06-11","domains":["Multimodal","Vision","Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2102.05918","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:9fda6ae97e965c8bd26d","name":"Denoising Diffusion Probabilistic Models (LSUN Bedroom)","organization":"University of California (UC) Berkeley","date":"2021-06-11","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2006.11239","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:168b6ed88944f850d26a","name":"Delta RNN (+ full context)","organization":"IDSIA,SUPSI,King Abdullah University of Science and Technology (KAUST)","date":"2021-06-11","domains":["Language"],"country":"Switzerland,Switzerland,Saudi Arabia","reference":"https://proceedings.neurips.cc/paper/2021/file/3f9e3767ef3b10a0de4c256d7ef9805d-Paper.pdf","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:f9bb1dfc5767d9125a62","name":"DeBERTa","organization":"Microsoft","date":"2021-06-10","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2006.03654","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d9f825732dd4a7990627","name":"CoAtNet","organization":"Google,Google Research,Google Brain","date":"2021-06-09","domains":["Vision"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2106.04803v2","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:9f10e2317b9f372f94d7","name":"EMDR","organization":"Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),McGill University,DeepMind","date":"2021-06-09","domains":["Language"],"country":"Canada,Canada,United Kingdom","reference":"https://arxiv.org/abs/2106.05346v2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d1734bebc72c8059dace","name":"ViT-G/14","organization":"Google Brain,Google Research","date":"2021-06-08","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2106.04560","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:78161b493c3e73e44c37","name":"AFP+FPI (PTB)","organization":"University of Sheffield","date":"2021-06-04","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2106.02417","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:827f28caa8f5cd963bc5","name":"AFP+FPI (WT2)","organization":"University of Sheffield","date":"2021-06-04","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2106.02417","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:98c51579a43d72ce66d4","name":"GPT2-Large+LHOPT","organization":"OpenAI","date":"2021-06-02","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2106.00958","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:76d8f8a5bf6ae0cf9190","name":"Wu Dao 2.0","organization":"Beijing Academy of Artificial Intelligence / BAAI","date":"2021-05-31","domains":["Multimodal","Language","Vision","Image generation"],"country":"China","reference":"https://www.engadget.com/chinas-gigantic-multi-modal-ai-is-no-one-trick-pony-211414388.html","confidence":"Speculative","accessibility":"API access"},{"id":"epoch:71afb06857f9ae332d05","name":"CODA","organization":"The University of Hong Kong,Sun Yat-sen University,Shanghai AI Lab","date":"2021-05-31","domains":["Language"],"country":"Hong Kong,China,China","reference":"https://arxiv.org/abs/2105.14850","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e88efb28c847f8e58298","name":"ByT5-XXL","organization":"Google,Google Research","date":"2021-05-28","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2105.13626","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:cb298a0db2af0564eb05","name":"Transformer local-attention (NesT-B)","organization":"Google Cloud,Google Research","date":"2021-05-26","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2105.12723v4","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7cb2381e889efb8ac85b","name":"CogView","organization":"Tsinghua University,Alibaba DAMO Academy","date":"2021-05-26","domains":["Image generation"],"country":"China,China","reference":"https://arxiv.org/abs/2105.13290","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1c279ec7e218c79b73da","name":"DeepFRI","organization":"Flatiron Institute,University of California San Diego,Jagiellonian University,New York University (NYU),Broad Institute,University of Auckland,Massachusettes General Hospital,Harvard Medical School,Massachusetts Institute of Technology (MIT)","date":"2021-05-26","domains":["Biology"],"country":"United States of America,United States of America,Poland,United States of America,United States of America,New Zealand,United States of America,United States of America,United States of America","reference":"https://www.nature.com/articles/s41467-021-23303-9","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0b910aba578a7140abf7","name":"ConSERT","organization":"Meituan University,Beijing University of Posts and Telecommunications","date":"2021-05-25","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2105.11741","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4c4a986bac6dd1196a1d","name":"MedBERT","organization":"Peng Cheng Laboratory,University of Texas at Houston","date":"2021-05-20","domains":["Medicine"],"country":"China,United States of America","reference":"https://www.nature.com/articles/s41746-021-00455-y","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:7927f906ac9b95c7db3b","name":"Multitask Unified Model (MUM)","organization":"Google","date":"2021-05-18","domains":["Language","Multimodal","Vision"],"country":"United States of America","reference":"https://blog.google/products/search/introducing-mum/","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:44351cdb70b2340314f9","name":"Fairseq + UID: variance","organization":"Google AI,ETH Zurich,University of Cambridge","date":"2021-05-15","domains":["Language"],"country":"United States of America,Switzerland,United Kingdom","reference":"https://arxiv.org/abs/2105.07144","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:5b6d204521b2c63747db","name":"ADM","organization":"OpenAI","date":"2021-05-11","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2105.05233","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:fbbcc0907359ae7f4333","name":"ProtT5-XXL","organization":"Technical University of Munich,Med AI Technology,Nvidia,Oak Ridge National Laboratory,Google,Seoul National University","date":"2021-05-04","domains":["Biology"],"country":"Germany,China,United States of America,United States of America,United States of America,South Korea","reference":"https://www.biorxiv.org/content/10.1101/2020.07.12.199554v3 or \nhttps://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9477085","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3d96e66b848c0bca3847","name":"ProtT5-XXL-BFD","organization":"Technical University of Munich,Med AI Technology,Nvidia,Oak Ridge National Laboratory,Google,Seoul National University","date":"2021-05-04","domains":["Biology"],"country":"Germany,China,United States of America,United States of America,United States of America,South Korea","reference":"https://www.biorxiv.org/content/10.1101/2020.07.12.199554v3 or \nhttps://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9477085","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:158a09c0e3cc91a59677","name":"ProtBERT-BFD","organization":"Technical University of Munich,Nvidia,Seoul National University,Google,Oak Ridge National Laboratory,Med AI Technology","date":"2021-05-04","domains":["Biology"],"country":"Germany,United States of America,South Korea,United States of America,United States of America,China","reference":"https://www.biorxiv.org/content/10.1101/2020.07.12.199554v3 or \nhttps://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9477085","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e8c089371e8b79eb4d12","name":"ProtBERT-UniRef","organization":"Technical University of Munich,Nvidia,Seoul National University,Google,Oak Ridge National Laboratory,Med AI Technology","date":"2021-05-04","domains":["Biology"],"country":"Germany,United States of America,South Korea,United States of America,United States of America,China","reference":"https://www.biorxiv.org/content/10.1101/2020.07.12.199554v3 or \nhttps://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9477085","confidence":"Confident","accessibility":""},{"id":"epoch:6632ad7998464f9b2076","name":"ProtT5-XL-U50","organization":"Technical University of Munich,Med AI Technology,Nvidia,Oak Ridge National Laboratory,Google,Seoul National University","date":"2021-05-04","domains":["Biology"],"country":"Germany,China,United States of America,United States of America,United States of America,South Korea","reference":"https://www.biorxiv.org/content/10.1101/2020.07.12.199554v3.full.pdf","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:bbdc91309f932ea318e2","name":"Transformer-XL + SIS","organization":"INRIA","date":"2021-05-03","domains":["Language"],"country":"France","reference":"http://proceedings.mlr.press/v139/verma21b/verma21b.pdf","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:d2023742a2781a500efa","name":"GPT-J-6B","organization":"EleutherAI,LAION","date":"2021-05-01","domains":["Language"],"country":"United States of America,Germany","reference":"https://arankomatsuzaki.wordpress.com/2021/06/04/gpt-j/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8b6f439d861352bae3b0","name":"ViT + DINO","organization":"INRIA,Facebook AI Research","date":"2021-04-29","domains":["Vision"],"country":"France,United States of America,France","reference":"https://arxiv.org/abs/2104.14294","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:11592f15ddc41ae9e970","name":"SPALM + kNN","organization":"DeepMind","date":"2021-04-26","domains":["Language"],"country":"United Kingdom","reference":"https://direct.mit.edu/tacl/article/doi/10.1162/tacl_a_00371/100688/Adaptive-Semiparametric-Language-Models","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:62a9505c2474b5a42438","name":"PanGu-\u03b1","organization":"Huawei Noah's Ark Lab","date":"2021-04-25","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2104.12369","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:3e10d0e4d59f9d7fb950","name":"DiffQ Transformer (16L)","organization":"Meta AI","date":"2021-04-20","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2104.09987","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c31eed6cf14a2e05589f","name":"PLUG","organization":"Alibaba","date":"2021-04-19","domains":["Language"],"country":"China","reference":"https://mp.weixin.qq.com/s/DAQomIkDa52Sef-ruyH5qg","confidence":"Confident","accessibility":"Hosted access (no API)"},{"id":"epoch:b8c49813cc63c688860b","name":"DLRM-12T","organization":"Meta AI,Carnegie Mellon University (CMU)","date":"2021-04-12","domains":["Recommendation"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2104.05158","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:cc1f1a2edb527bd66b4c","name":"Megatron-LM (1T)","organization":"Microsoft Research,Nvidia,Stanford University","date":"2021-04-09","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2104.04473","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:27a000f39674c0c967e6","name":"Transformer-C","organization":"University of Massachusetts Amherst","date":"2021-04-08","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2104.03474","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4848f958e4fc4f08de0d","name":"TransfoRNN(d=1024)(2-layer) (WT2)","organization":"Lenovo Research","date":"2021-04-04","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2104.01572","confidence":"","accessibility":"Unreleased"},{"id":"epoch:b1940c8ca389006dd27b","name":"TransfoRNN(d=1024)(2-layer) (PTB)","organization":"Lenovo Research","date":"2021-04-04","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2104.01572","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b967f3b909ab81fa4daf","name":"GraphMS","organization":"Dalian University of Technology,Dongbei University of Technology,Baidu,China National Health Development Research Center","date":"2021-04-04","domains":["Biology"],"country":"China,China,China,China","reference":"https://www.mdpi.com/2076-3417/11/7/3239","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:1e3a37b017c1b03fbb6c","name":"T2R + Random Init","organization":"University of Washington,Microsoft,DeepMind,Allen Institute for AI","date":"2021-03-24","domains":["Language"],"country":"United States of America,United States of America,United Kingdom,United States of America","reference":"https://arxiv.org/abs/2103.13076","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:4dc2183deefc49bdbf9f","name":"T2R 75% + Pretrain (WT-103)","organization":"University of Washington,Microsoft,DeepMind","date":"2021-03-24","domains":["Language"],"country":"United States of America,United States of America,United Kingdom","reference":"https://arxiv.org/abs/2103.13076","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:a796f1c423a78f9a6ea5","name":"T2R + Pretrain","organization":"University of Washington,Microsoft,DeepMind","date":"2021-03-24","domains":["Language"],"country":"United States of America,United States of America,United Kingdom","reference":"https://arxiv.org/abs/2103.13076","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2d3f400aaaf56e2110a6","name":"Unicorn","organization":"Allen Institute for AI","date":"2021-03-24","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2103.13009","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2a52c455267d043a0d99","name":"GPT-Neo-2.7B","organization":"EleutherAI","date":"2021-03-21","domains":["Language"],"country":"United States of America","reference":"https://github.com/EleutherAI/gpt-neo","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f3dc3c61ffa026cf6ffa","name":"GPT-Neo-2.7B (finetuned)","organization":"EleutherAI","date":"2021-03-21","domains":["Language"],"country":"United States of America","reference":"https://github.com/EleutherAI/gpt-neo","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:bed48b025311241e5713","name":"GPT-Neo-2.7B (finetuned on PTB)","organization":"EleutherAI","date":"2021-03-21","domains":["Language"],"country":"United States of America","reference":"https://github.com/EleutherAI/gpt-neo","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:46a96ce681cc2da45790","name":"GPT-Neo-125M","organization":"EleutherAI","date":"2021-03-21","domains":["Language"],"country":"United States of America","reference":"https://github.com/EleutherAI/gpt-neo","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:658f7eab29e56ebb6c6a","name":"GPT-Neo-1.3B","organization":"EleutherAI","date":"2021-03-21","domains":["Language"],"country":"United States of America","reference":"https://github.com/EleutherAI/gpt-neo","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5329d4bc65c7bdd308d8","name":"GPT-Neo-125M (finetuned)","organization":"EleutherAI","date":"2021-03-21","domains":["Language"],"country":"United States of America","reference":"https://github.com/EleutherAI/gpt-neo, https://arxiv.org/pdf/2302.03773.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:81ff4716787db4b7a2f1","name":"GPT-Neo-1.3B (finetuned)","organization":"EleutherAI","date":"2021-03-21","domains":["Language"],"country":"United States of America","reference":"https://github.com/EleutherAI/gpt-neo","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e9a4458b38daf12af0d3","name":"U-Net GAN (FFHQ)","organization":"Bosch Center for Artificial Intelligence,Max Planck Institute for Informatics","date":"2021-03-19","domains":["Image generation"],"country":"Germany,Germany","reference":"https://arxiv.org/abs/2002.12655","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:95e66bc08aa56a4c22a4","name":"GLM-10B","organization":"Tsinghua University,Beijing Academy of Artificial Intelligence / BAAI,Massachusetts Institute of Technology (MIT),Shanghai Qi Zhi institute","date":"2021-03-18","domains":["Language"],"country":"China,China,United States of America,China","reference":"https://arxiv.org/abs/2103.10360","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:13c40d0997fcc2ab4ffb","name":"GLM-2B","organization":"Tsinghua University,Beijing Academy of Artificial Intelligence / BAAI,Massachusetts Institute of Technology (MIT)","date":"2021-03-18","domains":["Language"],"country":"China,China,United States of America","reference":"https://arxiv.org/abs/2103.10360","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e58acb5b84161ccbd298","name":"GLM-10B-bidirectional","organization":"Tsinghua University,Beijing Academy of Artificial Intelligence / BAAI,Massachusetts Institute of Technology (MIT)","date":"2021-03-18","domains":["Language"],"country":"China,China,United States of America","reference":"https://arxiv.org/abs/2103.10360","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1855b46a88aad5a8a743","name":"GLM-10B-unidirectional","organization":"Tsinghua University,Beijing Academy of Artificial Intelligence / BAAI,Massachusetts Institute of Technology (MIT)","date":"2021-03-18","domains":["Language"],"country":"China,China,United States of America","reference":"https://arxiv.org/abs/2103.10360","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:15899072b33d0453dea9","name":"Very Deep VAEs (ImageNet-64)","organization":"OpenAI","date":"2021-03-16","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2011.10650","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ae11a9712d69e2372e03","name":"ResNet-RS","organization":"Google Brain,University of California (UC) Berkeley","date":"2021-03-13","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2103.07579","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0c650a93c696d51c87b0","name":"AraELECTRA","organization":"American University of Beirut","date":"2021-03-07","domains":["Language"],"country":"Lebanon","reference":"https://arxiv.org/abs/2012.15516","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:addbb32780d207d965f9","name":"M6-T","organization":"Alibaba","date":"2021-03-05","domains":["Multimodal","Language","Vision"],"country":"China","reference":"https://arxiv.org/abs/2105.15082","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:8d734df6039bb55e33ff","name":"Generative BST","organization":"Facebook AI Research","date":"2021-03-05","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/2004.13637","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:282a1f7130ec84ec76a7","name":"DCTransformer (ImageNet)","organization":"DeepMind","date":"2021-03-05","domains":["Image generation"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2103.03841","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:29c61267d53c46c9999f","name":"ProteinGAN","organization":"Vilnius University,Chalmers University of Technology","date":"2021-03-04","domains":["Biology"],"country":"Lithuania,Sweden","reference":"https://www.nature.com/articles/s42256-021-00310-5","confidence":"Confident","accessibility":""},{"id":"epoch:bcd714428b729d36ecaf","name":"RFA-GATE-Gaussian-Stateful Big","organization":"University of Washington,DeepMind,Allen Institute for AI,Hebrew University of Jerusalem,The University of Hong Kong","date":"2021-03-03","domains":["Language"],"country":"United States of America,United Kingdom,United States of America,Israel,Hong Kong","reference":"https://arxiv.org/abs/2103.02143","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:6076832c40500d5cfb96","name":"Wu Dao - Wen Hui","organization":"Beijing Academy of Artificial Intelligence / BAAI","date":"2021-03-01","domains":["Language","Multimodal","Video","Image generation"],"country":"China","reference":"https://medium.com/syncedreview/chinas-gpt-3-baai-introduces-superscale-intelligence-model-wu-dao-1-0-98a573fc4d70","confidence":"Confident","accessibility":""},{"id":"epoch:7cb4cf8154d7dfd794af","name":"Wu Dao - Wen Lan","organization":"Beijing Academy of Artificial Intelligence / BAAI","date":"2021-03-01","domains":["Multimodal","Vision","Language"],"country":"China","reference":"https://medium.com/syncedreview/chinas-gpt-3-baai-introduces-superscale-intelligence-model-wu-dao-1-0-98a573fc4d70","confidence":"Confident","accessibility":""},{"id":"epoch:b821d3197f333bb3ff2a","name":"Wu Dao - Wen Su","organization":"Beijing Academy of Artificial Intelligence / BAAI","date":"2021-03-01","domains":["Biology"],"country":"China","reference":"https://medium.com/syncedreview/chinas-gpt-3-baai-introduces-superscale-intelligence-model-wu-dao-1-0-98a573fc4d70","confidence":"Unknown","accessibility":"Open weights (non-commercial)"},{"id":"epoch:c7222f7e86beb544ab56","name":"Meta Pseudo Labels","organization":"Google Brain,Google AI","date":"2021-03-01","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2003.10580","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:12655724cabfa32e998d","name":"M6-10B","organization":"Tsinghua University,Alibaba","date":"2021-03-01","domains":["Multimodal","Image generation","Vision","Language"],"country":"China,China","reference":"https://arxiv.org/abs/2103.00823","confidence":"","accessibility":""},{"id":"epoch:694ecf1e0c3440162b24","name":"M6-100B","organization":"Tsinghua University,Alibaba","date":"2021-03-01","domains":["Multimodal","Image generation","Vision","Language"],"country":"China,China","reference":"https://arxiv.org/abs/2103.00823","confidence":"","accessibility":""},{"id":"epoch:98a667db34e87bc27a95","name":"SRU++ Large","organization":"ASAPP","date":"2021-02-24","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2102.12459","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:87c6f0ecdd760c42435f","name":"SRU++ Base","organization":"ASAPP","date":"2021-02-24","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2102.12459","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c483510cb078e08fdb06","name":"SRU++ Large only 2 attention layers (k=5) (WT103)","organization":"ASAPP","date":"2021-02-24","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2102.12459","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:3a684fc6d4f41c9736bc","name":"Linear Transformer (large)","organization":"IDSIA","date":"2021-02-22","domains":["Language"],"country":"Switzerland","reference":"https://arxiv.org/abs/2102.11174","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c0c1c5a1e05738039866","name":"Linear Transformer (small)","organization":"IDSIA,SUPSI","date":"2021-02-22","domains":["Language"],"country":"Switzerland,Switzerland","reference":"https://arxiv.org/abs/2102.11174","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:cdd7dd3d535cb26f1e1b","name":"Rational DQN Average","organization":"TU Darmstadt","date":"2021-02-18","domains":["Games"],"country":"Germany","reference":"https://arxiv.org/abs/2102.09407v3","confidence":"","accessibility":"Unreleased"},{"id":"epoch:244820cc6e17de13c4a4","name":"MSA Transformer","organization":"Facebook AI Research,University of California (UC) Berkeley,New York University (NYU)","date":"2021-02-13","domains":["Biology"],"country":"United States of America,France,United States of America,United States of America","reference":"https://proceedings.mlr.press/v139/rao21a/rao21a.pdf","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:266d1266561c3db9ff9b","name":"top-down frozen classifier","organization":"University of Edinburgh,Toshiba Cambridge Research Laboratory","date":"2021-02-09","domains":["Speech"],"country":"United Kingdom,United Kingdom","reference":"https://arxiv.org/abs/2102.04697","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:b6fdaf376c590faf045f","name":"DLWP","organization":"University of Washington,Microsoft Research","date":"2021-02-09","domains":["Earth science"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2102.05107","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b13b06e111dceb8cd94b","name":"CryoDRGN","organization":"Massachusetts Institute of Technology (MIT)","date":"2021-02-04","domains":["Biology"],"country":"United States of America","reference":"https://www.nature.com/articles/s41592-020-01049-4","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:2651e81e912b916058ff","name":"Culturome","organization":"","date":"2021-01-31","domains":[],"country":"","reference":"https://github.com/YongxinLiu/Culturome","confidence":"Unknown","accessibility":""},{"id":"epoch:81ee2f02421de8c3b1a7","name":"Selfish-RNN (SNT-ASGD) Stacked LSTMs","organization":"Eindhoven University of Technology,University of Twente","date":"2021-01-22","domains":["Language"],"country":"Netherlands,Netherlands","reference":"https://arxiv.org/abs/2101.09048","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:d982d8570226766ab226","name":"Selfish-RNN (ON-LSTM)","organization":"Eindhoven University of Technology","date":"2021-01-22","domains":["Language"],"country":"Netherlands","reference":"https://arxiv.org/abs/2101.09048","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:7a8f437a7b8ae9c126dc","name":"Selfish-RNN (SNT-ASGD)RHNs","organization":"Eindhoven University of Technology","date":"2021-01-22","domains":["Language"],"country":"Netherlands","reference":"https://arxiv.org/abs/2101.09048","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:83d92804a1526ab121fd","name":"Selfish-RNN (AWD-LSTM-MoS)","organization":"Eindhoven University of Technology","date":"2021-01-22","domains":["Language"],"country":"Netherlands","reference":"https://arxiv.org/abs/2101.09048","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:10367c26598160f4dd8d","name":"DeiT-B","organization":"Meta AI,Sorbonne University","date":"2021-01-15","domains":["Vision"],"country":"United States of America,France","reference":"https://arxiv.org/abs/2012.12877","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2ec4ad64a099389e608e","name":"Switch","organization":"Google","date":"2021-01-11","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2101.03961","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:24f2af8a2d3cea1c1e42","name":"Wu Dao - Wen Yuan","organization":"Beijing Academy of Artificial Intelligence / BAAI","date":"2021-01-11","domains":["Language"],"country":"China","reference":"https://mp.weixin.qq.com/s/BUQWZ5EdR19i40GuFofpBg","confidence":"Confident","accessibility":"API access"},{"id":"epoch:902fa5737c5b275519da","name":"BigSSL","organization":"Google,Apple","date":"2021-01-10","domains":["Speech"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2109.13226","confidence":"","accessibility":"Unreleased"},{"id":"epoch:7db1c6603c4bf627ffba","name":"NVAE (CIFAR 10)","organization":"Nvidia","date":"2021-01-08","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2007.03898","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:a318e029589a76ccd332","name":"NVAE (FFHQ)","organization":"Nvidia","date":"2021-01-08","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2007.03898","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:9cca06037f6c9ad0df8d","name":"NVAE (Celeba HQ)","organization":"Nvidia","date":"2021-01-08","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/2007.03898","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:1ce6889dabab09cbbf01","name":"DALL-E","organization":"OpenAI","date":"2021-01-05","domains":["Image generation"],"country":"United States of America","reference":"https://openai.com/blog/dall-e/\n\nhttps://arxiv.org/abs/2102.12092","confidence":"Likely","accessibility":"API access"},{"id":"epoch:36e3217471ef0dbaca1e","name":"CLIP (ViT L/14@336px)","organization":"OpenAI","date":"2021-01-05","domains":["Multimodal","Vision","Language","Video"],"country":"United States of America","reference":"https://arxiv.org/abs/2103.00020\nhttps://huggingface.co/openai/clip-vit-large-patch14-336","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7d6bdb1ac05d3828b865","name":"CLIP (ResNet-50)","organization":"OpenAI","date":"2021-01-05","domains":["Multimodal","Vision","Language","Video"],"country":"United States of America","reference":"https://arxiv.org/abs/2103.00020","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e2e9bda85e5602c914b5","name":"Transformer-XL + AutoDropout (WT2)","organization":"Google Brain,Carnegie Mellon University (CMU)","date":"2021-01-05","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2101.01761","confidence":"","accessibility":"Unreleased"},{"id":"epoch:107acd094686bdce01ff","name":"Transformer-XL + AutoDropout (PTB)","organization":"Google Research","date":"2021-01-05","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2101.01761","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:ac97cf55027d547127d4","name":"Subformer (122M)","organization":"National Institute of Advanced Industrial Science and Technology (AIST),University of Tokyo","date":"2021-01-01","domains":["Language"],"country":"Japan,Japan","reference":"https://arxiv.org/abs/2101.00234","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e51e84bb9ae76ea361ed","name":"Subformer (83M)","organization":"University of Tokyo,National Institute of Advanced Industrial Science and Technology (AIST)","date":"2021-01-01","domains":["Language"],"country":"Japan,Japan","reference":"https://arxiv.org/abs/2101.00234","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:4cbdea9cf5e765c9c577","name":"Subformer (96M)","organization":"University of Tokyo,National Institute of Advanced Industrial Science and Technology (AIST)","date":"2021-01-01","domains":["Language"],"country":"Japan,Japan","reference":"https://arxiv.org/abs/2101.00234","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:7c3314c136fbb58b5acb","name":"AraGPT2-Mega","organization":"American University of Beirut","date":"2020-12-31","domains":["Language"],"country":"Lebanon","reference":"https://arxiv.org/abs/2012.15520","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:17c6bb142c34161dfb60","name":"Shortformer","organization":"University of Washington,Facebook AI Research,Allen Institute for AI","date":"2020-12-31","domains":["Language"],"country":"United States of America,United States of America,France,United States of America","reference":"https://arxiv.org/abs/2012.15832","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:a93e30a618c403e9261f","name":"ERNIE-Doc (247M)","organization":"Baidu","date":"2020-12-31","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2012.15688","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:58d3423d61d1e57cc461","name":"ERNIE-Doc Base (151M, WT103)","organization":"Baidu","date":"2020-12-31","domains":["Language"],"country":"China","reference":"https://arxiv.org/pdf/2012.15688","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0190fbb316332f5486ca","name":"CT-MoS (WT2)","organization":"Google,National Tsing Hua University","date":"2020-12-25","domains":["Language"],"country":"United States of America,Taiwan","reference":"https://arxiv.org/abs/2012.13575","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:fdd4b0ca2bcb129bc682","name":"CT-MoS + DynamicEval (WT2)","organization":"National Tsing Hua University,Google","date":"2020-12-25","domains":["Language"],"country":"Taiwan,United States of America","reference":"https://arxiv.org/abs/2012.13575","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2e2340febba6e90bf7fd","name":"CT-MoS (PTB)","organization":"National Tsing Hua University,Google","date":"2020-12-25","domains":["Language"],"country":"Taiwan,United States of America","reference":"https://arxiv.org/pdf/2012.13575","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c33e8a9c0d3ad3a806a9","name":"CT-MoS + DynamicEval (PTB)","organization":"National Tsing Hua University,Google","date":"2020-12-25","domains":["Language"],"country":"Taiwan,United States of America","reference":"https://arxiv.org/pdf/2012.13575","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:5e9c2b40a5058cf5e415","name":"DensePhrases","organization":"Korea University,Princeton University","date":"2020-12-23","domains":["Language"],"country":"South Korea,United States of America","reference":"https://arxiv.org/abs/2012.12624v3","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5514d3c2c646bd7c100c","name":"RaSoR","organization":"Korea University,Princeton University","date":"2020-12-23","domains":["Language"],"country":"South Korea,United States of America","reference":"https://arxiv.org/abs/1611.01436","confidence":"Confident","accessibility":""},{"id":"epoch:f7b1faeb5466d80f72e4","name":"VQGAN + CLIP","organization":"Heidelberg University","date":"2020-12-17","domains":["Image generation"],"country":"Germany","reference":"https://arxiv.org/abs/2012.09841","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4f8bdd4451902411af4d","name":"ESM1b","organization":"Facebook AI Research,New York University (NYU)","date":"2020-12-15","domains":["Biology"],"country":"United States of America,France,United States of America","reference":"https://www.pnas.org/doi/abs/10.1073/pnas.2016239118","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0a81cd87652b74a35ecd","name":"RoBERTa (PFAM)","organization":"IBM Research,ETH Zurich","date":"2020-12-05","domains":["Biology"],"country":"United States of America,Switzerland","reference":"https://arxiv.org/abs/2012.03084","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4a6195752af4bab2be74","name":"OC-GAN (Visual Genome)","organization":"Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Universit\u00e9 de Montr\u00e9al,Microsoft Research,CIFAR AI Research","date":"2020-12-03","domains":["Image generation"],"country":"Canada,Canada,United States of America,Canada","reference":"https://arxiv.org/abs/2003.07449","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:b524d230688ac0334db4","name":"OC-GAN (COCO-Stuff)","organization":"Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Universit\u00e9 de Montr\u00e9al,Microsoft Research,CIFAR AI Research","date":"2020-12-03","domains":["Image generation"],"country":"Canada,Canada,United States of America,Canada","reference":"https://arxiv.org/abs/2003.07449","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:c23bf006ec9381a26fb4","name":"CPM-Large","organization":"Tsinghua University,Beijing Academy of Artificial Intelligence / BAAI","date":"2020-12-01","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2012.00413","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fd143bbc8e7a04f74ea5","name":"Profile Prediction","organization":"University of Washington,Salesforce Research","date":"2020-12-01","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2012.00195","confidence":"Confident","accessibility":""},{"id":"epoch:05e1a83492f6c2fe76e5","name":"AlphaFold 2","organization":"DeepMind","date":"2020-11-30","domains":["Biology"],"country":"United Kingdom","reference":"https://www.nature.com/articles/s41586-021-03819-2","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c6f1b943cfc3ed27eb9d","name":"KEPLER","organization":"Tsinghua University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),HEC,CIFAR AI Research,Princeton University,University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2020-11-23","domains":["Language"],"country":"China,Canada,France,Canada,United States of America,Canada","reference":"https://arxiv.org/abs/1911.06136","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0e0c466c45078951e9b4","name":"AWD-FWM (WT2)","organization":"IDSIA,Microsoft Research","date":"2020-11-16","domains":["Language"],"country":"Switzerland,United States of America","reference":"https://arxiv.org/abs/2011.07831","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:9dba0d79a4a573e5b5c8","name":"AWD-FWM (PTB)","organization":"IDSIA,Microsoft Research","date":"2020-11-16","domains":["Language"],"country":"Switzerland,United States of America","reference":"https://arxiv.org/abs/2011.07831","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:3430d49d181fdcd4ed05","name":"Machine learning a model for RNA structure prediction","organization":"International School for Advanced Studies,Institute of Structural Biology,Technical University of Munich","date":"2020-11-16","domains":["Biology"],"country":"Italy,Germany,Germany","reference":"https://academic.oup.com/nargab/article/2/4/lqaa090/5983421","confidence":"Confident","accessibility":""},{"id":"epoch:2f899a0787f8aa2512c6","name":"CPCProt","organization":"University of Toronto","date":"2020-11-10","domains":["Biology"],"country":"Canada","reference":"https://www.biorxiv.org/content/10.1101/2020.09.04.283929v2.abstract","confidence":"Confident","accessibility":""},{"id":"epoch:b4ee923e8ec426817135","name":"SimCLRv2","organization":"Google Brain","date":"2020-10-26","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2006.10029","confidence":"","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e4d9b6cea82d74e81b8c","name":"HiPPO-LegS","organization":"Stanford University,University at Buffalo","date":"2020-10-23","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2008.07669","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:09d38eb1e7694bd6938a","name":"ChemBERTa","organization":"University of Toronto,Reverie Labs,DeepChem","date":"2020-10-23","domains":["Biology"],"country":"Canada,United States of America,United States of America","reference":"https://arxiv.org/abs/2010.09885","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e470cda56c8437ff6355","name":"ViT-Huge/14","organization":"Google Brain,Google Research","date":"2020-10-22","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2010.11929","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3524bbd4bc0f355d9e96","name":"wave2vec 2.0 LARGE","organization":"Facebook","date":"2020-10-22","domains":["Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2006.11477","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b87ea54bb1337002f6d4","name":"ViT-Base/32","organization":"Google Brain","date":"2020-10-22","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2010.11929","confidence":"","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a5cf2ff8be137fdcbdec","name":"CryptoGRU","organization":"Indiana University Bloomington","date":"2020-10-22","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2010.11796","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:af841ccbcc5cea16810e","name":"GBERT-Large","organization":"deepset,Bayerische Staatsbibliothek Muenchen","date":"2020-10-21","domains":["Language"],"country":"Germany,Germany","reference":"https://arxiv.org/abs/2010.10906","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fa292e8a6b283a7c98a2","name":"German ELECTRA Large","organization":"deepset,Bayerische Staatsbibliothek Muenchen","date":"2020-10-21","domains":["Language"],"country":"Germany,Germany","reference":"https://arxiv.org/abs/2010.10906","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8d4434e242583fac82c6","name":"Conformer + Wav2vec 2.0 + Noisy Student","organization":"Google,Google Research,Google Brain","date":"2020-10-20","domains":["Speech"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2010.10504v2","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:1dbf13ef8506490addd3","name":"mT5-XXL","organization":"Google,Google Research","date":"2020-10-20","domains":["Language"],"country":"United States of America,United States of America","reference":"https://aclanthology.org/2021.naacl-main.41/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0a979b78061925d9c526","name":"TinyBert","organization":"Huazhong University of Science and Technology,Huawei Noah's Ark Lab,Huawei","date":"2020-10-16","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/abs/1909.10351","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a6e8578c740d742cff41","name":"Memformer (4 encoder + 16 decoder)","organization":"UC Davis,Westlake University,Facebook AI","date":"2020-10-14","domains":["Language"],"country":"United States of America,China,United States of America","reference":"https://arxiv.org/abs/2010.06891","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:24d1196e048287d8108e","name":"LUKE","organization":"University of Washington,National Institute of Informatics","date":"2020-10-02","domains":["Language"],"country":"United States of America,Japan","reference":"https://arxiv.org/abs/2010.01057v1","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ee6e396ae7567b829ad9","name":"Frage-AWD-LSTM-MemoryAug-NeuralCache (WT2)","organization":"Johns Hopkins University,Xiaomi Corp","date":"2020-09-29","domains":["Language"],"country":"United States of America,China","reference":"https://arxiv.org/abs/2009.13774","confidence":"","accessibility":"Unreleased"},{"id":"epoch:989aeafe91294d1c5d3d","name":"Frage-AWD-LSTM-MemoryAug-NeuralCache (PTB)","organization":"Johns Hopkins University,Xiaomi Corp","date":"2020-09-29","domains":["Language"],"country":"United States of America,China","reference":"https://arxiv.org/abs/2009.13774","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:44ee7fcb014ad7e8a2ff","name":"LSTM-MemoryAug (WT2)","organization":"Johns Hopkins University,Xiaomi Corp","date":"2020-09-29","domains":["Language"],"country":"United States of America,China","reference":"https://arxiv.org/abs/2009.13774","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2aa85a81d99edfb5c741","name":"LSTM-MemoryAug (PTB)","organization":"Johns Hopkins University,Xiaomi Corp","date":"2020-09-29","domains":["Language"],"country":"United States of America,China","reference":"https://arxiv.org/abs/2009.13774","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:eac015e7c61f73a4b136","name":"Integer Transformer","organization":"Northeastern University (China),NiuTrans Research,Chinese Academy of Sciences","date":"2020-09-17","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/abs/2009.08034","confidence":"","accessibility":"Unreleased"},{"id":"epoch:32fb71ac77a2e7b126f6","name":"PAR Transformer Large","organization":"Nvidia","date":"2020-09-09","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2009.04534","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:bfe80d36926c62a1755c","name":"ProBERTa","organization":"University of Illinois Urbana-Champaign (UIUC),Reed College","date":"2020-09-01","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://dl.acm.org/doi/10.1145/3388440.3412467","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:3bb512ef16c8f6b37ed6","name":"ESM1-670M (UR50/S)","organization":"Facebook AI Research,New York University (NYU)","date":"2020-08-31","domains":["Biology"],"country":"United States of America,France,United States of America","reference":"https://www.pnas.org/doi/abs/10.1073/pnas.2016239118","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3e10c4f0a699ff50254b","name":"ESM1-670M (UR50/D)","organization":"Facebook AI Research,New York University (NYU)","date":"2020-08-31","domains":["Biology"],"country":"United States of America,France,United States of America","reference":"https://www.pnas.org/doi/abs/10.1073/pnas.2016239118","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:cc508b5e69c5870a16d0","name":"ESM1-670M (UR100)","organization":"Facebook AI Research,New York University (NYU)","date":"2020-08-31","domains":["Biology"],"country":"United States of America,France,United States of America","reference":"https://www.pnas.org/doi/abs/10.1073/pnas.2016239118","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e14ff3911af1827481b2","name":"ESM1-85M","organization":"Facebook AI Research,New York University (NYU)","date":"2020-08-31","domains":["Biology"],"country":"United States of America,France,United States of America","reference":"https://www.pnas.org/doi/abs/10.1073/pnas.2016239118","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:40522bc5b3e2ebbfec63","name":"ESM1-43M","organization":"Facebook AI Research,New York University (NYU)","date":"2020-08-31","domains":["Biology"],"country":"United States of America,France,United States of America","reference":"https://www.pnas.org/doi/abs/10.1073/pnas.2016239118","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:cb4268cdc7d270d984b7","name":"Transformer+Recurrent Windows of Context","organization":"Toyota Technological Institute at Chicago,University of Chicago","date":"2020-08-16","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2008.07027","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:bb4179727732d2dc7c4b","name":"ERNIE-GEN (large)","organization":"Baidu","date":"2020-08-06","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2001.11314","confidence":"Speculative","accessibility":"Open weights (non-commercial)"},{"id":"epoch:1dd12c50c40d0d205818","name":"DeLighT","organization":"University of Washington,Allen Institute for AI,Facebook AI Research","date":"2020-08-03","domains":["Language"],"country":"United States of America,United States of America,United States of America,France","reference":"https://arxiv.org/abs/2008.00623","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:c3f6059297f011e8e142","name":"mBART-50","organization":"Facebook AI","date":"2020-08-02","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2008.00401\n\nhttps://huggingface.co/facebook/mbart-large-50","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:802898be586c32ca4712","name":"Grown to Prune Two-layer stacked LSTM","organization":"University of Chicago,Toyota Technological Institute at Chicago","date":"2020-07-30","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2007.15353","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:0896deeed060f82f0a90","name":"TransformerXL-LayerFusion-CA","organization":"University of Liverpool,University of Southern California","date":"2020-07-29","domains":["Language"],"country":"United Kingdom,United States of America","reference":"https://arxiv.org/abs/2007.14917","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:9c4a0ab3782516bbf0a0","name":"GPT2-LayerFusion-WS","organization":"University of Liverpool,University of Southern California","date":"2020-07-29","domains":["Language"],"country":"United Kingdom,United States of America","reference":"https://arxiv.org/abs/2007.14917","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:d98db577a57ce012e639","name":"EfficientDet","organization":"Google Brain","date":"2020-07-27","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1911.09070","confidence":"","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0590486a9506f0e2d00c","name":"Hopfield Networks (2020)","organization":"Johannes Kepler University Linz,Institute of Advanced Research in Artificial Intelligence,University of Oslo","date":"2020-07-16","domains":["Biology","Vision","Language","Medicine"],"country":"Austria,Austria,Norway","reference":"https://arxiv.org/abs/2008.02217","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ee4ad85fdb1450d90805","name":"1-layer-LSTM","organization":"Harvard University","date":"2020-07-13","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2007.06389","confidence":"","accessibility":"Unreleased"},{"id":"epoch:a560d08f779d24d1888d","name":"SemExp","organization":"Carnegie Mellon University (CMU),Facebook AI Research","date":"2020-07-02","domains":["Robotics"],"country":"United States of America,United States of America,France","reference":"https://proceedings.neurips.cc/paper/2020/file/2c75cf2681788adaca63aa95ae028b22-Paper.pdf\nhttps://arxiv.org/abs/2007.00643","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e04956aa34d6b689777a","name":"DLRM-2021","organization":"Meta AI","date":"2020-07-01","domains":["Recommendation"],"country":"United States of America","reference":"https://arxiv.org/abs/2104.05158","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:11c7ff68e4e1fb8e105b","name":"GShard (dense)","organization":"Google","date":"2020-06-30","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2006.16668","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0950ef59dda702f47070","name":"GShard (600B)","organization":"Google","date":"2020-06-30","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2006.16668","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e904d1ab3d25dbce7cdc","name":"GPT-3 6.7B","organization":"OpenAI","date":"2020-06-22","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2005.14165","confidence":"Confident","accessibility":""},{"id":"epoch:6684f356882a0980747d","name":"GPT-3 XL","organization":"OpenAI","date":"2020-06-22","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2005.14165","confidence":"Confident","accessibility":""},{"id":"epoch:9690d1f47fe040daf9fe","name":"GPT-3 Small","organization":"OpenAI","date":"2020-06-22","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2005.14165","confidence":"Confident","accessibility":""},{"id":"epoch:6ba49112fdbdd03cf602","name":"GPT-3 Medium","organization":"OpenAI","date":"2020-06-22","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2005.14165","confidence":"Confident","accessibility":""},{"id":"epoch:a08435f5c951ae46f3ae","name":"GPT-3 Large","organization":"OpenAI","date":"2020-06-22","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2005.14165","confidence":"Confident","accessibility":""},{"id":"epoch:a88c26838d31cd926259","name":"GPT-3 2.7B","organization":"OpenAI","date":"2020-06-22","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2005.14165","confidence":"Confident","accessibility":""},{"id":"epoch:1ae789f2543baed88459","name":"GPT-3 13B","organization":"OpenAI","date":"2020-06-22","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2005.14165","confidence":"Confident","accessibility":""},{"id":"epoch:75bc94ec2ab97dd825d1","name":"iGPT-XL","organization":"OpenAI","date":"2020-06-17","domains":["Vision","Image generation"],"country":"United States of America","reference":"https://openai.com/research/image-gpt","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3f4bd1e64aa4f4c06905","name":"iGPT-L","organization":"OpenAI","date":"2020-06-17","domains":["Image generation","Vision"],"country":"United States of America","reference":"https://openai.com/blog/image-gpt/","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f3c08cc420de98eb067c","name":"Transformer-XL+AdamP","organization":"Naver AI Lab,Naver Clova","date":"2020-06-15","domains":["Language"],"country":"South Korea,South Korea","reference":"https://arxiv.org/abs/2006.08217","confidence":"","accessibility":"Unreleased"},{"id":"epoch:9936b7f12fd375c78a52","name":"Transformer-XL+WN+AdamP","organization":"Naver AI Lab,Naver Clova","date":"2020-06-15","domains":["Language"],"country":"South Korea,South Korea","reference":"https://arxiv.org/abs/2006.08217","confidence":"","accessibility":"Unreleased"},{"id":"epoch:2cbbf35695663a3e0870","name":"SqueezeBERT","organization":"University of California (UC) Berkeley","date":"2020-06-10","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2006.11316","confidence":"","accessibility":""},{"id":"epoch:f8de8170c6750b99b599","name":"6-Layer-Tensor-Transformer+AdaHessian","organization":"NERSC, Lawrence Berkeley National Laboratory,University of California (UC) Berkeley","date":"2020-06-01","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2006.00719","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c93fd0bcc9c20af6c4f6","name":"3-Layer-Tensor-Transformer+AdaHessian","organization":"University of California (UC) Berkeley,NERSC, Lawrence Berkeley National Laboratory","date":"2020-06-01","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2006.00719","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e094cfe34cc090abb3fe","name":"GPT-3 175B (davinci)","organization":"OpenAI","date":"2020-05-28","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2005.14165","confidence":"Confident","accessibility":"API access"},{"id":"epoch:fcb06caf1bf69516d79f","name":"GPT3-6.7B (rerun of original)","organization":"Microsoft,OpenAI","date":"2020-05-28","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2203.03466","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e4ba797ce4ecfe64267b","name":"DETR","organization":"Facebook","date":"2020-05-26","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2005.12872","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:769ac90d053d56465f5b","name":"Retrieval-Augmented Generator","organization":"Facebook,New York University (NYU),University College London (UCL)","date":"2020-05-22","domains":["Language"],"country":"United States of America,United States of America,United Kingdom","reference":"https://arxiv.org/abs/2005.11401v4","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c9a3cf01b2ef0a9edd2a","name":"rTop-k(distributed setting)","organization":"Stanford University","date":"2020-05-21","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2005.10761","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:491f8cbbe49a42edba80","name":"Specter","organization":"Allen Institute for AI,University of Washington","date":"2020-05-20","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2004.07180","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:138dd3d67b287a9236e0","name":"Decay RNN","organization":"Indian Institute of Technology Delhi","date":"2020-05-17","domains":["Language"],"country":"India","reference":"https://arxiv.org/abs/2005.08199","confidence":"","accessibility":"Unreleased"},{"id":"epoch:af338a663fbf2728f3df","name":"Conformer","organization":"Google","date":"2020-05-16","domains":["Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2005.08100v1","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0bf3c2d354b10a9fe7cc","name":"ONLSTM-SYD","organization":"Westlake University,Institute for Advanced Study,McGill University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),CIFAR AI Research,University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2020-05-12","domains":["Language"],"country":"China,United States of America,Canada,Canada,Canada,Canada","reference":"https://arxiv.org/abs/2005.05864","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:47877fd9ba6da1ce3f35","name":"ContextNet","organization":"Google","date":"2020-05-07","domains":["Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2005.03191v3","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:d47de3eaba7cee620c59","name":"NAS+ESS (156M)","organization":"Northeastern University (China),Chinese Academy of Sciences,NiuTrans Research,Kingsoft","date":"2020-05-06","domains":["Language"],"country":"China,China,China,China","reference":"https://arxiv.org/abs/2005.02593","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:7dae58133ce42b3d828d","name":"NAS+ESS (23M)","organization":"Northeastern University (China),NiuTrans Research,Kingsoft","date":"2020-05-06","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/pdf/2005.02593","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:a4a696178fce68b414f5","name":"UnifiedQA","organization":"Allen Institute for AI,University of Washington","date":"2020-05-02","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2005.00700v3","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:09a8bfee80758efe38ac","name":"Segatron XL large, M=384","organization":"University of Waterloo,Peking University,RSVP.ai","date":"2020-04-30","domains":["Language"],"country":"Canada,China,China","reference":"https://arxiv.org/abs/2004.14996","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:87c38c18d0b8b8d75dd0","name":"Segatron XL base, M=384","organization":"University of Waterloo,RSVP.ai,Peking University","date":"2020-04-30","domains":["Language"],"country":"Canada,China,China","reference":"https://arxiv.org/abs/2004.14996","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:7d4545c9db0c311ecc76","name":"Once for All","organization":"MIT-IBM Watson AI Lab,Massachusetts Institute of Technology (MIT),IBM","date":"2020-04-29","domains":["Vision"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/1908.09791","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5ae4aed36591569f5902","name":"Go-explore","organization":"Uber AI,OpenAI","date":"2020-04-27","domains":["Games"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2004.12919","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:17d17781c68825376175","name":"Cube-Space AutoEncoder","organization":"MIT-IBM Watson AI Lab","date":"2020-04-27","domains":["Vision","Search"],"country":"United States of America","reference":"https://arxiv.org/abs/2004.12850","confidence":"Confident","accessibility":""},{"id":"epoch:576486cbbb1243764c74","name":"CURL","organization":"University of California (UC) Berkeley","date":"2020-04-08","domains":["Games"],"country":"United States of America","reference":"https://arxiv.org/abs/2004.04136v4","confidence":"","accessibility":"Open weights (unrestricted)"},{"id":"epoch:c3bb188582e2893c7be5","name":"MobileBERT","organization":"Carnegie Mellon University (CMU),Google Brain","date":"2020-04-06","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2004.02984","confidence":"","accessibility":"Open weights (unrestricted)"},{"id":"epoch:072d3c6b8c095f3de181","name":"DiffStk-MRNN","organization":"Pennsylvania State University,Rochester Institute of Technology","date":"2020-04-04","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2004.07623","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:7bcaa4a54c740673e9b4","name":"Agent57","organization":"DeepMind","date":"2020-03-30","domains":["Games"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2003.13350","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:cecd2fcaeaf610c81874","name":"AraBERT","organization":"American University of Beirut","date":"2020-03-30","domains":["Language"],"country":"Lebanon","reference":"https://arxiv.org/abs/2003.00104v2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ebfde7da14e00260ec03","name":"AraBERT LArge v2","organization":"American University of Beirut","date":"2020-03-30","domains":["Language"],"country":"Lebanon","reference":"https://huggingface.co/aubmindlab/bert-large-arabertv2","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:276ed5dd9808d3ef899a","name":"MetNet","organization":"Google","date":"2020-03-24","domains":["Earth science"],"country":"United States of America","reference":"https://arxiv.org/abs/2003.12140","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:a16b64438e269f24af8e","name":"ELECTRA","organization":"Stanford University,Google,Google Brain","date":"2020-03-23","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/2003.10555v1","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a2328cd71773fd140da1","name":"Tensor-Transformer(1core)+PN (WT103)","organization":"University of California (UC) Berkeley","date":"2020-03-17","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2003.07845","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:11654f337680d9f8c7a5","name":"Tensor-Transformer(1core)+PN (PTB)","organization":"University of California (UC) Berkeley","date":"2020-03-17","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/pdf/2003.07845","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:55bdb43af3ec94dd4c56","name":"WDC20 / DLWP","organization":"University of Washington,Microsoft Research","date":"2020-03-15","domains":["Earth science"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/2003.11927","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:40fb5d88efbcab2a3732","name":"ProGen","organization":"Salesforce Research,Stanford University","date":"2020-03-13","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://www.biorxiv.org/content/10.1101/2020.03.07.982272v2","confidence":"Likely","accessibility":""},{"id":"epoch:f48990ebace3f82ec5e1","name":"Routing Transformer (WT-103)","organization":"Google Research","date":"2020-03-12","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2003.05997","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8e4346d9607fb6250cd5","name":"Local Transformer (WT103)","organization":"Google Research","date":"2020-03-12","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2003.05997","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f2d633ea7ab2bf1d4629","name":"TransformerXL + spectrum control","organization":"University of California Los Angeles (UCLA),JD.com","date":"2020-03-11","domains":["Language"],"country":"United States of America,China","reference":"https://openreview.net/forum?id=ByxY8CNtvr","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:b2a38e71cbf6f096c84e","name":"MuPIPR","organization":"University of California Los Angeles (UCLA),University of Pennsylvania","date":"2020-03-05","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://academic.oup.com/nargab/article/2/2/lqaa015/5781175","confidence":"Confident","accessibility":""},{"id":"epoch:d080c2a4f16f509f3778","name":"LSTM-3-layer+Gadam","organization":"University of Oxford,University of Bristol,University of Cambridge","date":"2020-03-02","domains":["Language"],"country":"United Kingdom,United Kingdom,United Kingdom","reference":"https://arxiv.org/abs/2003.01247","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:7c48f2c57c8bd74c7f7d","name":"TCAN (WT2)","organization":"Nanjing University,Ant Group","date":"2020-02-28","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/2002.12530","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:74921702da8e493aed79","name":"TCAN (PTB)","organization":"Ant Group","date":"2020-02-28","domains":["Language"],"country":"China","reference":"https://arxiv.org/abs/2002.12530","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:13fc6f983d829587433b","name":"Feedback Transformer","organization":"LORIA,University of Lorraine,Facebook AI Research","date":"2020-02-21","domains":["Language"],"country":"France,France,United States of America,France","reference":"https://arxiv.org/abs/2002.09402","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:a05e0be300e82af6fe94","name":"FFN SwiGLU","organization":"Google","date":"2020-02-14","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2002.05202","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4f94997bae01afc2b750","name":"SimCLR","organization":"Google Brain","date":"2020-02-13","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/2002.05709","confidence":"","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d18079e5aa70ffa6cc01","name":"Turing-NLG","organization":"Microsoft","date":"2020-02-13","domains":["Language"],"country":"United States of America","reference":"https://www.microsoft.com/en-us/research/blog/turing-nlg-a-17-billion-parameter-language-model-by-microsoft/","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:4f21e61a7b9455d031fd","name":"ALBERT-xxlarge","organization":"Toyota Technological Institute at Chicago,Google","date":"2020-02-09","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1909.11942","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a4d176f6f3c86d827167","name":"Perceiver IO (optical flow)","organization":"DeepMind","date":"2020-02-08","domains":["Multimodal","Language","Vision"],"country":"United Kingdom","reference":"https://arxiv.org/abs/2107.14795","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:da35ddc31def2f2f45ae","name":"TaLK Convolution","organization":"Carleton University","date":"2020-02-08","domains":["Language"],"country":"Canada","reference":"https://arxiv.org/abs/2002.03184","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:d15e5765466034bb6247","name":"Theseus 6/768","organization":"University of California San Diego,Beihang University,Microsoft","date":"2020-02-07","domains":["Language"],"country":"United States of America,China,United States of America","reference":"https://arxiv.org/abs/2002.02925","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:063419e6526fb101675d","name":"Meena","organization":"Google Brain","date":"2020-01-28","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/2001.09977","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4e7932711eea329d0a16","name":"ContextNet + Noisy Student","organization":"Google","date":"2020-01-19","domains":["Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/2005.09629v2","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b0ba69e06f534e9ff457","name":"AlphaFold","organization":"DeepMind","date":"2020-01-15","domains":["Biology"],"country":"United Kingdom","reference":"https://www.nature.com/articles/s41586-019-1923-7","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:a4d2d80f704bd3e5726d","name":"MicroNet (Adaptive, Cache)","organization":"Massachusetts Institute of Technology (MIT),Harvard University","date":"2020-01-01","domains":["Language"],"country":"United States of America,United States of America","reference":"http://proceedings.mlr.press/v123/yan20a.html?ref=https://githubhelp.com","confidence":"","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ca14a68c89fd9a3df952","name":"Big Transfer (BiT-L)","organization":"Google Brain","date":"2019-12-24","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1912.11370","confidence":"","accessibility":"Unreleased"},{"id":"epoch:acd622488ee12f749e7b","name":"Big Transfer (BiT-M)","organization":"Google Brain","date":"2019-12-24","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1912.11370","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:738ee90a971e428ad795","name":"DD-PPO","organization":"Georgia Institute of Technology,Facebook AI Research,Oregon State University,Simon Fraser University","date":"2019-12-19","domains":["Robotics"],"country":"United States of America,United States of America,France,United States of America,Canada","reference":"https://openreview.net/forum?id=H1gX8C4YPr\nhttps://arxiv.org/abs/1911.00357","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:5fa959f68138c1ffd76b","name":"SeqVec","organization":"Technical University of Munich","date":"2019-12-17","domains":["Biology"],"country":"Germany","reference":"https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-019-3220-8","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3127d2533b13a856d92b","name":"OpenAI Five","organization":"OpenAI","date":"2019-12-13","domains":["Games"],"country":"United States of America","reference":"https://arxiv.org/abs/1912.06680","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:066ea7ccabff42090962","name":"OpenAI Five Rerun","organization":"OpenAI","date":"2019-12-13","domains":["Games"],"country":"United States of America","reference":"https://cdn.openai.com/dota-2.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:8f56a29b4c347d484e18","name":"MMLSTM (WT-103)","organization":"Beijing University of Posts and Telecommunications,University of West London","date":"2019-12-05","domains":["Language"],"country":"China,United Kingdom","reference":"http://repository.uwl.ac.uk/id/eprint/6490/1/Loo_etal_IEEE_TNNLS_2019_Major-minor_long_short-term_memory_for_word-level_language_model.pdf","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:6b009687e5049c763486","name":"MMLSTM (WT-2)","organization":"Beijing University of Posts and Telecommunications,University of West London","date":"2019-12-05","domains":["Language"],"country":"China,United Kingdom","reference":"http://repository.uwl.ac.uk/id/eprint/6490/1/Loo_etal_IEEE_TNNLS_2019_Major-minor_long_short-term_memory_for_word-level_language_model.pdf","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:fee0ab3f06b60c4f59d9","name":"MMLSTM (PTB)","organization":"Beijing University of Posts and Telecommunications,University of West London","date":"2019-12-05","domains":["Language"],"country":"China,United Kingdom","reference":"http://repository.uwl.ac.uk/id/eprint/6490/1/Loo_etal_IEEE_TNNLS_2019_Major-minor_long_short-term_memory_for_word-level_language_model.pdf","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:b14389535f253e63f70b","name":"StarGAN v2","organization":"NAVER,Yonsei University,Swiss Federal Institute of Technology","date":"2019-12-04","domains":["Vision","Image generation"],"country":"South Korea,South Korea,Switzerland","reference":"https://arxiv.org/abs/1912.01865","confidence":"Unknown","accessibility":"Open weights (non-commercial)"},{"id":"epoch:e127ce4a6dafc4a8b7df","name":"StyleGAN2","organization":"Nvidia,Aalto University","date":"2019-12-03","domains":["Image generation"],"country":"United States of America,Finland","reference":"https://arxiv.org/pdf/1912.04958","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:efda5f71fd1637c40897","name":"bRSM + cache","organization":"Numenta,Incubator 491","date":"2019-12-02","domains":["Language"],"country":"United States of America,Australia","reference":"https://arxiv.org/abs/1912.01116","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:bae3659bde27c67df9d6","name":"Transformer-XL DeFINE (141M)","organization":"University of Washington,Allen Institute for AI","date":"2019-11-27","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1911.12385","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:83e71e8252542e145856","name":"Transformer-XL DeFINE (107M)","organization":"University of Washington","date":"2019-11-27","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1911.12385","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b562b8aeeb7311296410","name":"Adaptive LSTM + DeFINE","organization":"University of Washington","date":"2019-11-27","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1911.12385","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0bd6ea4b7013a6391e83","name":"AWD-LSTM + DeFINE","organization":"University of Washington","date":"2019-11-27","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1911.12385","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0f735e6278221cdb4507","name":"Photo-Geometric Autoencoder","organization":"University of Oxford","date":"2019-11-25","domains":["3D modeling","Vision"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1911.11130","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b8e3b5c35192fa4fcd13","name":"FastSpeech","organization":"Zhejiang University (ZJU),Microsoft Research","date":"2019-11-20","domains":["Speech"],"country":"China,United States of America","reference":"https://arxiv.org/abs/1905.09263","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:ad2f737912ac18768387","name":"MuZero","organization":"DeepMind","date":"2019-11-19","domains":["Games"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1911.08265v2","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:bf741d44f3368f6b57f9","name":"Transformer - LibriVox + Decoding/Rescoring","organization":"Facebook","date":"2019-11-19","domains":["Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/1911.08460v3","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:67e61a442db13808e5ab","name":"MoCo","organization":"Facebook AI","date":"2019-11-13","domains":["Vision","Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/1911.05722","confidence":"","accessibility":"Open weights (non-commercial)"},{"id":"epoch:88884b8d26e63a2e9548","name":"Long-range sequence Compressive Transformers","organization":"DeepMind","date":"2019-11-13","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1911.05507","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0e230d9dd449b14eef9f","name":"Noisy Student (L2)","organization":"Carnegie Mellon University (CMU),Google","date":"2019-11-11","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1911.04252v4","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:dc548c774db10f79545c","name":"Sandwich Transformer","organization":"Allen Institute for AI,Facebook AI Research","date":"2019-11-10","domains":["Language"],"country":"United States of America,United States of America,France","reference":"https://arxiv.org/abs/1911.03864","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:63fc584f4f80ab678159","name":"CamemBERT","organization":"Facebook,INRIA,Sorbonne University","date":"2019-11-10","domains":["Language"],"country":"United States of America,France,France","reference":"https://arxiv.org/abs/1911.03894","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1a49d518e48fb5b46925","name":"Self-Attention and Convolutional Layers","organization":"Ecole Polytechnique F\u00b4ed\u00b4erale de Lausanne (EPFL)","date":"2019-11-08","domains":["Vision"],"country":"Switzerland","reference":"https://arxiv.org/abs/1911.03584","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:671249e8e678187f3619","name":"XLM-RoBERTa","organization":"Facebook AI","date":"2019-11-05","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1911.02116","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:aeaf33e79c13021c5f7f","name":"Base LM + kNN LM + Continuous Cache","organization":"Stanford University,Facebook AI Research","date":"2019-11-01","domains":["Language"],"country":"United States of America,United States of America,France","reference":"https://arxiv.org/abs/1911.00172","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:1c2915c799fcef853d6d","name":"AlphaStar","organization":"DeepMind","date":"2019-10-30","domains":["Games"],"country":"United Kingdom","reference":"https://www.deepmind.com/blog/alphastar-grandmaster-level-in-starcraft-ii-using-multi-agent-reinforcement-learning","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:def1ba2c0fbac3c7206b","name":"BART-large","organization":"Facebook AI","date":"2019-10-29","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1910.13461","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:48753f7176cd2bde9713","name":"Amended-DARTS","organization":"Tsinghua University,Huawei,Tongji University","date":"2019-10-25","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/abs/1910.11831","confidence":"","accessibility":"Unreleased"},{"id":"epoch:f12d15a009cc5690cbec","name":"T5-3B","organization":"Google","date":"2019-10-23","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1910.10683","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:53aafc0472ac0c16be86","name":"T5-11B","organization":"Google","date":"2019-10-23","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1910.10683","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:717c5c7253eaf9cdcfef","name":"LSTM(large)+Sememe+cell","organization":"Tsinghua University,Beijing University of Posts and Telecommunications,Huawei Noah's Ark Lab","date":"2019-10-20","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/abs/1910.08910","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:4c70e322de645f779f8e","name":"LSTM(medium)+Sememe+cell (WT2)","organization":"Tsinghua University,Beijing University of Posts and Telecommunications,Huawei Noah's Ark Lab","date":"2019-10-20","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/abs/1910.08910","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b32ebc483e1a9bdb709e","name":"RMSNorm (Transformer-base)","organization":"University of Edinburgh,University of Zurich","date":"2019-10-16","domains":["Language"],"country":"United Kingdom,Switzerland","reference":"https://arxiv.org/abs/1910.07467","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b96bef798a6176aa0438","name":"Rubik's cube ADR robot","organization":"OpenAI","date":"2019-10-15","domains":["Robotics"],"country":"United States of America","reference":"https://arxiv.org/abs/1910.07113","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:840a8ec5fb265ccfcbf6","name":"dense-IndRNN+dynamic eval","organization":"Shandong University,University of Wollongong","date":"2019-10-11","domains":["Language"],"country":"China,Australia","reference":"https://arxiv.org/abs/1910.06251","confidence":"","accessibility":"Unreleased"},{"id":"epoch:0b17379045f74ec3a840","name":"M4-50B","organization":"Google","date":"2019-10-11","domains":["Language"],"country":"United States of America","reference":"https://blog.research.google/2019/10/exploring-massively-multilingual.html","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c3226e251dcdc474d064","name":"AlphaX-1","organization":"Facebook AI Research,Brown University","date":"2019-10-02","domains":["Vision"],"country":"United States of America,France,United States of America","reference":"https://arxiv.org/abs/1903.11059","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:d715f0c0642bc24b9700","name":"DistilBERT","organization":"Hugging Face","date":"2019-10-02","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1910.01108","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4ebbd7cb62bb5d3d02e7","name":"ALBERT","organization":"Toyota Technological Institute at Chicago,Google Research","date":"2019-09-26","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1909.11942","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:dfc344e8f467dccc592f","name":"Adaptive Inputs + LayerDrop","organization":"Facebook AI Research,LORIA","date":"2019-09-25","domains":["Language"],"country":"United States of America,France,France","reference":"https://arxiv.org/abs/1909.11556","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:05a88684c49f7589e378","name":"Transformer-XL+AdamGapAware(GA)","organization":"Technion - Israel Institute of Technology","date":"2019-09-24","domains":["Language"],"country":"Israel","reference":"https://arxiv.org/abs/1909.10802","confidence":"","accessibility":"Unreleased"},{"id":"epoch:6ff8f101f373a7164544","name":"Alleviated TOI 10 (WT103)","organization":"Ecole Polytechnique F\u00b4ed\u00b4erale de Lausanne (EPFL),Swisscom,University of Freiburg","date":"2019-09-18","domains":["Language"],"country":"Switzerland,Switzerland,Germany","reference":"https://arxiv.org/abs/1909.08700","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:ce5f0c55d4a5a4e2ab1c","name":"Alleviated TOI 10 (PTB)","organization":"Ecole Polytechnique F\u00b4ed\u00b4erale de Lausanne (EPFL),Swisscom,University of Freiburg","date":"2019-09-18","domains":["Language"],"country":"Switzerland,Switzerland,Germany","reference":"https://arxiv.org/abs/1909.08700","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:be6c6c2d25855a83e174","name":"Alleviated TOI 10 (WT2)","organization":"Ecole Polytechnique F\u00b4ed\u00b4erale de Lausanne (EPFL),Swisscom,University of Freiburg","date":"2019-09-18","domains":["Language"],"country":"Switzerland,Switzerland,Germany","reference":"https://arxiv.org/abs/1909.08700","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0c0ca030c044e2148250","name":"Megatron-BERT","organization":"Nvidia","date":"2019-09-17","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1909.08053","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4a5307d8fe2144a0105b","name":"Megatron-LM (8.3B)","organization":"Nvidia","date":"2019-09-17","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1909.08053","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:dd5e645fefad36a3bcab","name":"Hide and Seek","organization":"OpenAI","date":"2019-09-17","domains":["Games"],"country":"United States of America","reference":"https://arxiv.org/abs/1909.07528","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:5c2da27bd26efe39259f","name":"Megatron-LM (2.5B)","organization":"Nvidia","date":"2019-09-17","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1909.08053","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:4fb19eaf4cde40372fc6","name":"Megatron-LM (355M)","organization":"Nvidia","date":"2019-09-17","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1909.08053","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ccc366489272da74d77e","name":"Megatron-LM (1.2B)","organization":"Nvidia","date":"2019-09-17","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1909.08053","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e98ef0f2ef70bf7d9295","name":"Xiaoice","organization":"Microsoft Research Asia","date":"2019-09-14","domains":["Language","Vision","Multimodal"],"country":"China","reference":"https://arxiv.org/abs/1812.08989","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:13265b11a0b4c2dfc998","name":"ResNet-152 + ObjectNet","organization":"Massachusetts Institute of Technology (MIT)","date":"2019-09-06","domains":["Vision"],"country":"United States of America","reference":"https://papers.nips.cc/paper/2019/file/97af07a14cacba681feacf3012730892-Paper.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e26525fe6325fa26d147","name":"Mogrifier (d2, MoS2, MC) + dynamic eval","organization":"DeepMind,University of Oxford","date":"2019-09-04","domains":["Language"],"country":"United Kingdom,United Kingdom","reference":"https://arxiv.org/abs/1909.01792","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:89c875b9c03d998ba878","name":"Mogrifier (d2, MC) + dynamic eval","organization":"DeepMind","date":"2019-09-04","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1909.01792","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0acbf58f6f307cfb612c","name":"UDSMProt","organization":"Fraunhofer Heinrich Hertz Institute","date":"2019-09-04","domains":["Biology"],"country":"Germany","reference":"https://www.biorxiv.org/content/10.1101/704874v2.full.pdf","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e9e408fab15cc4da10dc","name":"DEQ-Transformer (Medium, Adaptive Embedding)","organization":"Carnegie Mellon University (CMU),Intel Labs","date":"2019-09-03","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1909.01377","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:092d033b319b4bea4168","name":"DEQ-TrellisNet (PTB)","organization":"Carnegie Mellon University (CMU),Intel Labs","date":"2019-09-03","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1909.01377","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b15896e549be2077702b","name":"DEQ-TrellisNet (WT-103)","organization":"Carnegie Mellon University (CMU),Intel Labs","date":"2019-09-03","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1909.01377","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:70798de97710fca59847","name":"LSTM-Large+Behaviorial-Gating","organization":"University of Southern California","date":"2019-08-31","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1909.00107","confidence":"","accessibility":"Unreleased"},{"id":"epoch:c1c0b68ee7eb91c87796","name":"AWD-LSTM+Behaviorial-Gating","organization":"University of Southern California","date":"2019-08-31","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1909.00107","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:6c971294e1a803389eb8","name":"LSTM-Medium+Behaviorial-Gating (PTB)","organization":"University of Southern California","date":"2019-08-31","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1909.00107","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e4009499ca268e62da85","name":"trRosetta","organization":"Nankai University,University of Washington,Tianjin University,Harvard University","date":"2019-08-22","domains":["Biology"],"country":"China,United States of America,China,United States of America","reference":"https://www.pnas.org/doi/10.1073/pnas.1914677117","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3acfccf65c34322e6e76","name":"TripletRes","organization":"University of Michigan","date":"2019-08-13","domains":["Biology"],"country":"United States of America","reference":"https://onlinelibrary.wiley.com/doi/10.1002/prot.25798","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:1379a9c379ec4e50799e","name":"L_UL-seq","organization":"New York University (NYU),Facebook AI Research,CIFAR AI Research","date":"2019-08-12","domains":["Language"],"country":"United States of America,United States of America,France,Canada","reference":"https://arxiv.org/abs/1908.04319","confidence":"","accessibility":"Open weights (non-commercial)"},{"id":"epoch:f8190cfe6d7cc08af523","name":"EN^2AS with performance reward","organization":"Beijing Institute of Technology,University of Technology Sydney,Monash University","date":"2019-07-22","domains":["Language"],"country":"China,Australia,Australia","reference":"https://arxiv.org/abs/1907.09109","confidence":"","accessibility":"Unreleased"},{"id":"epoch:e1a5d13567fce0807b20","name":"LSTM+Adam+Lookahead","organization":"University of Toronto","date":"2019-07-19","domains":["Language"],"country":"Canada","reference":"https://arxiv.org/abs/1907.08610","confidence":"","accessibility":"Unreleased"},{"id":"epoch:bc2a58e66f63374eb2ef","name":"RNN + char4-MS-vec","organization":"NTT Communication Science Laboratories,Tohoku University","date":"2019-07-17","domains":["Language"],"country":"Japan,Japan","reference":"https://arxiv.org/abs/1906.05506","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:eea94664af697ecc56c9","name":"Graph-based Semi-Supervised Learning (GSSL) Model on MNIST","organization":"West Virginia University","date":"2019-07-17","domains":["Vision"],"country":"United States of America","reference":"https://ojs.aaai.org/index.php/AAAI/article/view/4438","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:4632e385fdb3f400daae","name":"RNN + char3-MS-vec","organization":"NTT Communication Science Laboratories,Tohoku University","date":"2019-07-16","domains":["Language"],"country":"Japan,Japan","reference":"https://arxiv.org/abs/1906.05506","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:a793f97d1fb12ec7a541","name":"RNN Baseline","organization":"Massachusetts Institute of Technology (MIT),Rey Juan Carlos University","date":"2019-07-14","domains":["Language"],"country":"United States of America,Spain","reference":"https://ojs.aaai.org/index.php/AAAI/article/view/4437","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:02b7953ad4d097c244ec","name":"R-Transformer","organization":"Michigan State University,TAL Education Group (Xueersi)","date":"2019-07-12","domains":["Language"],"country":"United States of America,China","reference":"https://arxiv.org/abs/1907.05572","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:55ffec07f00c109b2d8a","name":"Pluribus","organization":"Facebook AI Research","date":"2019-07-11","domains":["Games"],"country":"United States of America,France","reference":"https://www.science.org/cms/asset/910714a7-ee2a-486e-9970-42fb893b08d9/pap.pdf","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:b19732d064e172525b7f","name":"BigBiGAN","organization":"Google","date":"2019-07-04","domains":["Vision","Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/1907.02544","confidence":"","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d0318d9db2501ae80a7f","name":"All-attention network + adaptive span","organization":"Facebook AI Research","date":"2019-07-02","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1907.01470","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:675f2555455c55eec8d0","name":"RoBERTa Large","organization":"Facebook,University of Washington","date":"2019-07-01","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1907.11692","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fc8b1b29cc206e16c23c","name":"RoBERTa Base","organization":"Facebook,University of Washington","date":"2019-07-01","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1907.11692","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:addf39e8153114d78007","name":"Tensorized Transformer (257M)","organization":"Tianjin University,Microsoft Research Asia,Beijing Institute of Technology","date":"2019-06-24","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/abs/1906.09777","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0393b36a6e7788879e3e","name":"Tensorized Transformer (OBW)","organization":"Tianjin University,Microsoft Research Asia,Beijing Institute of Technology","date":"2019-06-24","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/abs/1906.09777","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:33e7c4a77686bc94a075","name":"Tensorized Transformer (PTB)","organization":"Tianjin University,Microsoft Research Asia,Beijing Institute of Technology","date":"2019-06-24","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/abs/1906.09777","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:066dec631c8cee4ddca1","name":"Tensorized Transformer (large PTB)","organization":"Tianjin University,Microsoft Research Asia,Beijing Institute of Technology","date":"2019-06-24","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/abs/1906.09777","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:9c7eeb7567cb2077e381","name":"Tensorized Transformer (W103)","organization":"Tianjin University,Microsoft Research Asia,Beijing Institute of Technology","date":"2019-06-24","domains":["Language"],"country":"China,China,China","reference":"https://arxiv.org/abs/1906.09777","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:a4389aa91d52751c3539","name":"Walking Minotaur robot","organization":"University of California (UC) Berkeley,Google Brain","date":"2019-06-19","domains":["Robotics"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1812.11103","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:dbcd65cbb2509c609eb9","name":"TAPE Transformer","organization":"University of California (UC) Berkeley,Covariant,Google,Chan Zuckerberg Initiative","date":"2019-06-19","domains":["Biology"],"country":"United States of America,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/1906.08230","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e36f0547ab76a2a7f039","name":"Stacked-LSTM+PruningNo","organization":"University of Electronic Science and Technology of China,Chinese University of Hong Kong (CUHK),Peng Cheng Laboratory","date":"2019-06-17","domains":["Language"],"country":"China,Hong Kong,China","reference":"https://arxiv.org/abs/1906.06847","confidence":"","accessibility":"Unreleased"},{"id":"epoch:2cbbddea5f24f8ca4910","name":"LaNet-L (CIFAR-10)","organization":"Brown University,Facebook","date":"2019-06-17","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1906.06832","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:1b055cb7be9d0b49a761","name":"PG-SWGAN","organization":"ETH Zurich","date":"2019-06-15","domains":["Image generation"],"country":"Switzerland","reference":"https://openaccess.thecvf.com/content_CVPR_2019/html/Wu_Sliced_Wasserstein_Generative_Models_CVPR_2019_paper.html","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:d9c9cc59325dc87bafeb","name":"FixRes ResNeXt-101 WSL","organization":"Facebook AI","date":"2019-06-14","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1906.06423","confidence":"","accessibility":"Open weights (non-commercial)"},{"id":"epoch:150c06c008495f72bdd7","name":"SAGAN","organization":"Rutgers University,Google Research","date":"2019-06-14","domains":["Image generation"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1805.08318","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:6d379167d519bb1170f3","name":"Char-CNN-BiLSTM","organization":"Capital One","date":"2019-06-13","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1906.05678","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:5b6024fd6928f9a18d42","name":"RNN + char2-MS-vec","organization":"NTT Communication Science Laboratories,Tohoku University","date":"2019-06-13","domains":["Language"],"country":"Japan,Japan","reference":"https://arxiv.org/abs/1906.05506","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:ab27e10192324fd67b15","name":"AWD-LSTM + MoS + Partial Shuffled","organization":"University of Texas at Austin","date":"2019-06-10","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1906.03805","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:41974eb6bef6940e9b04","name":"AdvSoft + 4 layer QRNN + dynamic evaluation (WT103)","organization":"University of Texas at Austin","date":"2019-06-10","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1906.03805","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b601c4070af81ef0be40","name":"Adversarial + AWD-LSTM-MoS + partial shuffled","organization":"University of Texas at Austin","date":"2019-06-10","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1906.03805","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:bbf6621e0092f4be2618","name":"4 layer QRNN + dynamic evaluation","organization":"University of Texas at Austin","date":"2019-06-10","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1906.03805","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:1e191bff0766f5e77356","name":"Transformer-XL Large + Phrase Induction","organization":"Massachusetts Institute of Technology (MIT),University of Illinois Urbana-Champaign (UIUC)","date":"2019-06-04","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1906.01702","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:8a2bd90c36f69a738d2e","name":"AWD-LSTM + Phrase Induction + finetuning (PTB)","organization":"Massachusetts Institute of Technology (MIT),University of Illinois Urbana-Champaign (UIUC)","date":"2019-06-04","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1906.01702","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:8d50a42f2fa650da188b","name":"AMDIM","organization":"Microsoft Research","date":"2019-06-03","domains":["Vision","Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/1906.00910","confidence":"","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9142db6efefe9bafa405","name":"VQ-VAE-2 (ImageNet)","organization":"Google","date":"2019-06-02","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/1906.00446","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:03bedd615fac278f3a77","name":"VQ-VAE-2 (FFHQ)","organization":"Google","date":"2019-06-02","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/1906.00446","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:92015e04135f264a9885","name":"XLNet","organization":"Carnegie Mellon University (CMU),Google Brain","date":"2019-06-01","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1906.08237","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:279825751ae21b461c51","name":"XLM","organization":"Facebook","date":"2019-06-01","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1901.07291","confidence":"","accessibility":"Open weights (non-commercial)"},{"id":"epoch:56ac27f639e1d7295c4d","name":"DLRM-2020","organization":"Facebook AI","date":"2019-05-31","domains":["Recommendation"],"country":"United States of America","reference":"https://arxiv.org/abs/1906.00091","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:288ec19ec05d5c1d3561","name":"Grover-Mega","organization":"University of Washington","date":"2019-05-29","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1905.12616","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d2d1acf929cf1d6f0038","name":"MnasNet-A3","organization":"Google","date":"2019-05-29","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1807.11626","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8a89a1f327621a223014","name":"MnasNet-A1 + SSDLite","organization":"Google","date":"2019-05-29","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1807.11626","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:318fe8e31b034c02e886","name":"EfficientNet-L2","organization":"Google","date":"2019-05-28","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1905.11946","confidence":"","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7be0818bf96b583400d5","name":"RSM","organization":"Cerenaut","date":"2019-05-28","domains":["Language"],"country":"Australia","reference":"https://arxiv.org/abs/1905.11589","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:77f17256b5fb03328c7a","name":"EfficientNet-B1","organization":"Google","date":"2019-05-28","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1905.11946","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:75f6200949392446f10f","name":"CPC v2","organization":"DeepMind,University of California (UC) Berkeley","date":"2019-05-22","domains":["Vision"],"country":"United Kingdom,United States of America","reference":"https://arxiv.org/abs/1905.09272","confidence":"","accessibility":"Unreleased"},{"id":"epoch:92ef7dc0666827029445","name":"Flow++ (CIFAR10)","organization":"University of California (UC) Berkeley,Covariant","date":"2019-05-15","domains":["Image generation"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1902.00275","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4f658c98c83b89d4cc97","name":"AWD-LSTM-DRILL + dynamic evaluation\u2020 (WT2)","organization":"IDIAP","date":"2019-05-14","domains":["Language"],"country":"Switzerland","reference":"https://arxiv.org/abs/1905.05513","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:be45e03d0275e91ea09f","name":"AWD-LSTM-DRILL + dynamic evaluation\u2020 (PTB)","organization":"IDIAP","date":"2019-05-14","domains":["Language"],"country":"Switzerland","reference":"https://arxiv.org/abs/1905.05513","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3bd61a6cccb2a0d2b9cb","name":"ResNet-50 Billion-scale","organization":"Facebook AI","date":"2019-05-02","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1905.00546","confidence":"","accessibility":"Open weights (non-commercial)"},{"id":"epoch:be7f5f657d6cf1de8aed","name":"ResNeXt-101 Billion-scale","organization":"Facebook AI","date":"2019-05-02","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1905.00546","confidence":"","accessibility":"Open weights (non-commercial)"},{"id":"epoch:af8bf9d4bbdf8cec2662","name":"RaptorX-Contact","organization":"Toyota Technological Institute at Chicago","date":"2019-05-02","domains":["Biology"],"country":"United States of America","reference":"https://www.biorxiv.org/content/biorxiv/early/2019/05/02/624460.full.pdf","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:6ea528349d30082019d6","name":"Neuro-Symbolic Concept Learner","organization":"Massachusetts Institute of Technology (MIT),Tsinghua University,MIT-IBM Watson AI Lab,DeepMind","date":"2019-04-26","domains":["Vision","Language"],"country":"United States of America,China,United States of America,United Kingdom","reference":"https://arxiv.org/abs/1904.12584","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:91fb13c80ad90ae4b078","name":"MuseNet","organization":"OpenAI","date":"2019-04-25","domains":["Audio"],"country":"United States of America","reference":"https://openai.com/index/musenet/","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:6c6f54bc7704caceab2d","name":"Sparse Transformer (Enwik8)","organization":"OpenAI","date":"2019-04-23","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1904.10509","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:94cdc0f95206c400c3f8","name":"Sparse Transformer (CIFAR10)","organization":"OpenAI","date":"2019-04-23","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/1904.10509","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:9eb0fff4d4b6c6cd3f1a","name":"Sparse Transformer (ImageNet)","organization":"OpenAI","date":"2019-04-23","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/1904.10509","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:39a8e40beac414353f73","name":"DANet","organization":"Chinese Academy of Sciences","date":"2019-04-21","domains":["Vision"],"country":"China","reference":"https://openaccess.thecvf.com/content_CVPR_2019/html/Fu_Dual_Attention_Network_for_Scene_Segmentation_CVPR_2019_paper.html","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e9e6b3f2ea05e5fff256","name":"BERT-Large-CAS (PTB+WT2+WT103)","organization":"Amazon","date":"2019-04-20","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1904.09408","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:85c44a488827e44af264","name":"BERT-Large-CAS (WT103)","organization":"Amazon","date":"2019-04-20","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1904.09408","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f670ba6db0b1a3afb34f","name":"BERT-Large-CAS (WT2)","organization":"Amazon","date":"2019-04-20","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1904.09408","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:04819609ccbd4e5428e6","name":"SpecAugment","organization":"Google Brain","date":"2019-04-18","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1904.08779","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:1ff2d79481a341aaa8e5","name":"LTM","organization":"Murdoch University","date":"2019-04-18","domains":["Language"],"country":"Australia","reference":"https://arxiv.org/abs/1904.08936","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:e3375e8b9c83fefca8e0","name":"Transformer-XL + RMS dynamic eval","organization":"University of Edinburgh","date":"2019-04-17","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1904.08378","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:dd9c4dca549c91ffc430","name":"MEGNet (molecule model)","organization":"University of California San Diego","date":"2019-04-10","domains":["Materials science"],"country":"United States of America","reference":"https://arxiv.org/abs/1812.05055","confidence":"Speculative","accessibility":""},{"id":"epoch:2359bb090e1baa8e13fa","name":"MEGNet (crystal formation energy model)","organization":"University of California San Diego","date":"2019-04-10","domains":["Materials science"],"country":"United States of America","reference":"https://arxiv.org/abs/1812.05055","confidence":"Likely","accessibility":""},{"id":"epoch:f5f1a98e2a06545edab5","name":"MEGNet (crystal band gap model)","organization":"University of California San Diego","date":"2019-04-10","domains":["Materials science"],"country":"United States of America","reference":"https://arxiv.org/abs/1812.05055","confidence":"Likely","accessibility":""},{"id":"epoch:6c617201b59607fd91cb","name":"MEGNet (crystal elasticity model)","organization":"University of California San Diego","date":"2019-04-10","domains":["Materials science"],"country":"United States of America","reference":"https://arxiv.org/abs/1812.05055","confidence":"Likely","accessibility":""},{"id":"epoch:204af3f99cd39e970dde","name":"True-Regularization+Finetune","organization":"Mobvoi,Williams College","date":"2019-04-08","domains":["Language"],"country":"China,United States of America","reference":"https://arxiv.org/abs/1904.04163","confidence":"","accessibility":"Unreleased"},{"id":"epoch:b8569468a7e09ed559a9","name":"True-Regularization+Finetune+Dynamic-Eval","organization":"Mobvoi,Williams College","date":"2019-04-08","domains":["Language"],"country":"China,United States of America","reference":"https://arxiv.org/abs/1904.04163","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:de95a39f9bcf91abb46d","name":"WeNet (WT2)","organization":"Amazon","date":"2019-04-08","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1904.03819","confidence":"","accessibility":"Unreleased"},{"id":"epoch:3a35ee05b042451ca367","name":"WeNet (PTB)","organization":"Amazon","date":"2019-04-08","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/pdf/1904.03819","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:76e344630fcf88388b82","name":"WeNet (Penn Treebank)","organization":"Amazon","date":"2019-04-08","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1904.03819","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:892cbfb37acb3b58e796","name":"Cross-lingual alignment","organization":"Tel Aviv University,Massachusetts Institute of Technology (MIT)","date":"2019-04-04","domains":["Language"],"country":"Israel,United States of America","reference":"https://arxiv.org/abs/1902.09492","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5ff85e8799ed9b40092f","name":"FAIRSEQ Adaptive Inputs","organization":"Facebook AI Research,Google Brain","date":"2019-04-01","domains":["Language"],"country":"United States of America,France,United States of America","reference":"https://arxiv.org/abs/1904.01038","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:9cfd9103b71194d01478","name":"UniRep","organization":"Harvard University","date":"2019-03-26","domains":["Biology"],"country":"United States of America","reference":"https://www.nature.com/articles/s41592-019-0598-1","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:d7fbd90c0c014df3ee7e","name":"SciBERT","organization":"Allen Institute for AI","date":"2019-03-26","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1903.10676","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:a3798f0e4fa23e91d27e","name":"DOC + Finetune\u2217 + Partial Shuffle (WT2)","organization":"University of Washington","date":"2019-03-11","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1903.04167","confidence":"","accessibility":"Unreleased"},{"id":"epoch:59f6a7146b7491f1e12b","name":"DOC + Finetune\u2217 + Partial Shuffle (PTB)","organization":"University of Washington","date":"2019-03-11","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1903.04167","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:ec32c3e1f97cea6a8deb","name":"NMT Transformer 437M","organization":"Google,Bar-Ilan University","date":"2019-02-28","domains":["Language"],"country":"United States of America,Israel","reference":"https://arxiv.org/abs/1903.00089","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:7d6ab42a2522f3e310d5","name":"KataGo","organization":"Jane Street","date":"2019-02-27","domains":["Games"],"country":"United States of America","reference":"https://arxiv.org/abs/1902.10565","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d88d8e35c7b73fb3e2ad","name":"ProxylessNAS","organization":"Massachusetts Institute of Technology (MIT)","date":"2019-02-23","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1812.00332","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:078ad6f93adfc618b868","name":"SSA","organization":"Massachusetts Institute of Technology (MIT)","date":"2019-02-22","domains":["Biology"],"country":"United States of America","reference":"https://arxiv.org/abs/1902.08661","confidence":"Confident","accessibility":""},{"id":"epoch:9bc2627fa0719e0c751c","name":"code2seq","organization":"Technion - Israel Institute of Technology,Facebook AI Research","date":"2019-02-21","domains":["Language"],"country":"Israel,United States of America,France","reference":"https://arxiv.org/abs/1808.01400","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0baa73425d4fddd30a0f","name":"GPT-2 (1.5B)","organization":"OpenAI","date":"2019-02-14","domains":["Language"],"country":"United States of America","reference":"https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f9a68d50a50e57e258ae","name":"GPT-2 (124M)","organization":"OpenAI","date":"2019-02-14","domains":["Language"],"country":"United States of America","reference":"https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b443e765a93ab5bb8a2f","name":"GPT-2 (774M)","organization":"OpenAI","date":"2019-02-14","domains":["Language"],"country":"United States of America","reference":"https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1113d7b2f89db6a1de24","name":"GPT-2 (355M)","organization":"OpenAI","date":"2019-02-14","domains":["Language"],"country":"United States of America","reference":"https://openai.com/blog/better-language-models/","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:56b1ae8f879b48b50987","name":"SDE","organization":"Carnegie Mellon University (CMU),Google Brain,Monash University","date":"2019-02-09","domains":["Language"],"country":"United States of America,United States of America,Australia","reference":"https://arxiv.org/abs/1902.03499","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:325b004652165553cef7","name":"Compress-LSTM (66M)","organization":"Samsung R&D Institute Russia,National Research University Higher School of Economics","date":"2019-02-06","domains":["Language"],"country":"Russia,Russia","reference":"https://arxiv.org/abs/1902.02380","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:1600eb3e1f047d4e3bfb","name":"Compress-LSTM (4.6M)","organization":"Samsung R&D Institute Russia,National Research University Higher School of Economics","date":"2019-02-06","domains":["Language"],"country":"Russia,Russia","reference":"https://arxiv.org/abs/1902.02380","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:95e2c97de83e46d9a2f9","name":"Hanabi 4 player","organization":"DeepMind,University of Oxford,Carnegie Mellon University (CMU),Google Brain","date":"2019-02-01","domains":["Games"],"country":"United Kingdom,United Kingdom,United States of America,United States of America","reference":"https://arxiv.org/abs/1902.00506","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:7256c4a6c5db9393c89c","name":"MT-DNN","organization":"Microsoft","date":"2019-01-31","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1901.11504","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:fa7292243c57fabba467","name":"TSLM+MoS (WT2)","organization":"Tianjin University,Beijing Institute of Technology","date":"2019-01-31","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/1901.11167","confidence":"","accessibility":"Unreleased"},{"id":"epoch:2df0ae1d29279e534652","name":"TSLM+MoS (PTB)","organization":"Tianjin University,Beijing Institute of Technology","date":"2019-01-31","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/1901.11167","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:57d23c9feb4aee015624","name":"H-LSTM+wg+rcp+rcg+wp","organization":"Princeton University,Alibaba","date":"2019-01-30","domains":["Language"],"country":"United States of America,China","reference":"https://arxiv.org/abs/1901.10997","confidence":"","accessibility":"Unreleased"},{"id":"epoch:370afe0f5593ecbaf096","name":"Mono3D++","organization":"University of California Los Angeles (UCLA),Megvii Inc","date":"2019-01-11","domains":["3D modeling","Vision"],"country":"United States of America,China","reference":"https://arxiv.org/abs/1901.03446","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:07a3316be03c0c6f3bf6","name":"Transformer-XL (257M)","organization":"Carnegie Mellon University (CMU),Google Brain","date":"2019-01-09","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1901.02860","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ed61fb40182c1dbe072f","name":"Transformer-XL-ptb","organization":"Carnegie Mellon University (CMU),Google Brain","date":"2019-01-09","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1901.02860","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:5291b19abf5d1c799c79","name":"Decoupled weight decay regularization","organization":"University of Freiburg","date":"2019-01-04","domains":["Vision"],"country":"Germany","reference":"https://arxiv.org/abs/1711.05101","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3ca618f198550e6cb130","name":"Transformer ELMo","organization":"Allen Institute for AI,University of Washington","date":"2019-01-01","domains":["Language"],"country":"United States of America,United States of America","reference":"https://www.semanticscholar.org/paper/Dissecting-Contextual-Word-Embeddings%3A-Architecture-Peters-Neumann/ac11062f1f368d97f4c826c317bf50dcc13fdb59","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e52831229c3377d908c8","name":"Transformer + Average Attention Network","organization":"University of Electronic Science and Technology of China","date":"2019-01-01","domains":["Language"],"country":"China","reference":"https://ieeexplore.ieee.org/abstract/document/9067534","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:5a24a0ae5dc7713574a7","name":"StyleGAN","organization":"Nvidia","date":"2018-12-12","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/1812.04948","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:4aaa5638046d081129f5","name":"Vine copula (breast cancer)","organization":"Massachusetts Institute of Technology (MIT),Rey Juan Carlos University","date":"2018-12-04","domains":["Vision"],"country":"United States of America,Spain","reference":"https://arxiv.org/abs/1812.01226","confidence":"Speculative","accessibility":""},{"id":"epoch:7158ddac4b9193f81f29","name":"Vine copula (wine quality)","organization":"Massachusetts Institute of Technology (MIT),Rey Juan Carlos University","date":"2018-12-04","domains":[],"country":"United States of America,Spain","reference":"https://arxiv.org/abs/1812.01226","confidence":"Speculative","accessibility":""},{"id":"epoch:66c01ce223c7c60e435b","name":"Vine copula (crime)","organization":"Massachusetts Institute of Technology (MIT),Rey Juan Carlos University","date":"2018-12-04","domains":[],"country":"United States of America,Spain","reference":"https://arxiv.org/abs/1812.01226","confidence":"Speculative","accessibility":""},{"id":"epoch:745d89d32b70cdc2039b","name":"SPN (ImageNet 128)","organization":"Google Brain,DeepMind","date":"2018-12-04","domains":["Image generation"],"country":"United States of America,United Kingdom","reference":"https://arxiv.org/abs/1812.01608","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0fc70b2f5afdeecb2b0c","name":"SPN (CelebA HQ)","organization":"Google Brain,DeepMind","date":"2018-12-04","domains":["Image generation"],"country":"United States of America,United Kingdom","reference":"https://arxiv.org/abs/1812.01608","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:99344e2ad3bdf774dcef","name":"DMPFold","organization":"University College London (UCL)","date":"2018-11-29","domains":["Biology"],"country":"United Kingdom","reference":"https://www.nature.com/articles/s41467-019-11994-0","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f6b995264d131449d039","name":"GPipe (Transformer)","organization":"Google","date":"2018-11-16","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1811.06965","confidence":"Likely","accessibility":""},{"id":"epoch:3bf54514c0cc556cab72","name":"GPipe (Amoeba)","organization":"Google","date":"2018-11-16","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1811.06965","confidence":"","accessibility":""},{"id":"epoch:b5dfc1b5c6658e760f2e","name":"Multi-cell LSTM","organization":"University of Hyderabad","date":"2018-11-15","domains":["Language"],"country":"India","reference":"https://arxiv.org/abs/1811.06477","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:393f15bece7e879056c8","name":"Fine-tuned-AWD-LSTM-DOC (fin)","organization":"Samsung R&D Institute Russia","date":"2018-11-12","domains":["Language"],"country":"Russia","reference":"https://arxiv.org/abs/1811.04623","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:da25d29fd6d9af942cc1","name":"Discriminator-tuned LSTM","organization":"Samsung R&D Institute Russia","date":"2018-11-12","domains":["Language"],"country":"Russia","reference":"https://arxiv.org/abs/1811.04623","confidence":"Likely","accessibility":""},{"id":"epoch:b12a7e9738aa9b06cde5","name":"Mesh-TensorFlow Transformer 4.9B (language)","organization":"Google Brain","date":"2018-11-05","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1811.02084","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:7b605384d74798d07b0f","name":"Mesh-TensorFlow Transformer 2.9B (translation)","organization":"Google Brain","date":"2018-11-05","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1811.02084","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:2c37d0f3a41eec505b97","name":"MemoReader","organization":"Samsung,Korea University","date":"2018-10-31","domains":["Language"],"country":"South Korea,South Korea","reference":"https://aclanthology.org/D18-1237/","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:62d20023eadf865885b5","name":"code2vec","organization":"Technion - Israel Institute of Technology,Facebook AI Research","date":"2018-10-30","domains":["Language"],"country":"Israel,United States of America,France","reference":"https://arxiv.org/abs/1803.09473","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:49f335358b525e0a719c","name":"TrellisNet","organization":"Carnegie Mellon University (CMU),Bosch Center for Artificial Intelligence,Intel Labs","date":"2018-10-15","domains":["Language"],"country":"United States of America,Germany,United States of America","reference":"https://arxiv.org/abs/1810.06682","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:62215a085149538943f3","name":"TrellisNet-MoS (1.4x larger) PTB","organization":"Carnegie Mellon University (CMU),Intel Labs,Bosch Center for Artificial Intelligence","date":"2018-10-15","domains":["Language"],"country":"United States of America,United States of America,Germany","reference":"https://arxiv.org/abs/1810.06682","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:60f527da230e7bfe732e","name":"BERT-Large","organization":"Google","date":"2018-10-11","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1810.04805","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2a5a4d0836fdb16353a4","name":"MetaMimic","organization":"Google","date":"2018-10-11","domains":["Robotics"],"country":"United States of America","reference":"https://arxiv.org/abs/1810.05017","confidence":"","accessibility":""},{"id":"epoch:97ea7d00b75dd244658a","name":"DeepConPred2","organization":"Tsinghua University","date":"2018-10-11","domains":["Biology"],"country":"China","reference":"https://www.sciencedirect.com/science/article/pii/S2001037018301466","confidence":"Confident","accessibility":""},{"id":"epoch:a96353b904c84531c372","name":"BigGAN-deep 512x512","organization":"Heriot-Watt University,DeepMind","date":"2018-09-28","domains":["Image generation"],"country":"United Kingdom,United Kingdom","reference":"https://arxiv.org/abs/1809.11096","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:af5609536d65aba7b296","name":"Transformer (Adaptive Input Embeddings) WT103","organization":"Facebook AI Research","date":"2018-09-28","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1809.10853","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:dcd9367a3bdbe472bc36","name":"LSTM+NeuralCache","organization":"KU Leuven,ESAT - PSI,Apple","date":"2018-09-24","domains":["Language"],"country":"Belgium,Belgium,United States of America","reference":"https://arxiv.org/abs/1809.08826","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:2c827351cc7195ffa0e6","name":"AWD-LSTM-MoS + dynamic evaluation (WT2, 2018)","organization":"Peking University,Microsoft Research Asia","date":"2018-09-18","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/1809.06858","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:48a543818b34482312a8","name":"AWD-LSTM-MoS + dynamic evaluation (PTB, 2018)","organization":"Peking University,Microsoft Research Asia","date":"2018-09-18","domains":["Language"],"country":"China,China","reference":"https://arxiv.org/abs/1809.06858","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:cf703690894ccb98df53","name":"Talent Search and Recommendation Systems","organization":"LinkedIn","date":"2018-09-18","domains":["Recommendation"],"country":"United States of America","reference":"https://arxiv.org/abs/1809.06481","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:0667dc9e112246957a52","name":"Transformer + Simple Recurrent Unit","organization":"ASAPP,Cornell University,Google,Princeton University","date":"2018-09-17","domains":["Language"],"country":"United States of America,United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/1709.02755v5","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2b0e1b7d5423bbd308c6","name":"NetSurfP-2.0","organization":"Technical University of Denmark,University of Copenhagen,Universidad Nacional de San Mart\u00edn,AIMST University","date":"2018-09-10","domains":["Biology"],"country":"Denmark,Denmark,Argentina,Malaysia","reference":"https://www.biorxiv.org/content/10.1101/311209v3.full","confidence":"Unknown","accessibility":"Hosted access (no API)"},{"id":"epoch:b38fe9cf0c3a3b4c62c5","name":"ESRGAN","organization":"Chinese University of Hong Kong (CUHK),Chinese Academy of Sciences,Nanyang Technological University","date":"2018-09-01","domains":["Vision","Image generation"],"country":"Hong Kong,China,Singapore","reference":"https://arxiv.org/abs/1809.00219","confidence":"Unknown","accessibility":""},{"id":"epoch:452aede1a13c8c423038","name":"(ensemble): AWD-LSTM-DOC (fin) \u00d7 5 (WT2)","organization":"NTT Communication Science Laboratories,Tohoku University","date":"2018-08-30","domains":["Language"],"country":"Japan,Japan","reference":"https://arxiv.org/abs/1808.10143","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:b4733c34f5e1596fee4f","name":"(ensemble): AWD-LSTM-DOC (fin) \u00d7 5 (PTB)","organization":"NTT Communication Science Laboratories,Tohoku University","date":"2018-08-30","domains":["Language"],"country":"Japan,Japan","reference":"https://arxiv.org/abs/1808.10143","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:59ef49dbedabff68de67","name":"AWD-LSTM-DOC (fin) (37M)","organization":"NTT Communication Science Laboratories,Tohoku University","date":"2018-08-30","domains":["Language"],"country":"Japan,Japan","reference":"https://arxiv.org/abs/1808.10143","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:5a6be7d1281c06050c8e","name":"AWD-LSTM-DOC (fin) (23M)","organization":"NTT Communication Science Laboratories,Tohoku University","date":"2018-08-30","domains":["Language"],"country":"Japan,Japan","reference":"https://arxiv.org/abs/1808.10143","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6789e25c244233b18a61","name":"Big Transformer for Back-Translation","organization":"Facebook AI Research,Google Brain","date":"2018-08-28","domains":["Language"],"country":"United States of America,France,United States of America","reference":"https://arxiv.org/abs/1808.09381","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:9516e3fd2f2e20a21991","name":"AWD-LSTM-MoS+PDR + dynamic evaluation (WT2)","organization":"IBM","date":"2018-08-14","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1808.05908","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:069e0079bb118ca4de67","name":"AWD-LSTM-MoS+PDR + dynamic evaluation (PTB)","organization":"IBM","date":"2018-08-14","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1808.05908","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:3b3cb27b7e9e853c29a7","name":"RGC+ASQ (WT2)","organization":"Tsinghua University,University of California Los Angeles (UCLA)","date":"2018-08-13","domains":["Language"],"country":"China,United States of America","reference":"https://arxiv.org/abs/1808.04357","confidence":"","accessibility":"Unreleased"},{"id":"epoch:b2c66595b10027757a77","name":"RGC+ASQ (PTB)","organization":"Tsinghua University,University of California Los Angeles (UCLA)","date":"2018-08-13","domains":["Language"],"country":"China,United States of America","reference":"https://arxiv.org/abs/1808.04357","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:d7f7d6308943a2a4f48e","name":"Dexterous In-Hand Manipulation [control policy]","organization":"OpenAI","date":"2018-08-01","domains":["Robotics"],"country":"United States of America","reference":"http://arxiv.org/abs/1808.00177\nhttps://journals.sagepub.com/doi/full/10.1177/0278364919887447\nhttps://openai.com/index/learning-dexterity/","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:d2156ebd2833f044142b","name":"Big-Little Net","organization":"IBM","date":"2018-07-10","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1807.03848","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3ebff45ed346fbdbf0ad","name":"Big-Little Net (vision)","organization":"IBM","date":"2018-07-10","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1807.03848","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3ae59e24ef8be1b5607b","name":"Big-Little Net (speech)","organization":"IBM","date":"2018-07-10","domains":["Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/1807.03848","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:f133794108a1ea4567bc","name":"Glow (Celeba HQ)","organization":"OpenAI","date":"2018-07-10","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/1807.03039","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:d43ed604f5e9d91922d7","name":"RCAN","organization":"Northeastern University","date":"2018-07-08","domains":["Image generation","Vision"],"country":"United States of America","reference":"https://openaccess.thecvf.com/content_ECCV_2018/html/Yulun_Zhang_Image_Super-Resolution_Using_ECCV_2018_paper.html","confidence":"Unknown","accessibility":""},{"id":"epoch:1a0da3a6322109d8f380","name":"FTW (For The Win)","organization":"DeepMind","date":"2018-07-03","domains":["Games"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1807.01281","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:2e6ead926dfa8116272c","name":"ShuffleNet v2","organization":"Tsinghua University,Megvii Inc","date":"2018-06-30","domains":["Vision"],"country":"China,China","reference":"https://arxiv.org/abs/1807.11164","confidence":"","accessibility":""},{"id":"epoch:ee9a7d28b9b821e72273","name":"QT-Opt","organization":"Google Brain,University of California (UC) Berkeley","date":"2018-06-27","domains":["Robotics","Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1806.10293","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:f63e800aa13f2b480271","name":"S + I-Attention (3)","organization":"National Research University Higher School of Economics,Samsung R&D Institute Russia","date":"2018-06-26","domains":["Language"],"country":"Russia,Russia","reference":"https://arxiv.org/abs/1806.10090","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:fa3cea664e437835a811","name":"DARTS","organization":"DeepMind,Carnegie Mellon University (CMU)","date":"2018-06-24","domains":["Language"],"country":"United Kingdom,United States of America","reference":"https://arxiv.org/abs/1806.09055","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:117e237bc419baa69894","name":"DARTS (second order) (PTB)","organization":"Carnegie Mellon University (CMU),DeepMind","date":"2018-06-24","domains":["Language"],"country":"United States of America,United Kingdom","reference":"https://arxiv.org/abs/1806.09055","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:282903beaf9dcbc20554","name":"MobileNetV2","organization":"Google","date":"2018-06-18","domains":["Vision"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/8578572","confidence":"","accessibility":""},{"id":"epoch:6c1459c6ac9d8f75d597","name":"Image Transformer","organization":"","date":"2018-06-15","domains":["Vision"],"country":"","reference":"https://arxiv.org/abs/1802.05751, https://proceedings.mlr.press/v80/parmar18a.html","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:7d89298aec1d62f9f281","name":"Relational Memory Core","organization":"DeepMind,University College London (UCL)","date":"2018-06-05","domains":["Language"],"country":"United Kingdom,United Kingdom","reference":"https://arxiv.org/abs/1806.01822","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:faa78fbb2a4eb20edede","name":"GPT-1","organization":"OpenAI","date":"2018-06-01","domains":["Language"],"country":"United States of America","reference":"https://openai.com/blog/language-unsupervised/","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:4cf462861343e282db83","name":"2-layer skip-LSTM + dropout tuning (WT2)","organization":"DeepMind","date":"2018-05-23","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1805.09208","confidence":"","accessibility":"Unreleased"},{"id":"epoch:1f54d933565702f3db3b","name":"RHN+HSG(depth=40)","organization":"Ben-Gurion University","date":"2018-05-23","domains":["Language"],"country":"Israel","reference":"https://arxiv.org/abs/1805.09238","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:0999d8f48b652bbd5360","name":"RHN(depth=40)","organization":"Ben-Gurion University","date":"2018-05-23","domains":["Language"],"country":"Israel","reference":"https://arxiv.org/abs/1805.09238","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:1be045109d641ba31d80","name":"2-layer skip-LSTM + dropout tuning (PTB)","organization":"DeepMind","date":"2018-05-23","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1805.09208","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f268c684c4057dbb7097","name":"aLSTM(depth-2)+RecurrentPolicy (WT2)","organization":"University of Manchester,Alan Turing Institute","date":"2018-05-22","domains":["Language"],"country":"United Kingdom,United Kingdom","reference":"https://arxiv.org/abs/1805.08574","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b0b898486ff8a0768f11","name":"aLSTM(depth-2)+RecurrentPolicy (PTB)","organization":"University of Manchester","date":"2018-05-22","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1805.08574","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:910d334a0cad216a84af","name":"Dropout-LSTM+Noise(Bernoulli) (WT2)","organization":"Columbia University,New York University (NYU),Princeton University","date":"2018-05-03","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/1805.01500","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:e8d39d4154c9a36d9c2f","name":"LSTM+Noise(Beta)","organization":"Columbia University,New York University (NYU),Princeton University","date":"2018-05-03","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/1805.01500","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:f7bfa7a17f01e1e19d43","name":"AWD-LSTM-MoS+Noisin+dynamic evaluation (PTB)","organization":"Columbia University,New York University (NYU),Princeton University","date":"2018-05-03","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/1805.01500","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0bd8282d8b03c5aa8f7b","name":"Dropout-LSTM+Noise(Laplace) - medium (WT2)","organization":"Columbia University,New York University (NYU),Princeton University","date":"2018-05-03","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/1805.01500","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f3396e0d03ab2bb203d0","name":"Dropout-LSTM+Noise(Bernoulli) - large(PTB)","organization":"Columbia University,New York University (NYU),Princeton University","date":"2018-05-03","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/1805.01500","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:168073ee02adc16c5f13","name":"ResNeXt-101 32x48d","organization":"Facebook","date":"2018-05-02","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1805.00932","confidence":"Confident","accessibility":"Open weights (non-commercial)"},{"id":"epoch:4aa64a5d844131688765","name":"TF-LM-discourse LSTM (WT2)","organization":"ESAT - PSI","date":"2018-05-01","domains":["Language"],"country":"Belgium","reference":"https://aclanthology.org/L18-1470.pdf","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1942e44e67951e78425b","name":"TF-LM-discourse LSTM (PTB)","organization":"ESAT - PSI","date":"2018-05-01","domains":["Language"],"country":"Belgium","reference":"https://aclanthology.org/L18-1470.pdf","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e27ef69721de55a448fa","name":"DNCON2","organization":"University of Missouri","date":"2018-05-01","domains":["Biology"],"country":"United States of America","reference":"https://academic.oup.com/bioinformatics/article/34/9/1466/4708303?login=false","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:19bc00d153a3a8d41c5d","name":"RNNLM + Dynamic KL Regularization (WT2)","organization":"Northwestern University","date":"2018-04-27","domains":["Language"],"country":"United States of America","reference":"https://ojs.aaai.org/index.php/AAAI/article/view/11993","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:b66fbf855d6d59af522b","name":"RNMT+","organization":"Google AI","date":"2018-04-26","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1804.09849","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:0fd272f1e56e30d85626","name":"Diffractive Deep Neural Network","organization":"University of California Los Angeles (UCLA)","date":"2018-04-14","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1804.08711","confidence":"Likely","accessibility":""},{"id":"epoch:2c3ea1941d194fa36bab","name":"YOLOv3","organization":"University of Washington","date":"2018-04-08","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1804.02767","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:48debf81cbb20d8861ab","name":"LSTM (Hebbian, Cache, MbPA)","organization":"DeepMind,University College London (UCL)","date":"2018-03-27","domains":["Language"],"country":"United Kingdom,United Kingdom","reference":"https://arxiv.org/abs/1803.10049","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:93858c3ee35d306e4714","name":"4 layer QRNN (h=2500)","organization":"Salesforce Research","date":"2018-03-22","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1803.08240","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:8faca20695177af5f8f6","name":"Rotation","organization":"\u00c9cole des Ponts ParisTech","date":"2018-03-21","domains":["Image generation","Vision"],"country":"France","reference":"https://arxiv.org/abs/1803.07728","confidence":"","accessibility":""},{"id":"epoch:879d29558b987e52fc61","name":"LSTM (2018)","organization":"Intel Labs,Carnegie Mellon University (CMU)","date":"2018-03-04","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1803.01271","confidence":"","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7820eeb7aa631d1cd564","name":"Chinese - English translation","organization":"Microsoft","date":"2018-03-01","domains":["Language"],"country":"United States of America","reference":"https://www.microsoft.com/en-us/research/publication/achieving-human-parity-on-automatic-chinese-to-english-news-translation/","confidence":"Unknown","accessibility":""},{"id":"epoch:492ebfa089dfb6f856c8","name":"Residual Dense Network","organization":"Northeastern University,University of Rochester","date":"2018-02-24","domains":["Vision","Image generation"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1802.08797v2","confidence":"Unknown","accessibility":""},{"id":"epoch:dbcc3ecb71a75a8165dd","name":"Spectrally Normalized GAN","organization":"Preferred Networks Inc,Ritsumeikan University,National Institute of Informatics","date":"2018-02-16","domains":["Image generation"],"country":"Japan,Japan,Japan","reference":"https://arxiv.org/abs/1802.05957","confidence":"Unknown","accessibility":""},{"id":"epoch:c9c3a0f9a0aec286e09b","name":"Multipop Adaptive Continuous Stack (WT2)","organization":"DeepMind,University of Oxford","date":"2018-02-15","domains":["Language"],"country":"United Kingdom,United Kingdom","reference":"https://openreview.net/forum?id=SkFqf0lAZ","confidence":"","accessibility":"Unreleased"},{"id":"epoch:d12c484c2b3c46f1a549","name":"TCN (148M)","organization":"Carnegie Mellon University (CMU),Intel Labs","date":"2018-02-15","domains":["Language"],"country":"United States of America,United States of America","reference":"https://openreview.net/forum?id=rk8wKk-R-","confidence":"","accessibility":"Unreleased"},{"id":"epoch:8194a17e6b50071344aa","name":"TCN (13M)","organization":"Carnegie Mellon University (CMU),Intel Labs","date":"2018-02-15","domains":["Language"],"country":"United States of America,United States of America","reference":"https://openreview.net/forum?id=rk8wKk-R-\nhttps://arxiv.org/abs/1803.01271","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:87edd9096cb36a70ee9f","name":"Multipop Adaptive Continuous Stack (PTB)","organization":"DeepMind,University of Oxford","date":"2018-02-15","domains":["Language"],"country":"United Kingdom,United Kingdom","reference":"https://openreview.net/forum?id=SkFqf0lAZ","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b46b42187042fecdfcf7","name":"TCN (P-MNIST)","organization":"Carnegie Mellon University (CMU),Intel Labs","date":"2018-02-15","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://openreview.net/forum?id=rk8wKk-R-","confidence":"Confident","accessibility":""},{"id":"epoch:12e375f0c05fe84297ef","name":"ENAS","organization":"Google Brain,Carnegie Mellon University (CMU),Stanford University","date":"2018-02-09","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/1802.03268","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:ace3f84ec6965fa058aa","name":"DeepLabV3+","organization":"Google","date":"2018-02-07","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1802.02611v3","confidence":"Unknown","accessibility":""},{"id":"epoch:7b8433e86a7553910dbd","name":"AmoebaNet-A (F=448)","organization":"Google Brain","date":"2018-02-05","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1802.01548","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:30fb47e28a139feaf62f","name":"AmoebaNet-A (F=190)","organization":"Google Brain","date":"2018-02-05","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1802.01548","confidence":"","accessibility":""},{"id":"epoch:3301ace86ea82da753fc","name":"IMPALA","organization":"DeepMind","date":"2018-02-05","domains":["Games"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1802.01561","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:7a2c9b65e1e6082b4210","name":"ELMo","organization":"University of Washington,Allen Institute for AI","date":"2018-02-01","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1802.05365","confidence":"Speculative","accessibility":""},{"id":"epoch:b64ee40b30d409bb0736","name":"QRNN","organization":"Salesforce Research","date":"2018-02-01","domains":["Language"],"country":"United States of America","reference":"https://mlsys.org/Conferences/doc/2018/50.pdf","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:9e59cf0959bc6ab6b4a7","name":"T-DMCA","organization":"Google Brain","date":"2018-01-30","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1801.10198","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:e0a764339e9c91920f56","name":"DenseNet201","organization":"Tsinghua University,Facebook AI Research,Cornell University","date":"2018-01-28","domains":["Vision"],"country":"China,United States of America,France,United States of America","reference":"https://arxiv.org/abs/1608.06993","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:306e5d1dd7296c16e462","name":"ULM-FiT","organization":"University of San Francisco,Insight Centre NUI Galway,Fast.ai","date":"2018-01-18","domains":["Language"],"country":"United States of America,Ireland","reference":"https://arxiv.org/abs/1801.06146","confidence":"Speculative","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3803522a6259645af765","name":"Refined Part Pooling","organization":"Tsinghua University,University of Technology Sydney,University of Texas at San Antonio","date":"2018-01-09","domains":["Vision"],"country":"China,Australia,United States of America","reference":"https://arxiv.org/abs/1711.09349","confidence":"Confident","accessibility":""},{"id":"epoch:c9226072a25e187d7d0e","name":"RNNLM + Dynamic KL Regularization","organization":"Northwestern University","date":"2018-01-01","domains":["Language"],"country":"United States of America","reference":"https://ojs.aaai.org/index.php/AAAI/article/view/11993","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:e94f67aa17dcdeaca45d","name":"PixelSNAIL (CIFAR 10)","organization":"University of California (UC) Berkeley","date":"2017-12-28","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/1712.09763","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:875541e5c73866840990","name":"PixelSNAIL (ImageNet)","organization":"University of California (UC) Berkeley","date":"2017-12-28","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/1712.09763","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:0633912b29bcd02d5a6b","name":"Tacotron 2","organization":"Google,University of California (UC) Berkeley","date":"2017-12-19","domains":["Speech"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1712.05884","confidence":"Confident","accessibility":""},{"id":"epoch:974972aac9a50f7fadf7","name":"WGAN (Wasserstein GAN)","organization":"Facebook AI Research,Courant Institute of Mathematical Sciences","date":"2017-12-06","domains":["Image generation"],"country":"United States of America,France,United States of America","reference":"https://arxiv.org/abs/1701.07875","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:280e5d74a3be4eba1b46","name":"AlphaZero","organization":"DeepMind","date":"2017-12-05","domains":["Games"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1712.01815","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:c503f3b8dc463b6a7be4","name":"2-layer-LSTM+Deep-Gradient-Compression","organization":"Tsinghua University,Stanford University,Nvidia","date":"2017-12-05","domains":["Language"],"country":"China,United States of America,United States of America","reference":"https://arxiv.org/abs/1712.01887","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:b895e8a53add1db35ced","name":"PNAS-net","organization":"Johns Hopkins University,Google AI,Stanford University","date":"2017-12-02","domains":["Vision"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/1712.00559","confidence":"","accessibility":""},{"id":"epoch:0da34668eef513f181a1","name":"PNASNet-5","organization":"Johns Hopkins University,Google AI,Stanford University","date":"2017-12-02","domains":["Vision"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/1712.00559","confidence":"Likely","accessibility":""},{"id":"epoch:339bdf10b05ec9e3db73","name":"DL scaling speech","organization":"Baidu","date":"2017-12-01","domains":["Speech"],"country":"China","reference":"https://www.semanticscholar.org/paper/Deep-Learning-Scaling-is-Predictable%2C-Empirically-Hestness-Narang/a1c922be467d1c0c64b963e65dae41778b81b2a0","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:9426d930834ff136c87a","name":"DL scaling LM","organization":"Baidu","date":"2017-12-01","domains":["Language"],"country":"China","reference":"https://www.semanticscholar.org/paper/Deep-Learning-Scaling-is-Predictable%2C-Empirically-Hestness-Narang/a1c922be467d1c0c64b963e65dae41778b81b2a0","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:e69cb506fb34f507e0e3","name":"DL scaling Image","organization":"Baidu","date":"2017-12-01","domains":["Vision"],"country":"China","reference":"https://www.semanticscholar.org/paper/Deep-Learning-Scaling-is-Predictable%2C-Empirically-Hestness-Narang/a1c922be467d1c0c64b963e65dae41778b81b2a0","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:7813c32d869376dcf8c2","name":"TriNet","organization":"Visual Computing Institute,RWTH Aachen University","date":"2017-11-21","domains":["Video"],"country":"Germany,Germany","reference":"https://arxiv.org/abs/1703.07737","confidence":"Unknown","accessibility":""},{"id":"epoch:d4faf5d38eb941a0913e","name":"AWD-LSTM-MoS + dynamic evaluation (WT2, 2017)","organization":"Carnegie Mellon University (CMU)","date":"2017-11-10","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1711.03953","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:af85b1ee1dc91556e88d","name":"AWD-LSTM-MoS + dynamic evaluation (PTB, 2017)","organization":"Carnegie Mellon University (CMU)","date":"2017-11-10","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1711.03953","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:bfab2f60dc66afea36bf","name":"VQ-VAE","organization":"DeepMind","date":"2017-11-02","domains":["Vision"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1711.00937","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:b2713fa96a1e5ec126d4","name":"Fraternal dropout + AWD-LSTM 3-layer (WT2)","organization":"Jagiellonian University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2017-10-31","domains":["Language"],"country":"Poland,Canada,Canada","reference":"https://arxiv.org/abs/1711.00066","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:db4c738d3502e4a45a0f","name":"Fraternal dropout + AWD-LSTM 3-layer (PTB)","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms)","date":"2017-10-31","domains":["Language"],"country":"Canada,Canada","reference":"https://arxiv.org/abs/1711.00066","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:24bdabedfb06c781a716","name":"DCN+","organization":"Salesforce Research","date":"2017-10-31","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1711.00106v2","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e93aef26564e7491f683","name":"S-Norm","organization":"University of Washington,Allen Institute for AI","date":"2017-10-29","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1710.10723v2","confidence":"Confident","accessibility":""},{"id":"epoch:27d94fb8c45029366010","name":"PhraseCond","organization":"Carnegie Mellon University (CMU),University of Pittsburgh","date":"2017-10-28","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1710.10504v2","confidence":"Confident","accessibility":""},{"id":"epoch:31dcb68e7fa04241929d","name":"ProgressiveGAN","organization":"Nvidia","date":"2017-10-27","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1710.10196","confidence":"Unknown","accessibility":""},{"id":"epoch:9c8efb93ef4c8a297930","name":"CapsNet (MultiMNIST)","organization":"Google Brain","date":"2017-10-26","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1710.09829","confidence":"","accessibility":""},{"id":"epoch:19abbcb1fd60955026d3","name":"CapsNet (MNIST)","organization":"Google Brain","date":"2017-10-26","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1710.09829","confidence":"","accessibility":""},{"id":"epoch:5c275ffd7e2620e4cf1a","name":"LRSO-GAN","organization":"University of Technology Sydney","date":"2017-10-22","domains":["Vision"],"country":"Australia","reference":"https://arxiv.org/abs/1701.07717","confidence":"Unknown","accessibility":""},{"id":"epoch:35bfb841cc76af6ebcfc","name":"AlphaGo Master","organization":"DeepMind","date":"2017-10-19","domains":["Games"],"country":"United Kingdom","reference":"https://www.nature.com/articles/nature24270","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:47e1c681ec7a34e252e3","name":"AlphaGo Zero","organization":"DeepMind","date":"2017-10-18","domains":["Games"],"country":"United Kingdom","reference":"https://www.nature.com/articles/nature24270","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:703b4a5404e22f205304","name":"Rainbow DQN","organization":"DeepMind","date":"2017-10-06","domains":["Games"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1710.02298","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:bc7e35c8148b4e49ed00","name":"AWD-LSTM+WT+Cache+IOG (WT2)","organization":"NTT Communication Science Laboratories","date":"2017-09-26","domains":["Language"],"country":"Japan","reference":"https://arxiv.org/abs/1709.08907","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:8f6c45764693b0957eea","name":"AWD-LSTM+WT+Cache+IOG (PTB)","organization":"NTT Communication Science Laboratories","date":"2017-09-26","domains":["Language"],"country":"Japan","reference":"https://arxiv.org/abs/1709.08907","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:bcd5c56fc22d5dce8237","name":"LSTM + dynamic eval","organization":"University of Edinburgh","date":"2017-09-21","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1709.07432","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:f0bdf3563e49c7ccb27d","name":"AWD-LSTM + dynamic eval (PTB)","organization":"University of Edinburgh","date":"2017-09-21","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1709.07432","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:844908f191a18c38ad2b","name":"AWD-LSTM + dynamic eval (WT2)","organization":"University of Edinburgh","date":"2017-09-21","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1709.07432","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:6b02522833e7c3536b1b","name":"ISS","organization":"Duke University,Microsoft","date":"2017-09-15","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1709.05027","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:9c18aa84328dc0a59ca1","name":"PyramidNet","organization":"Korea Advanced Institute of Science and Technology (KAIST)","date":"2017-09-06","domains":["Vision"],"country":"South Korea","reference":"https://arxiv.org/abs/1610.02915v4","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:373adc1c372b146b2bc3","name":"SENet (ImageNet)","organization":"Chinese Academy of Sciences,University of Oxford","date":"2017-09-05","domains":["Vision"],"country":"China,United Kingdom","reference":"https://arxiv.org/abs/1709.01507","confidence":"","accessibility":""},{"id":"epoch:b83c5cdc8055c9e22d52","name":"GL-LWGC-AWD-MoS-LSTM + dynamic evaluation (WT2)","organization":"Ben-Gurion University of the Negev","date":"2017-08-29","domains":["Language"],"country":"Israel","reference":"https://arxiv.org/abs/1708.08863","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f2e13b6d9c47c99d4708","name":"GL-LWGC-AWD-MoS-LSTM + dynamic evaluation (PTB)","organization":"Ben-Gurion University","date":"2017-08-29","domains":["Language"],"country":"Israel","reference":"https://arxiv.org/abs/1708.08863","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:74f6b313fb2e3bb1d1b7","name":"NMM(LSTM+RNN)","organization":"Saarland University","date":"2017-08-23","domains":["Language"],"country":"Germany","reference":"https://arxiv.org/abs/1708.06989","confidence":"","accessibility":"Unreleased"},{"id":"epoch:fd9010cb80a91584dbb4","name":"D-LSRC(200)+KN5","organization":"Saarland University","date":"2017-08-22","domains":["Language"],"country":"Germany","reference":"https://arxiv.org/abs/1708.06555","confidence":"","accessibility":"Unreleased"},{"id":"epoch:9295e3c2640e6d2225d8","name":"D-LSRC(100)+KN5 (PTB)","organization":"Saarland University","date":"2017-08-22","domains":["Language"],"country":"Germany","reference":"https://arxiv.org/abs/1708.06555","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b3909e3db1f9bae91c92","name":"Libratus","organization":"Carnegie Mellon University (CMU)","date":"2017-08-19","domains":["Games"],"country":"United States of America","reference":"https://www.ijcai.org/proceedings/2017/0772.pdf","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:90164fc5b21192f4bea6","name":"GRU + p-tHSM (pretrain via Brown) (WT103)","organization":"Beihang University,University of Montreal / Universit\u00e9 de Montr\u00e9al,Chongqing University","date":"2017-08-19","domains":["Language"],"country":"China,Canada,China","reference":"https://www.ijcai.org/proceedings/2017/271","confidence":"","accessibility":"Unreleased"},{"id":"epoch:b708d8ce51f5e1863ffa","name":"GRU + p-tHSM (pretrain via Brown) (PTB)","organization":"Beihang University,University of Montreal / Universit\u00e9 de Montr\u00e9al,Chongqing University","date":"2017-08-19","domains":["Language"],"country":"China,Canada,China","reference":"https://www.researchgate.net/profile/Yikang-Shen-2/publication/318830618_Exploration_of_Tree-based_Hierarchical_Softmax_for_Recurrent_Language_Models/links/5b2c050aa6fdcc8506bc6f4a/Exploration-of-Tree-based-Hierarchical-Softmax-for-Recurrent-Language-Models.pdf","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:f3138ec0228597400e35","name":"GRU + p-tHSM (pretrain via Brown) (WT2)","organization":"Beihang University,University of Montreal / Universit\u00e9 de Montr\u00e9al,Chongqing University","date":"2017-08-19","domains":["Language"],"country":"China,Canada,China","reference":"https://www.researchgate.net/profile/Yikang-Shen-2/publication/318830618_Exploration_of_Tree-based_Hierarchical_Softmax_for_Recurrent_Language_Models/links/5b2c050aa6fdcc8506bc6f4a/Exploration-of-Tree-based-Hierarchical-Softmax-for-Recurrent-Language-Models.pdf","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:fb7f5b90c86d8ce42b91","name":"Adversarial Joint Adaptation Network (ResNet)","organization":"Tsinghua University,University of California (UC) Berkeley","date":"2017-08-17","domains":["Vision"],"country":"China,United States of America","reference":"https://arxiv.org/abs/1605.06636","confidence":"Speculative","accessibility":""},{"id":"epoch:100f53feb8e1bec872c1","name":"NeuMF (Pinterest)","organization":"Shandong University,Texas A&M,National University of Singapore,Columbia University","date":"2017-08-16","domains":["Recommendation"],"country":"China,United States of America,Singapore,United States of America","reference":"https://arxiv.org/abs/1708.05031","confidence":"Unknown","accessibility":""},{"id":"epoch:36ff1946b6fd86f58963","name":"Cutout-regularized net","organization":"University of Guelph,Vector Institute,CIFAR AI Research","date":"2017-08-15","domains":["Vision"],"country":"Canada,Canada,Canada","reference":"https://arxiv.org/abs/1708.04552","confidence":"Unknown","accessibility":""},{"id":"epoch:3a3ea4e8741aa8bfef3d","name":"EI-REHN-1000D","organization":"Korea Advanced Institute of Science and Technology (KAIST)","date":"2017-08-14","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/abs/1708.04116","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4bd2b83d5e215a93decd","name":"EI-REHN-1200D (PTB)","organization":"Korea Advanced Institute of Science and Technology (KAIST)","date":"2017-08-14","domains":["Language"],"country":"South Korea","reference":"https://arxiv.org/abs/1708.04116","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:8d063f67d8d19d6c2dbb","name":"OpenAI TI7 DOTA 1v1","organization":"OpenAI","date":"2017-08-11","domains":["Games"],"country":"United States of America","reference":"https://openai.com/research/dota-2","confidence":"Confident","accessibility":""},{"id":"epoch:392f7578856f43bcd337","name":"RetinaNet-R50","organization":"Facebook AI Research","date":"2017-08-07","domains":["Vision"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1708.02002","confidence":"","accessibility":""},{"id":"epoch:d8edb9dd0c0b20b7d160","name":"RetinaNet-R101","organization":"Facebook AI Research","date":"2017-08-07","domains":["Vision"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1708.02002","confidence":"Confident","accessibility":""},{"id":"epoch:47d9deaf4c5f898790ed","name":"AWD-LSTM - 3-layer LSTM (tied) + continuous cache pointer (WT2)","organization":"Salesforce Research","date":"2017-08-07","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1708.02182","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:20bc909460de1206fe91","name":"AWD-LSTM - 3-layer LSTM (tied) + continuous cache pointer (PTB)","organization":"Salesforce Research","date":"2017-08-07","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1708.02182","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:8f31d1afc99d09b57c67","name":"GSM","organization":"Peking University,Microsoft Research","date":"2017-07-30","domains":["Language"],"country":"China,United States of America","reference":"https://aclanthology.org/P17-1018/","confidence":"Likely","accessibility":""},{"id":"epoch:d8acad09d009409008c1","name":"ConvS2S (ensemble of 8 models)","organization":"Meta AI","date":"2017-07-25","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1705.03122","confidence":"Likely","accessibility":""},{"id":"epoch:34baab493e6e67cd3606","name":"NASNet-A","organization":"Google Brain","date":"2017-07-21","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1707.07012","confidence":"","accessibility":""},{"id":"epoch:623720b4b6eb47fc119f","name":"PSPNet","organization":"Chinese University of Hong Kong (CUHK)","date":"2017-07-21","domains":["Vision"],"country":"Hong Kong","reference":"https://ieeexplore.ieee.org/document/8100143","confidence":"Unknown","accessibility":""},{"id":"epoch:42bf316360331710d96a","name":"Densely Connected LSTM + Var. Dropout","organization":"Ghent University","date":"2017-07-19","domains":["Language"],"country":"Belgium","reference":"https://arxiv.org/abs/1707.06130","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:9af136e9bea896d548d6","name":"4 layer Densely Connected LSTM 14M (PTB)","organization":"Ghent University","date":"2017-07-19","domains":["Language"],"country":"Belgium","reference":"https://arxiv.org/pdf/1707.06130","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0232c213a873fd8c7948","name":"AWD-LSTM","organization":"DeepMind,University of Oxford","date":"2017-07-18","domains":["Language"],"country":"United Kingdom,United Kingdom","reference":"https://arxiv.org/abs/1707.05589","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:44a688640f704b1ac9bb","name":"JFT","organization":"Google Research,Carnegie Mellon University (CMU)","date":"2017-07-10","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1707.02968","confidence":"Confident","accessibility":""},{"id":"epoch:746a1449d0644ea8406e","name":"DeepLoc","organization":"Technical University of Denmark,University of Copenhagen","date":"2017-07-07","domains":["Biology"],"country":"Denmark,Denmark","reference":"https://academic.oup.com/bioinformatics/article/33/21/3387/3931857","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:9086dee73ed0102389c8","name":"ShuffleNet v1","organization":"Megvii Inc","date":"2017-07-03","domains":["Vision"],"country":"China","reference":"https://arxiv.org/abs/1707.01083","confidence":"","accessibility":""},{"id":"epoch:e36b544cfbe14b3c647f","name":"NoisyNet-Dueling","organization":"DeepMind","date":"2017-06-30","domains":["Games"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1706.10295v3","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:f4e590a21a1f9c012b77","name":"DeepLabV3","organization":"Google","date":"2017-06-17","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1706.05587","confidence":"Unknown","accessibility":""},{"id":"epoch:2d03522ce4ce3978102a","name":"HRA","organization":"Maluuba,Microsoft","date":"2017-06-13","domains":["Games"],"country":"Canada,United States of America","reference":"https://arxiv.org/abs/1706.04208","confidence":"Unknown","accessibility":""},{"id":"epoch:f0d2916f8d848e5616b3","name":"Transformer","organization":"Google Research,Google Brain","date":"2017-06-12","domains":["Language"],"country":"United States of America,United States of America","reference":"https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:a32653fa5fbcd394b9fe","name":"EDSR","organization":"Seoul National University","date":"2017-06-10","domains":["Vision","Image generation"],"country":"South Korea","reference":"https://arxiv.org/abs/1707.02921","confidence":"Unknown","accessibility":""},{"id":"epoch:20ba67cc6002ada34e2e","name":"Reading Twice for NLU","organization":"DeepMind","date":"2017-06-08","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1706.02596v3","confidence":"Unknown","accessibility":""},{"id":"epoch:50657c5974daf4b8bc8b","name":"PointNet++","organization":"Stanford University","date":"2017-06-07","domains":["3D modeling"],"country":"United States of America","reference":"https://arxiv.org/abs/1706.02413","confidence":"Unknown","accessibility":""},{"id":"epoch:e6b221a0e9a7eb03cfa2","name":"Inflated 3D ConvNet","organization":"DeepMind,University of Oxford","date":"2017-06-01","domains":["3D modeling"],"country":"United Kingdom,United Kingdom","reference":"https://arxiv.org/abs/1705.07750","confidence":"Unknown","accessibility":""},{"id":"epoch:1fa0ce9f92b3351b072b","name":"SRGAN","organization":"Twitter","date":"2017-05-25","domains":["Vision","Image generation"],"country":"United States of America","reference":"https://openaccess.thecvf.com/content_cvpr_2017/html/Ledig_Photo-Realistic_Single_Image_CVPR_2017_paper.html","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:453ac172a39467198466","name":"Low-Cost Collaborative Network","organization":"National University of Singapore,University of Technology Sydney,Qihoo 360 AI Institute","date":"2017-05-15","domains":["Vision"],"country":"Singapore,Australia,China","reference":"https://arxiv.org/abs/1703.08651","confidence":"Speculative","accessibility":""},{"id":"epoch:8cb702e9c138b1c9003c","name":"Mnemonic Reader","organization":"Fudan University,Microsoft Research","date":"2017-05-08","domains":["Language"],"country":"China,United States of America","reference":"https://arxiv.org/abs/1705.02798v6","confidence":"Confident","accessibility":""},{"id":"epoch:5bf2498a6295b9291460","name":"DeepLab (2017)","organization":"Johns Hopkins University,Google,University College London (UCL)","date":"2017-04-27","domains":["Vision"],"country":"United States of America,United States of America,United Kingdom","reference":"https://ieeexplore.ieee.org/abstract/document/7913730","confidence":"Unknown","accessibility":""},{"id":"epoch:946d6da190bcd0389898","name":"MobileNet","organization":"Google","date":"2017-04-17","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1704.04861","confidence":"","accessibility":""},{"id":"epoch:ff7264137a66da6890e1","name":"Tacotron","organization":"Google","date":"2017-04-06","domains":["Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/1703.10135","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:a7b6afbfc0a8b0aa2f96","name":"WGAN-GP","organization":"Courant Institute of Mathematical Sciences,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms)","date":"2017-03-31","domains":["Image generation"],"country":"United States of America,Canada","reference":"https://arxiv.org/abs/1704.00028","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:75d2c17cfc5862901627","name":"Mask R-CNN","organization":"Facebook AI Research","date":"2017-03-30","domains":["Vision"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1703.06870","confidence":"Unknown","accessibility":""},{"id":"epoch:437348494e814d3003e7","name":"AlexNet + coordinating filters","organization":"University of Pittsburgh,Duke University","date":"2017-03-28","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1703.09746","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:3a1ad36f681e9e014d97","name":"Word-Independent-SRNN+KN5","organization":"Saarland University","date":"2017-03-23","domains":["Language"],"country":"Germany","reference":"https://arxiv.org/abs/1703.08068","confidence":"","accessibility":"Unreleased"},{"id":"epoch:ba1434b15f17abc402fb","name":"Prototypical networks","organization":"University of Toronto,Twitter","date":"2017-03-15","domains":["Vision"],"country":"Canada,United States of America","reference":"https://arxiv.org/abs/1703.05175","confidence":"Unknown","accessibility":""},{"id":"epoch:cff45e2077a0c51e3e5b","name":"Variational Lossy Autoencoder (VLAE) MNIST","organization":"University of California (UC) Berkeley,OpenAI","date":"2017-03-04","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1611.02731","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:a29dddf488849b9414da","name":"SEST","organization":"Carnegie Mellon University (CMU)","date":"2017-03-02","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1703.00572v3","confidence":"Confident","accessibility":""},{"id":"epoch:68e4193df51105647b87","name":"DnCNN","organization":"Harbin Institute of Technology,Hong Kong Polytechnic University,ULSee Inc.,Xi\u2019an Jiaotong University","date":"2017-02-01","domains":["Vision","Image generation"],"country":"China,Hong Kong,China,China","reference":"https://ieeexplore.ieee.org/abstract/document/7839189","confidence":"Unknown","accessibility":""},{"id":"epoch:37a644232191c83ef3c9","name":"VDCNN (on Amazon Review Full dataset)","organization":"Facebook AI Research,University of Le Mans","date":"2017-01-27","domains":["Language"],"country":"United States of America,France,France","reference":"https://arxiv.org/abs/1606.01781","confidence":"Confident","accessibility":""},{"id":"epoch:46e4d0f5730c489022f9","name":"MoE-Multi","organization":"Jagiellonian University,Google Brain","date":"2017-01-23","domains":["Language"],"country":"Poland,United States of America","reference":"https://arxiv.org/abs/1701.06538","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:8a9e0056698a19bf3bfa","name":"PixelCNN++","organization":"OpenAI","date":"2017-01-19","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/1701.05517","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:65d05b12c5f7dd08c844","name":"OR-WideResNet","organization":"Duke University,University of Chinese Academy of Sciences","date":"2017-01-07","domains":["Vision"],"country":"United States of America,China","reference":"https://arxiv.org/abs/1701.01833v2","confidence":"Confident","accessibility":""},{"id":"epoch:acb29c57f39d9a736717","name":"DeepStack","organization":"University of Alberta,Charles University,Czech Technical University","date":"2017-01-06","domains":["Games"],"country":"Canada,Czechia,Czechia","reference":"https://arxiv.org/abs/1701.01724","confidence":"Speculative","accessibility":""},{"id":"epoch:f3a354e9878eaf2d0ae0","name":"YOLOv2","organization":"University of Washington,Allen Institute for AI","date":"2016-12-25","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1612.08242","confidence":"","accessibility":"Open weights (non-commercial)"},{"id":"epoch:fe781c4338bcb07051ef","name":"GCNN-14","organization":"Facebook AI Research","date":"2016-12-23","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1612.08083","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:b44ea3936c2080204792","name":"EnhanceNet","organization":"Max Planck Institute for Intelligent Systems","date":"2016-12-23","domains":["Vision"],"country":"Germany","reference":"https://www.semanticscholar.org/paper/EnhanceNet%3A-Single-Image-Super-Resolution-Through-Sajjadi-Scholkopf/fddc32f3880688238847077fd927ab3025db7a6a","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:22b6432c821e0bf056fa","name":"GCRN-M1, dropout","organization":"Ecole Polytechnique F\u00b4ed\u00b4erale de Lausanne (EPFL)","date":"2016-12-22","domains":["Language"],"country":"Switzerland","reference":"https://arxiv.org/abs/1612.07659","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e4a0b2ce9afd87dd3335","name":"3DMM-CNN","organization":"University of Southern California","date":"2016-12-15","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1612.04904","confidence":"Confident","accessibility":"Open weights (restricted use)"},{"id":"epoch:fb75e5d3fceea254e866","name":"Diabetic Retinopathy Detection Net","organization":"UT Austin,University of California (UC) Berkeley,Google","date":"2016-12-13","domains":["Vision"],"country":"United States of America,United States of America,United States of America","reference":"https://jamanetwork.com/journals/jama/article-abstract/2588763","confidence":"Unknown","accessibility":""},{"id":"epoch:4ed7f5956d9a59f855c8","name":"Neural cache model (size=2000) (300M)","organization":"Facebook AI Research","date":"2016-12-13","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1612.04426","confidence":"","accessibility":"Unreleased"},{"id":"epoch:4f73c53e488e761fd79b","name":"Neural cache model (size=2000)","organization":"Facebook AI Research","date":"2016-12-13","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1612.04426","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:6436029888302485159c","name":"LSTM (PTB)","organization":"Facebook AI Research","date":"2016-12-13","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1612.04426","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:5f9bbd49c46f5151b90d","name":"LSTM (WT2)","organization":"Facebook AI Research","date":"2016-12-13","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1612.04426","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:06e9ced9ded0943f662f","name":"LSTM (WT103)","organization":"Facebook AI Research","date":"2016-12-13","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1612.04426","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:d416ecae3911e8257517","name":"HR-ResNet101","organization":"Carnegie Mellon University (CMU)","date":"2016-12-13","domains":["Vision"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Finding-Tiny-Faces-Hu-Ramanan/71f51e1b6691343ed031c1ed42efc96a7f1a0619","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8ddc6fc737551d5135a5","name":"GAN-Advancer","organization":"OpenAI","date":"2016-12-05","domains":["Vision"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.5555/3157096.3157346","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:86500047d685f1b4aa29","name":"Layer-Norm Fast Weights RNN","organization":"University of Toronto,Google DeepMind,Google Brain","date":"2016-12-05","domains":["Vision"],"country":"Canada,United States of America,United States of America","reference":"https://arxiv.org/abs/1610.06258","confidence":"Confident","accessibility":""},{"id":"epoch:91965b64697228d2da55","name":"Elastic weight consolidation","organization":"DeepMind","date":"2016-12-02","domains":["Vision","Games"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1612.00796","confidence":"Unknown","accessibility":""},{"id":"epoch:edccb2ef20ab89332916","name":"PointNet","organization":"Stanford University","date":"2016-12-02","domains":["3D modeling"],"country":"United States of America","reference":"https://arxiv.org/abs/1612.00593","confidence":"Unknown","accessibility":""},{"id":"epoch:31878162011469fa8dae","name":"Image-to-image cGAN","organization":"University of California (UC) Berkeley","date":"2016-11-21","domains":["Vision","Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/1611.07004","confidence":"Unknown","accessibility":""},{"id":"epoch:405f5c605dba52a34076","name":"RefineNet","organization":"University of Adelaide,Australian Centre for Robotic Vision","date":"2016-11-20","domains":["Vision"],"country":"Australia,Australia","reference":"https://arxiv.org/abs/1611.06612v3","confidence":"Unknown","accessibility":""},{"id":"epoch:ec66b6d6932b0d65dc93","name":"PolyNet","organization":"Chinese University of Hong Kong (CUHK)","date":"2016-11-17","domains":["Vision"],"country":"Hong Kong","reference":"https://arxiv.org/abs/1611.05725","confidence":"Likely","accessibility":""},{"id":"epoch:0027267657c184a1e38f","name":"ResNeXt-50","organization":"University of California San Diego,Facebook","date":"2016-11-16","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1611.05431","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:42f37b995770e9cc5e61","name":"DAC-CSR","organization":"Jiangnan University,University of Surrey","date":"2016-11-16","domains":["Vision"],"country":"China,United Kingdom","reference":"https://arxiv.org/abs/1611.05396","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:48a43c728ed5d6415c8a","name":"ResNeXt-101 (64\u00d74d)","organization":"University of California San Diego,Facebook","date":"2016-11-16","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1611.05431","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:2bed12420c3a89c47f8a","name":"Deeply-recursive ConvNet","organization":"Seoul National University","date":"2016-11-11","domains":["Vision","Image generation"],"country":"South Korea","reference":"https://arxiv.org/abs/1511.04491","confidence":"Unknown","accessibility":""},{"id":"epoch:2fc0466d9e5e1c0fede0","name":"DTN (Domain Transfer Network)","organization":"Facebook AI Research","date":"2016-11-07","domains":["Vision","Image generation"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1611.02200","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:7e6c52250a1506e339b9","name":"DLDL (PASCAL)","organization":"University of Oxford","date":"2016-11-06","domains":["Vision"],"country":"United Kingdom","reference":"https://www.semanticscholar.org/paper/Deep-Label-Distribution-Learning-With-Label-Gao-Xing/02567fd428a675ca91a0c6786f47f3e35881bcbd","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e72ca12951e3d4fcc45d","name":"NASv3 (CIFAR-10)","organization":"Google Brain","date":"2016-11-05","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1611.01578","confidence":"Likely","accessibility":""},{"id":"epoch:b063f7d41e78c7637064","name":"NAS with base 8 and shared embeddings","organization":"Google Brain","date":"2016-11-05","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1611.01578","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2885bcb0ac7f676ad7cf","name":"BIDAF","organization":"University of Washington,Allen Institute for AI","date":"2016-11-05","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1611.01603v6","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:72f01c9a2ad8b494d266","name":"VD-LSTM+REAL Large","organization":"Salesforce Research,Stanford University","date":"2016-11-04","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1611.01462","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:3aab67cbc272dcd5f369","name":"VD-LSTM+REAL Medium","organization":"Stanford University,Salesforce Research","date":"2016-11-04","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1611.01462","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:8aea9f347bc0376bf8c2","name":"VD-LSTM+REAL Small","organization":"Stanford University,Salesforce Research","date":"2016-11-04","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1611.01462","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:777597845fbaa34a959b","name":"SPIDER2","organization":"Griffith University,University of Iowa,Dezhou University","date":"2016-10-28","domains":["Biology"],"country":"Australia,United States of America,China","reference":"https://link.springer.com/protocol/10.1007/978-1-4939-6406-2_6","confidence":"Likely","accessibility":"Open weights (non-commercial)"},{"id":"epoch:f79e251c883848e17f5b","name":"Differentiable neural computer","organization":"Google DeepMind","date":"2016-10-12","domains":["Language"],"country":"United States of America","reference":"https://www.nature.com/articles/nature20101","confidence":"Unknown","accessibility":""},{"id":"epoch:57ec1447b5de4c8fed9c","name":"GAWWN","organization":"University of Michigan,Max Planck Institute for Informatics","date":"2016-10-08","domains":["Vision"],"country":"United States of America,Germany","reference":"https://www.semanticscholar.org/paper/Learning-What-and-Where-to-Draw-Reed-Akata/cad4ac0d2389a89cf1955dd4788278c1e8ac1af9","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:717f76fb5c57e1f108cd","name":"Xception","organization":"Google","date":"2016-10-07","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1610.02357","confidence":"Confident","accessibility":""},{"id":"epoch:254bc33523ba687ee60a","name":"GNMT","organization":"Google","date":"2016-09-26","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1609.08144","confidence":"Likely","accessibility":"Hosted access (no API)"},{"id":"epoch:d648aaf33dbac7ee8f35","name":"Pointer Sentinel-LSTM (medium)","organization":"MetaMind Inc,Salesforce","date":"2016-09-26","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1609.07843","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:e61a6b1465f3f8574a52","name":"Zoneout + Variational LSTM (WT2)","organization":"MetaMind Inc,Salesforce","date":"2016-09-26","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1609.07843","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4b82b8a83a43f1078f37","name":"Byte-mLSTM+emb+WN+VD","organization":"University of Edinburgh,Toyota Technological Institute at Chicago","date":"2016-09-26","domains":["Language"],"country":"United Kingdom,United States of America","reference":"https://arxiv.org/abs/1609.07959","confidence":"","accessibility":"Unreleased"},{"id":"epoch:7cb22330b2bdc754fece","name":"Zoneout + Variational LSTM (PTB)","organization":"MetaMind Inc","date":"2016-09-26","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1609.07843","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f6e26c5713cef2482432","name":"Pointer Sentinel-LSTM (WT2)","organization":"MetaMind Inc","date":"2016-09-26","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1609.07843","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:3887e202f91aae9874ca","name":"Knowledge distillation student model","organization":"Harvard University","date":"2016-09-22","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1606.07947","confidence":"Speculative","accessibility":""},{"id":"epoch:4e980f9cc41cffc74a7e","name":"Wide Residual Network","organization":"Universit\u00e9 Paris-Est","date":"2016-09-19","domains":["Vision"],"country":"France","reference":"https://arxiv.org/abs/1605.07146","confidence":"Unknown","accessibility":""},{"id":"epoch:ca5525c383d2fe8e470d","name":"MS-CNN","organization":"IBM,University of California San Diego","date":"2016-09-17","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://link.springer.com/chapter/10.1007/978-3-319-46493-0_22","confidence":"Unknown","accessibility":""},{"id":"epoch:5ec019106db534761d9b","name":"Stacked hourglass network","organization":"University of Michigan","date":"2016-09-17","domains":["Vision"],"country":"United States of America","reference":"https://link.springer.com/chapter/10.1007/978-3-319-46484-8_29","confidence":"Unknown","accessibility":""},{"id":"epoch:08ac922cb6219d281f67","name":"TSN","organization":"ETH Zurich,Shenzhen Institute of Advanced Technology,Chinese University of Hong Kong (CUHK)","date":"2016-09-17","domains":["Video"],"country":"Switzerland,China,Hong Kong","reference":"https://link.springer.com/chapter/10.1007/978-3-319-46484-8_2","confidence":"Unknown","accessibility":""},{"id":"epoch:4385e7959a7986903b15","name":"ResNet-1001","organization":"Microsoft","date":"2016-09-17","domains":["Vision"],"country":"United States of America","reference":"https://link.springer.com/chapter/10.1007/978-3-319-46493-0_38","confidence":"","accessibility":""},{"id":"epoch:9705e03d9ba211e3783b","name":"ResNet-200","organization":"Microsoft Research Asia","date":"2016-09-17","domains":["Vision"],"country":"China","reference":"https://link.springer.com/chapter/10.1007/978-3-319-46493-0_38","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:532b5f21338870c4a33e","name":"Youtube recommendation model","organization":"Google","date":"2016-09-15","domains":["Recommendation"],"country":"United States of America","reference":"https://research.google/pubs/pub45530/","confidence":"Unknown","accessibility":""},{"id":"epoch:ea427eb6a387f51421a6","name":"WaveNet","organization":"Google DeepMind","date":"2016-09-12","domains":["Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/1609.03499","confidence":"Unknown","accessibility":""},{"id":"epoch:318038fd47e8b0c5da90","name":"MS-ensemble-speech-recognition","organization":"Microsoft","date":"2016-09-12","domains":["Speech"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/The-microsoft-2016-conversational-speech-system-Xiong-Droppo/ac94ef90be9b0c3bf744d6744e47b38855f9a4c7","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:63646df6e5a6f548429b","name":"LF-MMI","organization":"Johns Hopkins University,Cornell University","date":"2016-09-08","domains":["Speech"],"country":"United States of America,United States of America","reference":"https://www.semanticscholar.org/paper/Purely-Sequence-Trained-Neural-Networks-for-ASR-on-Povey-Peddinti/6ce6a9a30cd69bd2842a4b581cf48c6815bdfdd8","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:cfd0507a6067167e2cbf","name":"Multi-task Cascaded CNN","organization":"Chinese Academy of Sciences,Chinese University of Hong Kong (CUHK)","date":"2016-08-26","domains":["Vision"],"country":"China,Hong Kong","reference":"https://arxiv.org/abs/1604.02878","confidence":"Unknown","accessibility":""},{"id":"epoch:b52dc70b70cc0b25e79a","name":"DenseNet-264","organization":"Tsinghua University,Facebook AI Research,Cornell University","date":"2016-08-25","domains":["Vision"],"country":"China,United States of America,France,United States of America","reference":"https://arxiv.org/abs/1608.06993","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:009ea8ec0d421fc3dce7","name":"SimpleNet","organization":"Sensifai,Islamic Azad University,Technicolor R&I,Institute for Research in Fundamental Sciences (IPM)","date":"2016-08-22","domains":["Vision"],"country":"Belgium,Iran (Islamic Republic of),France,Iran (Islamic Republic of)","reference":"https://arxiv.org/abs/1608.06037","confidence":"Confident","accessibility":""},{"id":"epoch:198e3c31bd9fde79b17e","name":"Attend-Infer-Repeat","organization":"Google DeepMind","date":"2016-08-12","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1603.08575","confidence":"Confident","accessibility":""},{"id":"epoch:5cb38c9a324c135ef83f","name":"Order embeddings with layer norm","organization":"University of Toronto","date":"2016-07-21","domains":["Vision"],"country":"Canada","reference":"https://arxiv.org/abs/1607.06450","confidence":"Confident","accessibility":""},{"id":"epoch:b2ec045d875a1d5cc6f2","name":"Layer Normalization: The Attentive Reader","organization":"University of Toronto","date":"2016-07-21","domains":["Language"],"country":"Canada","reference":"https://arxiv.org/abs/1607.06450","confidence":"Confident","accessibility":""},{"id":"epoch:0565206cea7a67adfc0c","name":"Layer Normalization: Skip Thoughts","organization":"University of Toronto","date":"2016-07-21","domains":["Language"],"country":"Canada","reference":"https://arxiv.org/abs/1607.06450","confidence":"Confident","accessibility":""},{"id":"epoch:000f2a008e7f6575660e","name":"Layer Normalization: Draw","organization":"University of Toronto","date":"2016-07-21","domains":["Image generation"],"country":"Canada","reference":"https://arxiv.org/abs/1607.06450","confidence":"Confident","accessibility":""},{"id":"epoch:da3a90e81fdb94f70d52","name":"Layer Normalization: Handwriting sequence generation","organization":"University of Toronto","date":"2016-07-21","domains":["Image generation"],"country":"Canada","reference":"https://arxiv.org/abs/1607.06450","confidence":"Speculative","accessibility":""},{"id":"epoch:47c0eac5499801a6f499","name":"Character-enriched word2vec","organization":"Facebook AI Research","date":"2016-07-15","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1607.04606","confidence":"Unknown","accessibility":""},{"id":"epoch:03f28a809cb64ef41fe6","name":"VD-RHN","organization":"ETH Zurich,IDSIA","date":"2016-07-12","domains":["Language"],"country":"Switzerland,Switzerland","reference":"https://arxiv.org/abs/1607.03474","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:2e3116509dd7e15ad0b2","name":"Variational RHN + WT (PTB)","organization":"ETH Zurich,IDSIA","date":"2016-07-12","domains":["Language"],"country":"Switzerland,Switzerland","reference":"https://arxiv.org/abs/1607.03474","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:0d50990d6bcc40851ec2","name":"fastText","organization":"Facebook AI Research","date":"2016-07-06","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1607.01759","confidence":"Unknown","accessibility":""},{"id":"epoch:8c6f9910ffa3202b354c","name":"node2vec","organization":"Stanford University","date":"2016-07-03","domains":["Mathematics"],"country":"United States of America","reference":"https://arxiv.org/abs/1607.00653","confidence":"Speculative","accessibility":""},{"id":"epoch:d2d208e966acb1378f97","name":"CCL","organization":"SenseTime,Chinese University of Hong Kong (CUHK),Chinese Academy of Sciences","date":"2016-06-27","domains":["Vision"],"country":"Hong Kong,Hong Kong,China","reference":"https://www.semanticscholar.org/paper/Unconstrained-Face-Alignment-via-Cascaded-Learning-Zhu-Li/4396563ee3ec167932071e7936111f6a8899e905","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c0ba9fc8400977af3589","name":"Wide & Deep","organization":"Google","date":"2016-06-24","domains":["Recommendation"],"country":"United States of America","reference":"https://arxiv.org/abs/1606.07792","confidence":"Unknown","accessibility":""},{"id":"epoch:141ab822d413f7c14ae8","name":"R-FCN","organization":"Tsinghua University,Microsoft Research","date":"2016-06-21","domains":["Vision"],"country":"China,United States of America","reference":"https://arxiv.org/abs/1605.06409","confidence":"Confident","accessibility":""},{"id":"epoch:527aa5049d840bc51a4c","name":"DMN","organization":"Salesforce","date":"2016-06-20","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1506.07285","confidence":"Unknown","accessibility":""},{"id":"epoch:f8ddb534db862230c902","name":"Segmental RNN","organization":"University of Edinburgh,Carnegie Mellon University (CMU),University of Washington","date":"2016-06-20","domains":["Speech"],"country":"United Kingdom,United States of America,United States of America","reference":"https://arxiv.org/abs/1603.00223","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:29435812d8f5d1f8e747","name":"CMS-RCNN","organization":"IEEE","date":"2016-06-17","domains":["Vision"],"country":"Multinational","reference":"https://www.semanticscholar.org/paper/CMS-RCNN%3A-Contextual-Multi-Scale-Region-based-CNN-Zhu-Zheng/e8b2a98f87b7b2593b4a046464c1ec63bfd13b51","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:012e78eb326f28c35f50","name":"PixelCNN","organization":"Google DeepMind","date":"2016-06-16","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1606.05328","confidence":"Unknown","accessibility":""},{"id":"epoch:1f2172a57ffe64742add","name":"Spatiotemporal fusion ConvNet","organization":"Graz University of Technology,University of Oxford","date":"2016-06-01","domains":["Video"],"country":"Austria,United Kingdom","reference":"https://openaccess.thecvf.com/content_cvpr_2016/html/Feichtenhofer_Convolutional_Two-Stream_Network_CVPR_2016_paper.html","confidence":"","accessibility":""},{"id":"epoch:24f833b36568005541d4","name":"Part-of-sentence tagging model","organization":"Carnegie Mellon University (CMU)","date":"2016-05-29","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1603.01354","confidence":"Confident","accessibility":""},{"id":"epoch:c9be72f58d277c37062c","name":"LRR-4X","organization":"UC Irvine","date":"2016-05-08","domains":["Vision"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Laplacian-Pyramid-Reconstruction-and-Refinement-for-Ghiasi-Fowlkes/2796c448023b78fd77f3a4b57966f257c5e654c2","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:e038372211a14879fbec","name":"Gated HORNN (3rd order)","organization":"York University","date":"2016-04-30","domains":["Language"],"country":"Canada","reference":"https://arxiv.org/abs/1605.00064","confidence":"","accessibility":"Unreleased"},{"id":"epoch:2f1966e87dfd71bfeaeb","name":"Dueling DQN","organization":"Google DeepMind","date":"2016-04-05","domains":["Games"],"country":"United States of America","reference":"https://arxiv.org/abs/1511.06581","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:7b01a9deaaa85ef05d47","name":"Symmetric Residual Encoder-Decoder Net","organization":"Nanjing University,University of Adelaide","date":"2016-03-30","domains":["Vision","Image generation"],"country":"China,Australia","reference":"https://arxiv.org/abs/1603.09056v2","confidence":"Unknown","accessibility":""},{"id":"epoch:7e4e1de28b28d465d2b7","name":"Binarized Neural Network (MNIST)","organization":"Technion - Israel Institute of Technology,Columbia University,University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2016-03-17","domains":["Vision"],"country":"Israel,United States of America,Canada","reference":"https://arxiv.org/abs/1602.02830","confidence":"Speculative","accessibility":""},{"id":"epoch:50b0fd8d9791f25fd9e9","name":"Template Adaptation","organization":"University of Oxford","date":"2016-03-12","domains":["Vision"],"country":"United Kingdom","reference":"https://www.semanticscholar.org/paper/Template-Adaptation-for-Face-Verification-and-Crosswhite-Byrne/42515cd370f5edf4ac6c8d4bc3f6e0d848f6b73a","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:4784d9c42a60c1b3b424","name":"Named Entity Recognition model","organization":"Carnegie Mellon University (CMU)","date":"2016-03-04","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1603.01354","confidence":"Confident","accessibility":""},{"id":"epoch:884bca6c3d5783755bb2","name":"Double DQN","organization":"Google DeepMind","date":"2016-03-02","domains":["Games"],"country":"United States of America","reference":"https://ojs.aaai.org/index.php/AAAI/article/view/10295","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:a2e754309878f4294e56","name":"Order-Embeddings of Images and Language","organization":"University of Toronto","date":"2016-03-01","domains":["Vision"],"country":"Canada","reference":"https://arxiv.org/abs/1511.06361","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:7df4c20c3788ca120b2b","name":"SqueezeNet","organization":"DeepScale,University of California (UC) Berkeley,Stanford University","date":"2016-02-24","domains":["Vision"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/1602.07360","confidence":"","accessibility":""},{"id":"epoch:79d7174b174923928d15","name":"Inception-ResNet-V2","organization":"Google","date":"2016-02-23","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1602.07261","confidence":"","accessibility":""},{"id":"epoch:d17a5ddb1e2df9261883","name":"Inceptionv4","organization":"Google","date":"2016-02-23","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1602.07261","confidence":"","accessibility":""},{"id":"epoch:54fd7af4f271eb5d8ef1","name":"BIG LSTM+CNN INPUTS","organization":"Google Brain","date":"2016-02-11","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1602.02410","confidence":"Likely","accessibility":""},{"id":"epoch:10faca516e5f866b2565","name":"10 LSTMS + KN-5 (OPTIMAL WEIGHTS)","organization":"Google Brain","date":"2016-02-11","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1602.02410","confidence":"Likely","accessibility":""},{"id":"epoch:7997f03f1965204a72db","name":"A3C FF hs","organization":"Google,University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2016-02-04","domains":["Games"],"country":"United States of America,Canada","reference":"http://arxiv.org/abs/1602.01783v2","confidence":"Unknown","accessibility":""},{"id":"epoch:6b9c30592a9c2e4c874b","name":"Convolutional Pose Machines","organization":"Carnegie Mellon University (CMU)","date":"2016-01-30","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1602.00134","confidence":"Unknown","accessibility":""},{"id":"epoch:b52941072e1d41f490ae","name":"AlphaGo Lee","organization":"DeepMind","date":"2016-01-27","domains":["Games"],"country":"United Kingdom","reference":"https://www.nature.com/articles/nature16961","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:2052a6346e5d7657d838","name":"Variational (untied weights, MC) LSTM (Large)","organization":"University of Cambridge","date":"2015-12-16","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1512.05287","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:9e168cc6880992e1fc42","name":"Advantage Learning","organization":"Google DeepMind","date":"2015-12-15","domains":["Games"],"country":"United States of America","reference":"http://arxiv.org/abs/1512.04860v1","confidence":"Unknown","accessibility":""},{"id":"epoch:d9338a6b1f312beda9c3","name":"BPL","organization":"University of Toronto,New York University (NYU),Massachusetts Institute of Technology (MIT)","date":"2015-12-11","domains":["Image generation"],"country":"Canada,United States of America,United States of America","reference":"https://science.sciencemag.org/content/350/6266/1332/","confidence":"Unknown","accessibility":""},{"id":"epoch:5c94da8f78f5d0254203","name":"ResNet-152 (ImageNet)","organization":"Microsoft","date":"2015-12-10","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1512.03385","confidence":"Confident","accessibility":""},{"id":"epoch:e2be9ac83c3943b939c1","name":"ResNet-110 (CIFAR-10)","organization":"Microsoft","date":"2015-12-10","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1512.03385","confidence":"","accessibility":""},{"id":"epoch:08737f064f1922ba9f65","name":"ResNet-101 (ImageNet)","organization":"Microsoft","date":"2015-12-10","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1512.03385","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6659e18078054161429a","name":"DeepSpeech2 (English)","organization":"Baidu Research - Silicon Valley AI Lab","date":"2015-12-08","domains":["Speech"],"country":"United States of America","reference":"https://arxiv.org/abs/1512.02595","confidence":"Confident","accessibility":""},{"id":"epoch:8d92de535766eff6fab2","name":"SSD","organization":"","date":"2015-12-08","domains":["Vision"],"country":"","reference":"https://arxiv.org/pdf/1512.02325","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:ab99247fe891263a5eac","name":"Inception v3","organization":"Google,University College London (UCL)","date":"2015-12-02","domains":["Vision"],"country":"United States of America,United Kingdom","reference":"https://arxiv.org/abs/1512.00567","confidence":"Likely","accessibility":""},{"id":"epoch:a7e3f388f841267b8019","name":"Netflix Recommender System","organization":"Netflix","date":"2015-12-01","domains":["Recommendation"],"country":"United States of America","reference":"https://dl.acm.org/doi/pdf/10.1145/2843948","confidence":"Unknown","accessibility":""},{"id":"epoch:09224d63099fddda430b","name":"Multi-scale Dilated CNN","organization":"Princeton University,Intel Labs","date":"2015-11-23","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1511.07122","confidence":"Unknown","accessibility":""},{"id":"epoch:d8ba413bf204232c15d7","name":"Highway Network","organization":"IDSIA,SUPSI","date":"2015-11-23","domains":["Vision"],"country":"Switzerland,Switzerland","reference":"https://arxiv.org/abs/1507.06228","confidence":"Confident","accessibility":""},{"id":"epoch:e7ddfd6de790f5f78d3f","name":"3DDFA","organization":"Chinese Academy of Sciences,Michigan State University","date":"2015-11-23","domains":["Vision"],"country":"China,United States of America","reference":"https://arxiv.org/abs/1511.07212","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c2dd9fa75f8a44cd620f","name":"The Attentive Reader","organization":"Google DeepMind","date":"2015-11-19","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1506.03340","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8fcbf30faabda8b8c9a6","name":"SAF R-CNN","organization":"Beijing Institute of Technology,Sun Yat-sen University,Panasonic R&D,National University of Singapore","date":"2015-10-28","domains":["Vision"],"country":"China,China,Singapore,Singapore","reference":"https://arxiv.org/abs/1510.08160","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:9947a60096ef7335307a","name":"AlphaGo Fan","organization":"DeepMind","date":"2015-10-01","domains":["Games"],"country":"United Kingdom","reference":"https://www.nature.com/articles/nature16961","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:48493fb88945d694670f","name":"Deep Deterministic Policy Gradients","organization":"Google DeepMind","date":"2015-09-09","domains":["Robotics"],"country":"United States of America","reference":"https://arxiv.org/abs/1509.02971","confidence":"Unknown","accessibility":""},{"id":"epoch:ddfbb5b1e9a74020e0fc","name":"BPE","organization":"University of Edinburgh","date":"2015-08-31","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1508.07909","confidence":"","accessibility":""},{"id":"epoch:3519204ff0e0a111cf77","name":"LSTM-Char-Large","organization":"Harvard University,New York University (NYU)","date":"2015-08-26","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1508.06615","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:366027f767ae5ff121c6","name":"Listen, Attend and Spell","organization":"Google,Carnegie Mellon University (CMU)","date":"2015-08-20","domains":["Speech"],"country":"United States of America,United States of America","reference":"https://ieeexplore.ieee.org/document/7472621","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:3e33333dca03ae421dc9","name":"DCNN","organization":"University of Maryland,Rutgers University","date":"2015-08-07","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1508.01722","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:fc67fc7a09cb87640566","name":"Deep CNN + COTS","organization":"IEEE","date":"2015-07-26","domains":["Vision"],"country":"Multinational","reference":"https://www.semanticscholar.org/paper/Face-Search-at-Scale%3A-80-Million-Gallery-Wang-Otto/be38934db99182c0e3100be58cd0299151b30625","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:a60e570dce511754c430","name":"CompACT-Deep","organization":"University of California San Diego","date":"2015-07-19","domains":["Vision"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Learning-Complexity-Aware-Cascades-for-Deep-Cai-Saberian/c3bd3b9782dc504ee4f2b8a12bd9c562a5c0d7ad","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:73782472b5953476ccc3","name":"Search-Proven Best LSTM","organization":"Google","date":"2015-07-06","domains":["Language"],"country":"United States of America","reference":"https://proceedings.mlr.press/v37/jozefowicz15.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:ba8438db50612a943851","name":"Skip-Thoughts","organization":"University of Toronto,Massachusetts Institute of Technology (MIT),CIFAR AI Research","date":"2015-06-22","domains":["Language"],"country":"Canada,United States of America,Canada","reference":"https://arxiv.org/abs/1506.06726","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:6ce4de232b1fa39f3830","name":"BatchNorm","organization":"Google","date":"2015-06-15","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1502.03167","confidence":"Confident","accessibility":""},{"id":"epoch:7a8d4ec1b9bb7a5d6681","name":"YOLO","organization":"University of Washington,Allen Institute for AI,Facebook AI Research","date":"2015-06-08","domains":["Vision"],"country":"United States of America,United States of America,United States of America,France","reference":"https://arxiv.org/abs/1506.02640","confidence":"","accessibility":""},{"id":"epoch:3329d781825becd2a3a6","name":"CFSS","organization":"SenseTime,Chinese University of Hong Kong (CUHK),Shenzhen Institute of Advanced Technology","date":"2015-06-07","domains":["Vision"],"country":"Hong Kong,Hong Kong,China","reference":"https://www.semanticscholar.org/paper/Face-alignment-by-coarse-to-fine-shape-searching-Zhu-Li/dc661f478c5a245a919c4b3faae8ea4f598b6e64","confidence":"Likely","accessibility":"Open weights (unrestricted)"},{"id":"epoch:37feb923983810621859","name":"Faster R-CNN","organization":"Microsoft Research","date":"2015-06-04","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1506.01497","confidence":"Unknown","accessibility":"Open weights (unrestricted)"},{"id":"epoch:1684cabefb217356acad","name":"Draw","organization":"Google DeepMind","date":"2015-05-20","domains":["Image generation"],"country":"United States of America","reference":"https://arxiv.org/abs/1502.04623","confidence":"Confident","accessibility":""},{"id":"epoch:4aacb751d1786c378822","name":"Trajectory-pooled conv nets","organization":"Chinese University of Hong Kong (CUHK),Chinese Academy of Sciences","date":"2015-05-19","domains":["Video"],"country":"Hong Kong,China","reference":"https://www.cv-foundation.org/openaccess/content_cvpr_2015/html/Wang_Action_Recognition_With_2015_CVPR_paper.html","confidence":"","accessibility":""},{"id":"epoch:23951cd2ba14cc493e22","name":"U-Net","organization":"University of Freiburg","date":"2015-05-18","domains":["Vision"],"country":"Germany","reference":"https://www.semanticscholar.org/paper/U-Net%3A-Convolutional-Networks-for-Biomedical-Image-Ronneberger-Fischer/6364fdaa0a0eccd823a779fcdd489173f938e91a","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:58a64ecb2d17ce27c9bf","name":"2nd order FOFE-FNNLM","organization":"University of Science and Technology of China (USTC),York University","date":"2015-05-06","domains":["Language"],"country":"China,Canada","reference":"https://arxiv.org/abs/1505.01504","confidence":"","accessibility":"Unreleased"},{"id":"epoch:5d22e9b4948d174d7e69","name":"Deep LSTM video classifier","organization":"University of Texas at Austin,Google","date":"2015-05-01","domains":["Video"],"country":"United States of America,United States of America","reference":"https://www.cv-foundation.org/openaccess/content_cvpr_2015/html/Ng_Beyond_Short_Snippets_2015_CVPR_paper.html","confidence":"Unknown","accessibility":""},{"id":"epoch:86f1c9a2c8d5374096ec","name":"Fast R-CNN","organization":"Microsoft Research","date":"2015-04-30","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1504.08083","confidence":"Unknown","accessibility":""},{"id":"epoch:a3ffe4bd3015c35bfdfc","name":"TC-DNN-BLSTM-DNN","organization":"Carnegie Mellon University (CMU)","date":"2015-04-06","domains":["Speech"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Deep-Recurrent-Neural-Networks-for-Acoustic-Chan-Lane/ae18f67c10a5d4da91f128ec7a6cf7c784122cd5","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:74f4a897cb30a1d0da80","name":"genCNN + dyn eval","organization":"Chinese Academy of Sciences,Huawei Noah's Ark Lab,Dublin City University","date":"2015-03-17","domains":["Language"],"country":"China,China,Ireland","reference":"https://aclanthology.org/P15-1151/","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:05923a966d81eeece817","name":"Stack RNN","organization":"Facebook AI Research","date":"2015-03-03","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1503.01007","confidence":"","accessibility":"Unreleased"},{"id":"epoch:249d03e7324537e50385","name":"Constituency-Tree LSTM","organization":"MetaMind Inc,Stanford University","date":"2015-02-28","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1503.00075","confidence":"","accessibility":""},{"id":"epoch:a98816e061cf1ce2517e","name":"DQN-2015","organization":"Google","date":"2015-02-25","domains":["Games"],"country":"United States of America","reference":"https://www.nature.com/articles/nature14236","confidence":"","accessibility":""},{"id":"epoch:d6c7d43dd2194ba260bf","name":"TRPO","organization":"University of California (UC) Berkeley","date":"2015-02-19","domains":["Games"],"country":"United States of America","reference":"https://arxiv.org/pdf/1502.05477","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:4e07374fe10ec1b7f344","name":"CRF-RNN","organization":"University of Oxford,Stanford University,Baidu","date":"2015-02-11","domains":["Vision"],"country":"United Kingdom,United States of America,China","reference":"https://arxiv.org/abs/1502.03240","confidence":"Unknown","accessibility":""},{"id":"epoch:787858b685ff69a56935","name":"MSRA (C, PReLU)","organization":"Microsoft Research","date":"2015-02-06","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1502.01852","confidence":"Confident","accessibility":""},{"id":"epoch:86f312af2cb789cec1d9","name":"VGG-Face","organization":"University of Oxford","date":"2015-01-01","domains":["Vision"],"country":"United Kingdom","reference":"https://www.semanticscholar.org/paper/Deep-Face-Recognition-Parkhi-Vedaldi/162ea969d1929ed180cc6de9f0bf116993ff6e06","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:59823beb8dfed9d942ec","name":"SCRN (Structurally Constrained Recurrent Network)","organization":"Facebook AI Research","date":"2014-12-24","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1412.7753","confidence":"","accessibility":"Unreleased"},{"id":"epoch:206cd840f140952d9236","name":"N-gram+Cache (PTB)","organization":"Facebook AI Research","date":"2014-12-24","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1412.7753","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:6b03b03f7f90cc435a18","name":"N-gram (PTB)","organization":"Facebook AI Research","date":"2014-12-24","domains":["Language"],"country":"United States of America,France","reference":"https://arxiv.org/abs/1412.7753","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:0d689a1678ca02a55470","name":"ADAM (CIFAR-10)","organization":"University of Amsterdam,OpenAI,University of Toronto","date":"2014-12-22","domains":["Vision"],"country":"Netherlands,United States of America,Canada","reference":"https://arxiv.org/abs/1412.6980","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:d6cfb0eec8da6bb67263","name":"DeepLab","organization":"Google,University of California Los Angeles (UCLA)","date":"2014-12-22","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1412.7062","confidence":"Unknown","accessibility":""},{"id":"epoch:348578040036e69ae67a","name":"4-gram + 8 DENN","organization":"IBM","date":"2014-12-22","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1412.7063","confidence":"","accessibility":"Unreleased"},{"id":"epoch:eacc347884e5333048af","name":"Fractional Max-Pooling","organization":"University of Warwick","date":"2014-12-18","domains":["Vision"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1412.6071v4","confidence":"Likely","accessibility":""},{"id":"epoch:e23684f82b14e3fe9e82","name":"NTM","organization":"Google DeepMind","date":"2014-12-10","domains":["Other","Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1410.5401","confidence":"Unknown","accessibility":""},{"id":"epoch:c1baae40961dc434e024","name":"SNM-skip","organization":"Google","date":"2014-12-03","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1412.1454","confidence":"Speculative","accessibility":""},{"id":"epoch:4b2273d35ad59c8b748e","name":"TA-CNN","organization":"Chinese University of Hong Kong (CUHK)","date":"2014-11-29","domains":["Vision"],"country":"Hong Kong","reference":"https://arxiv.org/abs/1412.0069","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:70511e059c3b26936aee","name":"Cascaded LNet-ANet","organization":"Chinese University of Hong Kong (CUHK)","date":"2014-11-28","domains":["Vision"],"country":"Hong Kong","reference":"https://arxiv.org/abs/1411.7766","confidence":"Unknown","accessibility":""},{"id":"epoch:c9bd36d624ef8e17039b","name":"Fully Convolutional Networks","organization":"University of California (UC) Berkeley","date":"2014-11-14","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1411.4038","confidence":"Unknown","accessibility":""},{"id":"epoch:96c148b73fb4c8d697e6","name":"SC-NLM","organization":"University of Toronto","date":"2014-11-10","domains":["Multimodal","Vision","Language"],"country":"Canada","reference":"https://www.semanticscholar.org/paper/Unifying-Visual-Semantic-Embeddings-with-Multimodal-Kiros-Salakhutdinov/2e36ea91a3c8fbff92be2989325531b4002e2afc","confidence":"Confident","accessibility":""},{"id":"epoch:e409d88ff37eb2dc904b","name":"LRCN","organization":"UT Austin,University of Massachusetts Lowell,University of California (UC) Berkeley","date":"2014-11-07","domains":["Video"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/1411.4389","confidence":"","accessibility":""},{"id":"epoch:62987f1087e897907485","name":"Spatially-Sparse CNN","organization":"University of Warwick","date":"2014-09-23","domains":["Vision"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1409.6070v1","confidence":"Unknown","accessibility":""},{"id":"epoch:7d086f031de48eba946e","name":"Deeply-supervised nets","organization":"Microsoft Research","date":"2014-09-18","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1409.5185","confidence":"","accessibility":""},{"id":"epoch:f243365f1343b7b4582b","name":"GoogLeNet / InceptionV1","organization":"Google,University of Michigan,University of North Carolina","date":"2014-09-17","domains":["Vision"],"country":"United States of America,United States of America,United States of America","reference":"https://arxiv.org/abs/1409.4842","confidence":"Confident","accessibility":""},{"id":"epoch:b8a1a7cb34b7daaeb270","name":"SPN-4+KN5","organization":"Singapore University of Technology & Design,DSO National Laboratories","date":"2014-09-14","domains":["Language"],"country":"Singapore,Singapore","reference":"https://www.comp.nus.edu.sg/~skok/papers/is14.pdf","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:cac7b5478f714b05e606","name":"Seq2Seq LSTM","organization":"Google","date":"2014-09-10","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1409.3215","confidence":"Confident","accessibility":""},{"id":"epoch:84cf7a7c495d6ae5626d","name":"Large regularized LSTM","organization":"New York University (NYU),Google Brain","date":"2014-09-08","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/1409.2329","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:df736489086af36e8b4e","name":"VGG16","organization":"University of Oxford","date":"2014-09-04","domains":["Vision"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1409.1556","confidence":"Confident","accessibility":""},{"id":"epoch:4e8cbb1e375262b32016","name":"VGG19","organization":"University of Oxford","date":"2014-09-04","domains":["Vision"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1409.1556","confidence":"Likely","accessibility":""},{"id":"epoch:6cc235593bdedecf088e","name":"RNNsearch-50*","organization":"Jacobs University Bremen,University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2014-09-01","domains":["Language"],"country":"Germany,Canada","reference":"https://arxiv.org/abs/1409.0473","confidence":"Confident","accessibility":""},{"id":"epoch:5ec91b6e20e258f08032","name":"AdClickNet","organization":"Facebook","date":"2014-08-24","domains":["Recommendation"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.1145/2648584.2648589","confidence":"Unknown","accessibility":""},{"id":"epoch:bb110c7637e92374b543","name":"NPD","organization":"IEEE","date":"2014-08-06","domains":["Vision"],"country":"Multinational","reference":"https://www.semanticscholar.org/paper/A-Fast-and-Accurate-Unconstrained-Face-Detector-Liao-Jain/fb67880d99ca29125866087dad2934ccc221378a","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:247403985cc7db1edd41","name":"ACF-WIDER","organization":"Chinese Academy of Sciences","date":"2014-07-15","domains":["Vision"],"country":"China","reference":"https://arxiv.org/abs/1407.4023","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:dbb1ce69859f6fd4d9b3","name":"SmooCT","organization":"University College London (UCL)","date":"2014-07-01","domains":["Games"],"country":"United Kingdom","reference":"https://www.semanticscholar.org/paper/Self-play-Monte-Carlo-tree-search-in-computer-poker-Heinrich-Silver/7b687599b4425aa959036071030e1212a3b359c7","confidence":"Likely","accessibility":""},{"id":"epoch:2a14b45be254466a1728","name":"Multiresolution CNN","organization":"Google,Stanford University","date":"2014-06-23","domains":["Video"],"country":"United States of America,United States of America","reference":"https://ieeexplore.ieee.org/document/6909619","confidence":"","accessibility":""},{"id":"epoch:cdeee0ebfa29ae7c6f27","name":"DeepFace","organization":"Tel Aviv University,Facebook","date":"2014-06-23","domains":["Vision"],"country":"Israel,United States of America","reference":"https://ieeexplore.ieee.org/document/6909616","confidence":"Unknown","accessibility":""},{"id":"epoch:0c11c9c0ca4c5768469a","name":"RNN-WER","organization":"DeepMind,University of Toronto","date":"2014-06-22","domains":["Speech"],"country":"United Kingdom,Canada","reference":"https://proceedings.mlr.press/v32/graves14.html","confidence":"Likely","accessibility":""},{"id":"epoch:b9b7831a71a3ef061891","name":"Fragment embedding","organization":"Stanford University","date":"2014-06-21","domains":["Vision"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Deep-Fragment-Embeddings-for-Bidirectional-Image-Karpathy-Joulin/7f1b111f0bb703b0bd97aba505728a9b0d9b2a54","confidence":"Likely","accessibility":""},{"id":"epoch:fe7908d69393d9977497","name":"SPPNet","organization":"Microsoft,Xi\u2019an Jiaotong University,University of Science and Technology of China (USTC)","date":"2014-06-18","domains":["Vision"],"country":"United States of America,China,China","reference":"https://arxiv.org/abs/1406.4729","confidence":"Confident","accessibility":""},{"id":"epoch:dec08c8d37856be390cd","name":"GANs","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2014-06-10","domains":["Image generation"],"country":"Canada","reference":"https://arxiv.org/abs/1406.2661","confidence":"Speculative","accessibility":""},{"id":"epoch:a0be3e1415b179974701","name":"Two-stream ConvNets for action recognition","organization":"University of Oxford","date":"2014-06-09","domains":["Video"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1406.2199","confidence":"Unknown","accessibility":""},{"id":"epoch:828a3398f2bb60101582","name":"GRUs","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al,Jacobs University,University of Maine","date":"2014-06-03","domains":["Language"],"country":"Canada,Germany,United States of America","reference":"https://arxiv.org/abs/1406.1078","confidence":"Unknown","accessibility":""},{"id":"epoch:a593f9849bd1bd40e667","name":"Dropout: SVHN","organization":"University of Toronto","date":"2014-06-01","domains":["Vision"],"country":"Canada","reference":"https://jmlr.org/papers/v15/srivastava14a.html","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:19c449146516e6992a58","name":"AdaRNN","organization":"Beihang University","date":"2014-06-01","domains":["Language"],"country":"China","reference":"https://www.semanticscholar.org/paper/Adaptive-Recursive-Neural-Network-for-Twitter-Dong-Wei/06e122f475a21d92dba137609c40f35690217475","confidence":"Confident","accessibility":""},{"id":"epoch:08af608dfb9e451ca1a7","name":"Paragraph Vector","organization":"Google","date":"2014-05-14","domains":["Language"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Distributed-Representations-of-Sentences-and-Le-Mikolov/f3de86aeb442216a8391befcacb49e58b478f512","confidence":"Confident","accessibility":""},{"id":"epoch:b49e6c858b8b9bf8da07","name":"HyperNEAT","organization":"University of Texas at Austin","date":"2014-03-05","domains":["Games"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/abstract/document/6756960\n\nhttps://www.cs.utexas.edu/~mhauskn/papers/atari.pdf","confidence":"","accessibility":""},{"id":"epoch:43c9b860b18ef1c33d5f","name":"Clockwork RNN (CW-RNN)","organization":"IDSIA,SUPSI","date":"2014-02-14","domains":["Speech"],"country":"Switzerland,Switzerland","reference":"https://arxiv.org/abs/1402.3511","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c0a2f004a4a8b46951fb","name":"DBN for NLP","organization":"Microsoft,University of Toronto","date":"2014-02-11","domains":["Language"],"country":"United States of America,Canada","reference":"https://ieeexplore.ieee.org/document/6737243","confidence":"","accessibility":""},{"id":"epoch:10d3de97eae565375b83","name":"GloVe (6B)","organization":"Stanford University","date":"2014-01-01","domains":["Language"],"country":"United States of America","reference":"https://nlp.stanford.edu/projects/glove/","confidence":"","accessibility":""},{"id":"epoch:f9062e1f117d484c86ab","name":"GloVe (32B)","organization":"Stanford University","date":"2014-01-01","domains":["Language"],"country":"United States of America","reference":"https://nlp.stanford.edu/projects/glove/","confidence":"","accessibility":""},{"id":"epoch:e8a5da200494c212d6c7","name":"SPN-4","organization":"Singapore University of Technology & Design,DSO National Laboratories","date":"2014-01-01","domains":["Language"],"country":"Singapore,Singapore","reference":"https://www.isca-archive.org/interspeech_2014/cheng14_interspeech.html","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:61e5b87b841b0493492f","name":"OverFeat","organization":"New York University (NYU)","date":"2013-12-21","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1312.6229","confidence":"Confident","accessibility":""},{"id":"epoch:ba250ecd1b66281c4775","name":"Image generation","organization":"University of Amsterdam","date":"2013-12-20","domains":["Vision"],"country":"Netherlands","reference":"https://arxiv.org/abs/1312.6114","confidence":"Confident","accessibility":""},{"id":"epoch:14bf22551770cf46b69c","name":"DOT(S)-RNN","organization":"Aalto University,University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2013-12-20","domains":["Language"],"country":"Finland,Canada","reference":"https://arxiv.org/abs/1312.6026","confidence":"","accessibility":"Unreleased"},{"id":"epoch:7f1656b6a9283fc7e98b","name":"DQN","organization":"DeepMind","date":"2013-12-19","domains":["Games"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1312.5602","confidence":"Confident","accessibility":""},{"id":"epoch:7b049dc8ebe5627a528e","name":"Network in Network","organization":"National University of Singapore","date":"2013-12-16","domains":["Vision"],"country":"Singapore","reference":"https://arxiv.org/abs/1312.4400","confidence":"Unknown","accessibility":""},{"id":"epoch:946ad0c474552fe0ac29","name":"Deep RNN (PTB)","organization":"MetaMind Inc","date":"2013-12-11","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1609.07843","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:36a4487d2edbcee2be05","name":"RNN for 1B words","organization":"Google","date":"2013-12-11","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1312.3005","confidence":"Speculative","accessibility":""},{"id":"epoch:2958be60992099d90276","name":"DBLSTM","organization":"University of Toronto","date":"2013-12-08","domains":["Speech"],"country":"Canada","reference":"https://ieeexplore.ieee.org/document/6707742","confidence":"","accessibility":""},{"id":"epoch:a44a056467fef2b9d957","name":"TransE","organization":"Universite de Technologie de Compi\u00e8gne \u2013 CNRS,Google","date":"2013-12-05","domains":["Language"],"country":"France,United States of America","reference":"https://papers.nips.cc/paper/2013/hash/1cecc7a77928ca8133fa24680a88d2f9-Abstract.html","confidence":"Speculative","accessibility":""},{"id":"epoch:ad9933d2d599b7b0e337","name":"DeViSE","organization":"Google","date":"2013-12-05","domains":["Vision"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/DeViSE%3A-A-Deep-Visual-Semantic-Embedding-Model-Frome-Corrado/4aa4069693bee00d1b0759ca3df35e59284e9845","confidence":"Confident","accessibility":""},{"id":"epoch:6347b250ea287512080f","name":"TensorReasoner","organization":"Stanford University","date":"2013-12-01","domains":["Language"],"country":"United States of America","reference":"https://papers.nips.cc/paper/2013/hash/b337e84de8752b27eda3a12363109e80-Abstract.html","confidence":"Unknown","accessibility":""},{"id":"epoch:4baf3aedca4a24489855","name":"Visualizing CNNs","organization":"New York University (NYU)","date":"2013-11-12","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1311.2901","confidence":"Confident","accessibility":""},{"id":"epoch:82bb47558406835c8ec3","name":"R-CNN (T-net)","organization":"University of California (UC) Berkeley","date":"2013-11-11","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1311.2524","confidence":"","accessibility":""},{"id":"epoch:2602f44321318f038fee","name":"Word2Vec (small)","organization":"Google","date":"2013-10-16","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1310.4546","confidence":"","accessibility":""},{"id":"epoch:0fae6c7d89c0f9da6b8a","name":"Word2Vec (large)","organization":"Google","date":"2013-10-16","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1310.4546","confidence":"Likely","accessibility":""},{"id":"epoch:bd45a8049a01db40827c","name":"RNTN","organization":"Stanford University","date":"2013-10-01","domains":["Language"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Recursive-Deep-Models-for-Semantic-Compositionality-Socher-Perelygin/687bac2d3320083eb4530bf18bb8f8f721477600","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:495c80b2eafedced26ab","name":"RCTM","organization":"University of Oxford","date":"2013-10-01","domains":["Language"],"country":"United Kingdom","reference":"https://www.semanticscholar.org/paper/Recurrent-Continuous-Translation-Models-Kalchbrenner-Blunsom/944a1cfd79dbfb6fef460360a0765ba790f4027a","confidence":"Likely","accessibility":""},{"id":"epoch:e0a110107a98740518a8","name":"Mitosis","organization":"IDSIA","date":"2013-09-22","domains":["Vision","Medicine"],"country":"Switzerland","reference":"https://link.springer.com/chapter/10.1007/978-3-642-40763-5_51","confidence":"Likely","accessibility":""},{"id":"epoch:271819674708f308b1c8","name":"RNN+weight noise+dynamic eval","organization":"University of Toronto","date":"2013-08-04","domains":["Language"],"country":"Canada","reference":"https://arxiv.org/abs/1308.0850","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:710716e9ef61c94c0013","name":"Hierarchical Scene Labeling (Stanford Background)","organization":"New York University (NYU)","date":"2013-08-01","domains":["Vision"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/6338939","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:8b1b492d3999ec3d8fe5","name":"Fisher Vector image classifier","organization":"Universidad Nacional de Cordoba,Inteligent Systems Lab Amsterdam,University of Amsterdam,LEAR Team,INRIA,Xerox Research Centre Europe (XRCE)","date":"2013-06-12","domains":["Vision"],"country":"Argentina,Netherlands,Netherlands,France,France,France","reference":"https://hal.inria.fr/hal-00830491v2/document","confidence":"Confident","accessibility":""},{"id":"epoch:85c9abaf58e0a3364743","name":"SemVec","organization":"Microsoft Research","date":"2013-06-09","domains":["Language"],"country":"United States of America","reference":"https://www.aclweb.org/anthology/N13-1090/","confidence":"Unknown","accessibility":""},{"id":"epoch:2cb92b2ec3fcce530bb9","name":"Multilingual DNN","organization":"Google","date":"2013-05-26","domains":["Speech"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Multilingual-acoustic-models-using-distributed-deep-Heigold-Vanhoucke/a41b826d23957d6ad4e9e794d20a583a9b567c5d","confidence":"Confident","accessibility":""},{"id":"epoch:27379bc7c6dd47f14eba","name":"ReLU-Speech","organization":"Google,University of Toronto,New York University (NYU)","date":"2013-05-26","domains":["Speech"],"country":"United States of America,Canada,United States of America","reference":"https://www.semanticscholar.org/paper/On-rectified-linear-units-for-speech-processing-Zeiler-Ranzato/64da1980714cfc130632c5b92b9d98c2f6763de6","confidence":"Likely","accessibility":""},{"id":"epoch:431da863746e93b9ea53","name":"Selective Search","organization":"University of Trento,University of Amsterdam","date":"2013-04-02","domains":["Vision"],"country":"Italy,Netherlands","reference":"https://link.springer.com/article/10.1007/s11263-013-0620-5","confidence":"Unknown","accessibility":""},{"id":"epoch:262d094670f102b0038a","name":"PreTrans-3L-250H","organization":"University of Toronto","date":"2013-03-22","domains":["Speech"],"country":"Canada","reference":"https://arxiv.org/abs/1303.5778","confidence":"","accessibility":""},{"id":"epoch:e60fad3dcc42a7fa7614","name":"Maxout Networks","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2013-02-18","domains":["Vision"],"country":"Canada","reference":"https://arxiv.org/abs/1302.4389","confidence":"Unknown","accessibility":""},{"id":"epoch:22d2a8769808860c9e60","name":"Textual Imager","organization":"Stanford University","date":"2013-01-16","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1301.3666","confidence":"Unknown","accessibility":""},{"id":"epoch:57373bc5a8b64bd072c1","name":"DistBelief NNLM","organization":"Google","date":"2013-01-16","domains":["Language"],"country":"United States of America","reference":"https://arxiv.org/abs/1301.3781","confidence":"Likely","accessibility":""},{"id":"epoch:13f41055635fb7a76cef","name":"RNN (SGD+CLR)","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2012-12-14","domains":["Audio"],"country":"Canada","reference":"https://arxiv.org/abs/1212.0901","confidence":"Speculative","accessibility":""},{"id":"epoch:cb56b57edd4d51aeca34","name":"RNN (SGD+CLR) (PTB)","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2012-12-04","domains":["Language"],"country":"Canada","reference":"https://arxiv.org/abs/1212.0901","confidence":"","accessibility":"Unreleased"},{"id":"epoch:395611ef167d35db10cf","name":"DistBelief Speech","organization":"Google","date":"2012-12-03","domains":["Speech"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Large-Scale-Distributed-Deep-Networks-Dean-Corrado/3127190433230b3dc1abd0680bb58dced4bcd90e","confidence":"Speculative","accessibility":""},{"id":"epoch:55b6d46baf541acea0eb","name":"DistBelief Vision","organization":"Google","date":"2012-12-03","domains":["Vision"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Large-Scale-Distributed-Deep-Networks-Dean-Corrado/3127190433230b3dc1abd0680bb58dced4bcd90e","confidence":"Likely","accessibility":""},{"id":"epoch:cdde06923bf21b9cff67","name":"DNN EM segmentation","organization":"IDSIA,SUPSI","date":"2012-12-03","domains":["Vision"],"country":"Switzerland,Switzerland","reference":"https://www.semanticscholar.org/paper/Deep-Neural-Networks-Segment-Neuronal-Membranes-in-Ciresan-Giusti/09193e19b59fc8f05bee9d6efbfb1607ca5b6501","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:39ab14559c2f37edb7b8","name":"Bayesian automated hyperparameter tuning","organization":"University of Toronto,University of Sherbrooke,Harvard University","date":"2012-12-02","domains":["Other"],"country":"Canada,Canada,United States of America","reference":"https://arxiv.org/abs/1206.2944","confidence":"Unknown","accessibility":""},{"id":"epoch:af89bb791ca2b7b5d038","name":"RNN+LDA+KN5+cache","organization":"Microsoft,Brno University of Technology","date":"2012-12-01","domains":["Language"],"country":"United States of America,Czechia","reference":"https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/rnn_ctxt.pdf","confidence":"","accessibility":"Unreleased"},{"id":"epoch:e788091925495eb7ce0e","name":"RNN+LDA","organization":"Microsoft Research","date":"2012-12-01","domains":["Language"],"country":"United States of America","reference":"https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/rnn_ctxt.pdf","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:6af107c11383b4a89ca4","name":"RNN+LSA+KN5+cache (model combination w/ linear extrapolation)","organization":"Microsoft Research","date":"2012-12-01","domains":["Language"],"country":"United States of America","reference":"https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/rnn_ctxt.pdf","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:bf5e1d906f4c4c99fadf","name":"RNN","organization":"Microsoft Research","date":"2012-12-01","domains":["Language"],"country":"United States of America","reference":"https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/rnn_ctxt.pdf","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:3e49e54f8ed72a0a7d65","name":"AlexNet","organization":"University of Toronto","date":"2012-09-30","domains":["Vision"],"country":"Canada","reference":"https://proceedings.neurips.cc/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html","confidence":"Confident","accessibility":""},{"id":"epoch:72b3d8c9f7bec57f1863","name":"LSTM LM","organization":"RWTH Aachen University","date":"2012-09-09","domains":["Language"],"country":"Germany","reference":"https://www.semanticscholar.org/paper/LSTM-Neural-Networks-for-Language-Modeling-Sundermeyer-Schl%C3%BCter/f9a1b3850dfd837793743565a8af95973d395a4e","confidence":"Speculative","accessibility":""},{"id":"epoch:7fadecbb69d1d0ab3df2","name":"LSTM-300units","organization":"RWTH Aachen University","date":"2012-09-01","domains":["Language"],"country":"Germany","reference":"https://www.isca-archive.org/interspeech_2012/sundermeyer12_interspeech.pdf","confidence":"","accessibility":"Unreleased"},{"id":"epoch:4ecc1beed010a82841a2","name":"Context-dependent RNN","organization":"Microsoft Research,Brno University of Technology","date":"2012-07-27","domains":["Language"],"country":"United States of America,Czechia","reference":"https://www.microsoft.com/en-us/research/wp-content/uploads/2012/07/rnn_ctxt_TR.sav_.pdf","confidence":"Unknown","accessibility":""},{"id":"epoch:fa59fadcaecbe3f5d86d","name":"MV-RNN","organization":"Stanford University","date":"2012-07-12","domains":["Language"],"country":"United States of America","reference":"https://www.aclweb.org/anthology/D12-1110/","confidence":"","accessibility":""},{"id":"epoch:05fce1017812ee16dc8c","name":"Unsupervised High-level Feature Learner","organization":"Google","date":"2012-07-12","domains":["Vision"],"country":"United States of America","reference":"https://arxiv.org/abs/1112.6209","confidence":"Likely","accessibility":""},{"id":"epoch:2209b0b0a24135cd54de","name":"Ngram corpus","organization":"Google","date":"2012-07-08","domains":["Language"],"country":"United States of America","reference":"https://aclanthology.org/P12-3029/","confidence":"Unknown","accessibility":""},{"id":"epoch:d0875820c4200140e6fa","name":"LBL","organization":"University College London (UCL)","date":"2012-06-27","domains":["Language"],"country":"United Kingdom","reference":"https://arxiv.org/abs/1206.6426","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:75d041cff09f079bd928","name":"Dropout (ImageNet)","organization":"University of Toronto","date":"2012-06-03","domains":["Vision"],"country":"Canada","reference":"https://arxiv.org/abs/1207.0580","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:6fae4e450a83c461396e","name":"Dropout (CIFAR)","organization":"University of Toronto","date":"2012-06-03","domains":["Vision"],"country":"Canada","reference":"https://arxiv.org/abs/1207.0580","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:3c433c67bebfaa24b1e4","name":"Dropout (MNIST)","organization":"University of Toronto","date":"2012-06-03","domains":["Vision"],"country":"Canada","reference":"https://arxiv.org/abs/1207.0580","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:36e6935d60b6ec96ffd2","name":"Dropout (TIMIT)","organization":"University of Toronto","date":"2012-06-03","domains":["Speech"],"country":"Canada","reference":"https://arxiv.org/abs/1207.0580","confidence":"","accessibility":"Unreleased"},{"id":"epoch:830d937c53b74c54d267","name":"MCDNN (MNIST)","organization":"IDSIA,SUPSI","date":"2012-02-13","domains":["Vision"],"country":"Switzerland,Switzerland","reference":"https://arxiv.org/abs/1202.2745v1","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:2cd401f94c3c7ccd8451","name":"HOGWILD!","organization":"University of Wisconsin Madison","date":"2011-11-11","domains":["Other"],"country":"United States of America","reference":"https://arxiv.org/abs/1106.5730","confidence":"Unknown","accessibility":""},{"id":"epoch:441f4f3cbc90b61f3487","name":"NLP from scratch","organization":"NEC Laboratories,Princeton University","date":"2011-11-08","domains":["Language"],"country":"United States of America,United States of America","reference":"https://www.jmlr.org/papers/volume12/collobert11a/collobert11a.pdf","confidence":"","accessibility":""},{"id":"epoch:302fa58f1c2d3b4c0c3e","name":"Domain Adaptation","organization":"University of Maryland","date":"2011-11-06","domains":["Vision"],"country":"United States of America","reference":"http://ftp.idiap.ch/pub/courses/EE-700/material/05-12-2012/2011_ICCV_DomainAdaptation.pdf","confidence":"","accessibility":""},{"id":"epoch:3b9e1118c522f289e721","name":"Adaptive Subgrad","organization":"Technion - Israel Institute of Technology,Google,University of California (UC) Berkeley","date":"2011-10-03","domains":["Language"],"country":"Israel,United States of America,United States of America","reference":"https://dl.acm.org/doi/10.5555/1953048.2021068","confidence":"Unknown","accessibility":""},{"id":"epoch:951bc3b026abce7a5811","name":"CNN committee (traffic sign)","organization":"IDSIA","date":"2011-10-03","domains":["Vision"],"country":"Switzerland","reference":"https://www.semanticscholar.org/paper/A-committee-of-neural-networks-for-traffic-sign-Ciresan-Meier/dd7f8b53e6802787179a961e766760cbbe2d5011","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:9fbececd5f7f90dfe6b9","name":"CNN Committee (NIST)","organization":"IDSIA","date":"2011-09-18","domains":["Vision"],"country":"Switzerland","reference":"https://www.semanticscholar.org/paper/Convolutional-Neural-Network-Committees-for-Ciresan-Meier/260a7615bbffec052d67dffde5bcf9b4687b50ee","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:9d8c029b104589600617","name":"CNN Committee (MNIST)","organization":"IDSIA","date":"2011-09-18","domains":["Vision"],"country":"Switzerland","reference":"https://www.semanticscholar.org/paper/Convolutional-Neural-Network-Committees-for-Ciresan-Meier/260a7615bbffec052d67dffde5bcf9b4687b50ee","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:d66d97af5346024de338","name":"Bayesian Starcraft","organization":"Coll\u00e8ge de France","date":"2011-08-31","domains":["Games"],"country":"France","reference":"https://ieeexplore.ieee.org/document/6032006","confidence":"","accessibility":""},{"id":"epoch:f776b7ae1072c29e4593","name":"High Performance CNN (NORB)","organization":"IDSIA,SUPSI","date":"2011-07-16","domains":["Vision"],"country":"Switzerland,Switzerland","reference":"https://people.idsia.ch/~juergen/ijcai2011.pdf","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:c1bb1860245df316afa5","name":"Recursive sentiment autoencoder","organization":"Stanford University","date":"2011-07-01","domains":["Language"],"country":"United States of America","reference":"https://aclanthology.org/D11-1014/","confidence":"Unknown","accessibility":""},{"id":"epoch:ed75c46cf6ca0d947312","name":"Recursive Neural Network","organization":"Stanford University","date":"2011-06-28","domains":["Vision","Language"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Parsing-Natural-Scenes-and-Natural-Language-with-Socher-Lin/9c0ddf74f87d154db88d79c640578c1610451eec","confidence":"Confident","accessibility":""},{"id":"epoch:79fdb12eff0359b585e4","name":"Cross-Lingual POS Tagger","organization":"Carnegie Mellon University (CMU),Google Research","date":"2011-06-19","domains":["Language"],"country":"United States of America,United States of America","reference":"https://aclanthology.org/P11-1061/","confidence":"Unknown","accessibility":""},{"id":"epoch:79cb5c69783f02aa3e2a","name":"Vector Space Model","organization":"Stanford University","date":"2011-06-19","domains":["Language"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Learning-Word-Vectors-for-Sentiment-Analysis-Maas-Daly/1c61f9ef06fe74505775a833ff849185757199e7\n\nhttps://ai.stanford.edu/~ang/papers/acl11-WordVectorsSentimentAnalysis.pdf","confidence":"Confident","accessibility":""},{"id":"epoch:5eb91d6eabd160af4b30","name":"RNN-SpeedUp","organization":"Brno University of Technology,Johns Hopkins University","date":"2011-05-22","domains":["Language"],"country":"Czechia,United States of America","reference":"https://ieeexplore.ieee.org/document/5947611","confidence":"","accessibility":""},{"id":"epoch:505d6f10d591a5e755af","name":"Deep Autoencoders","organization":"University of Toronto","date":"2011-04-29","domains":["Vision"],"country":"Canada","reference":"https://www.semanticscholar.org/paper/Using-very-deep-autoencoders-for-content-based-Krizhevsky-Hinton/88080d28536f36588740737f3b7a1f6c1a409654","confidence":"Confident","accessibility":""},{"id":"epoch:b8901d6ad0866700b46a","name":"Deep rectifier networks","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2011-04-13","domains":["Vision"],"country":"Canada","reference":"https://proceedings.mlr.press/v15/glorot11a.html","confidence":"Unknown","accessibility":""},{"id":"epoch:0b9b11ff36990b5bdd7b","name":"Optimized Single-layer Net","organization":"University of Michigan,Stanford University","date":"2011-04-11","domains":["Vision"],"country":"United States of America,United States of America","reference":"http://proceedings.mlr.press/v15/coates11a.html","confidence":"Unknown","accessibility":""},{"id":"epoch:f16622f43e5d70b91f1f","name":"YouTube Video Recommendation System","organization":"Google","date":"2010-09-26","domains":["Recommendation"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.1145/1864708.1864770","confidence":"","accessibility":""},{"id":"epoch:33f4fb1d5b1bf0f678b9","name":"KN5 LM + RNN 400/10 (WSJ)","organization":"Brno University of Technology,Johns Hopkins University","date":"2010-09-26","domains":["Speech"],"country":"Czechia,United States of America","reference":"https://www.isca-archive.org/interspeech_2010/mikolov10_interspeech.html","confidence":"Confident","accessibility":""},{"id":"epoch:cdb5b14da7c422e7457a","name":"RNN 500/10 + RT09 LM (NIST RT05)","organization":"Brno University of Technology,Johns Hopkins University","date":"2010-09-26","domains":["Speech"],"country":"Czechia,United States of America","reference":"https://www.isca-archive.org/interspeech_2010/mikolov10_interspeech.html","confidence":"Confident","accessibility":""},{"id":"epoch:01a143b2abfe135c9962","name":"RNN LM","organization":"Johns Hopkins University","date":"2010-09-26","domains":["Language"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Recurrent-neural-network-based-language-model-Mikolov-Karafi%C3%A1t/9819b600a828a57e1cde047bbe710d3446b30da5","confidence":"Speculative","accessibility":""},{"id":"epoch:4115fb5b28203b6087b6","name":"RNN 1000/5 + RT09 LM (NIST RT05)","organization":"Brno University of Technology,Johns Hopkins University","date":"2010-09-26","domains":["Speech"],"country":"Czechia,United States of America","reference":"https://www.isca-archive.org/interspeech_2010/mikolov10_interspeech.html","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:b49596778c74b372b7bd","name":"Pooling CNN (Caltech 101)","organization":"University of Bonn","date":"2010-09-15","domains":["Vision"],"country":"Germany","reference":"https://www.semanticscholar.org/paper/Evaluation-of-Pooling-Operations-in-Convolutional-Scherer-M%C3%BCller/5d21006fa32ff69f6b0a646f26ce0db84f2f4d33","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:b9f7f445c0f7f29a85e6","name":"Pooling CNN (NORB)","organization":"University of Bonn","date":"2010-09-15","domains":["Vision"],"country":"Germany","reference":"https://www.semanticscholar.org/paper/Evaluation-of-Pooling-Operations-in-Convolutional-Scherer-M%C3%BCller/5d21006fa32ff69f6b0a646f26ce0db84f2f4d33","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:f792a5e27a304756267d","name":"Fisher-Boost","organization":"Xerox Research Centre Europe (XRCE)","date":"2010-09-05","domains":["Vision"],"country":"France","reference":"https://link.springer.com/chapter/10.1007/978-3-642-15561-1_11","confidence":"Unknown","accessibility":""},{"id":"epoch:e2439ccea891db0bfe8c","name":"SimuParallelSGD","organization":"Yahoo Research","date":"2010-07-01","domains":["Other"],"country":"United States of America","reference":"https://martin.zinkevich.org/publications/nips2010.pdf","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:1a856ba4b28496f0ce5f","name":"ReLU (LFW)","organization":"University of Toronto","date":"2010-06-15","domains":["Vision"],"country":"Canada","reference":"https://dl.acm.org/doi/10.5555/3104322.3104425","confidence":"Unknown","accessibility":""},{"id":"epoch:fa8a1743029cd23ce36e","name":"ReLU (NORB)","organization":"University of Toronto","date":"2010-06-15","domains":["Vision"],"country":"Canada","reference":"https://dl.acm.org/doi/10.5555/3104322.3104425","confidence":"","accessibility":""},{"id":"epoch:9eafd8f71ea10001e5d5","name":"Mid-level Features","organization":"INRIA,Ecole Normale Sup\u00e8rieure,New York University (NYU)","date":"2010-06-13","domains":["Vision"],"country":"France,France,United States of America","reference":"https://ieeexplore.ieee.org/document/5539963","confidence":"Unknown","accessibility":""},{"id":"epoch:bfe09a2ec14e53e69511","name":"Deconvolutional Network","organization":"New York University (NYU)","date":"2010-06-13","domains":["Vision"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/5539957","confidence":"Unknown","accessibility":""},{"id":"epoch:acd6ff85ad90fe9d5dde","name":"iCCCP","organization":"Massachusetts Institute of Technology (MIT)","date":"2010-06-13","domains":["Vision"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/5540096","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:8addd8256ea860b240e1","name":"Word Representations","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al,University of Illinois Urbana-Champaign (UIUC)","date":"2010-06-01","domains":["Language"],"country":"Canada,United States of America","reference":"https://aclanthology.org/P10-1040.pdf","confidence":"","accessibility":""},{"id":"epoch:06ebb3327828d3f534dc","name":"Feedforward NN","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2010-05-13","domains":["Vision"],"country":"Canada","reference":"https://proceedings.mlr.press/v9/glorot10a.html","confidence":"Confident","accessibility":""},{"id":"epoch:62ce646b920d980e9660","name":"6-layer MLP (MNIST)","organization":"IDSIA,University of Lugano,SUPSI","date":"2010-03-01","domains":["Vision"],"country":"Switzerland,Switzerland,Switzerland","reference":"https://arxiv.org/abs/1003.0358","confidence":"Likely","accessibility":""},{"id":"epoch:80c9d520e881eb091554","name":"Stacked Denoising Autoencoders","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al,University of Toronto","date":"2010-01-03","domains":["Other"],"country":"Canada,Canada","reference":"https://www.jmlr.org/papers/v11/vincent10a.html","confidence":"Unknown","accessibility":""},{"id":"epoch:422cacdf1478ebf21283","name":"Super-vector coding","organization":"University of Illinois Urbana-Champaign (UIUC),NEC Laboratories,Rutgers University","date":"2010-01-01","domains":["Vision"],"country":"United States of America,United States of America,United States of America","reference":"http://tongzhang-ml.org/papers/eccv10_supervect.pdf","confidence":"Speculative","accessibility":""},{"id":"epoch:dce3f8af9fd0463de2b7","name":"LCNP LabelMe","organization":"University of Bonn","date":"2009-11-22","domains":["Vision"],"country":"Germany","reference":"https://ieeexplore.ieee.org/document/5357786","confidence":"Confident","accessibility":""},{"id":"epoch:dfdd47a7ee380f6b739f","name":"LCNP MNIST","organization":"","date":"2009-11-22","domains":["Vision"],"country":"","reference":"https://ieeexplore.ieee.org/document/5357786","confidence":"Confident","accessibility":""},{"id":"epoch:f4f594efc8a69806e619","name":"LCNP NORB","organization":"","date":"2009-11-22","domains":["Vision"],"country":"","reference":"https://ieeexplore.ieee.org/document/5357786","confidence":"Confident","accessibility":""},{"id":"epoch:06cfa9f338e835733a9c","name":"3D city reconstruction","organization":"University of Washington,Microsoft Research,Cornell University","date":"2009-09-29","domains":["3D modeling"],"country":"United States of America,United States of America,United States of America","reference":"https://grail.cs.washington.edu/rome/","confidence":"Unknown","accessibility":""},{"id":"epoch:24804dc11dae950e5e99","name":"BellKor 2007","organization":"AT&T","date":"2009-09-21","domains":["Recommendation"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/The-BellKor-solution-to-the-Netflix-Prize-Bell-Koren/f4ebb542c752a0dc423f94fd121e2edb8f6275ba","confidence":"","accessibility":""},{"id":"epoch:4418b452d1a628713723","name":"Polarity Classifier","organization":"University of Edinburgh,University of Pittsburgh","date":"2009-09-01","domains":["Language"],"country":"United Kingdom,United States of America","reference":"https://aclanthology.org/J09-3003.pdf","confidence":"","accessibility":""},{"id":"epoch:0a01c76642f06f2f32b1","name":"Two Stage Feature Extraction (MNIST)","organization":"New York University (NYU)","date":"2009-09-01","domains":["Vision"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/What-is-the-best-multi-stage-architecture-for-Jarrett-Kavukcuoglu/1f88427d7aa8225e47f946ac41a0667d7b69ac52","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:1e4620607069eff00ae8","name":"ConvNet Processor","organization":"Courant Institute of Mathematical Sciences","date":"2009-08-31","domains":["Vision"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/abstract/document/5272559/","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:9339d95a08f319aa9864","name":"MatrixFac for Recommenders","organization":"Yahoo Research,AT&T","date":"2009-08-07","domains":["Recommendation"],"country":"United States of America,United States of America","reference":"https://ieeexplore.ieee.org/document/5197422","confidence":"","accessibility":""},{"id":"epoch:bcc2677be03d62160cb2","name":"RL mapping instructions (troubleshooting)","organization":"Massachusetts Institute of Technology (MIT)","date":"2009-08-02","domains":["Language"],"country":"United States of America","reference":"https://aclanthology.org/P09-1010/","confidence":"","accessibility":""},{"id":"epoch:1067b383c62f8bbee54c","name":"BellKor 2009","organization":"AT&T","date":"2009-08-01","domains":["Recommendation"],"country":"United States of America","reference":"https://www2.seas.gwu.edu/~simhaweb/champalg/cf/papers/KorenBellKor2009.pdf","confidence":"","accessibility":""},{"id":"epoch:be63886e49ebd2a63275","name":"RL mapping instructions (games)","organization":"Massachusetts Institute of Technology (MIT)","date":"2009-08-01","domains":["Language"],"country":"United States of America","reference":"https://aclanthology.org/P09-1010/","confidence":"","accessibility":""},{"id":"epoch:fe5b262103b777b765d9","name":"BellKor 2008","organization":"AT&T","date":"2009-08-01","domains":["Recommendation"],"country":"United States of America","reference":"https://cseweb.ucsd.edu/classes/fa17/cse291-b/reading/ProgressPrize2008_BellKor.pdf","confidence":"","accessibility":""},{"id":"epoch:2344659ea57ed6e20353","name":"Pragmatic Theory solution (Netflix 2009)","organization":"Pragmatic Theory Inc.","date":"2009-08-01","domains":["Recommendation"],"country":"Canada","reference":"https://www.asc.ohio-state.edu/statistics/statgen/joul_aut2009/PragmaticTheory.pdf","confidence":"Confident","accessibility":""},{"id":"epoch:b8494a05b082956ffc5a","name":"BigChaos OptiBlend","organization":"AT&T","date":"2009-08-01","domains":["Recommendation"],"country":"United States of America","reference":"https://www.asc.ohio-state.edu/statistics/statgen/joul_aut2009/BigChaos.pdf","confidence":"","accessibility":""},{"id":"epoch:530fc86dbfb4eb0cb890","name":"Conditional Maximum Entropy Model (Gigaworld)","organization":"Google","date":"2009-07-01","domains":["Language"],"country":"United States of America","reference":"https://proceedings.neurips.cc/paper_files/paper/2009/hash/d81f9c1be2e08964bf9f24b15f0e4900-Abstract.html","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:498d6dfc7c6cbe1b61e3","name":"GPU DBNs","organization":"Stanford University","date":"2009-06-15","domains":["Other"],"country":"United States of America","reference":"https://dl.acm.org/doi/abs/10.1145/1553374.1553486?utm_campaign=The+Batch&utm_source=hs_email&utm_medium=email&_hsenc=p2ANqtz-95Z7-X4Dl-RJK6gYKvjyDIrYaGhBeqWoc0ldqyPEKni0Ip6UE7452hr-ygY52z00LBpYgM","confidence":"Confident","accessibility":""},{"id":"epoch:66842daafef7a13497aa","name":"Conv-DBN","organization":"Stanford University","date":"2009-06-14","domains":["Vision"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.1145/1553374.1553453","confidence":"Unknown","accessibility":""},{"id":"epoch:b6d99624eaa34539dc09","name":"Deep Boltzmann Machines","organization":"University of Toronto","date":"2009-04-16","domains":["Other"],"country":"Canada","reference":"https://www.sciencedirect.com/topics/computer-science/deep-boltzmann-machine","confidence":"Unknown","accessibility":""},{"id":"epoch:90d64ce71255a7bf2b0d","name":"RBM Image Classifier","organization":"University of Toronto","date":"2009-04-08","domains":["Vision"],"country":"Canada","reference":"https://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf","confidence":"Likely","accessibility":""},{"id":"epoch:3b4938279bbaa3b12ed8","name":"Long-Range Autonomous Off-Road Driving System","organization":"Courant Institute of Mathematical Sciences","date":"2009-01-08","domains":["Vision","Robotics","Driving"],"country":"United States of America","reference":"https://onlinelibrary.wiley.com/doi/abs/10.1002/rob.20276","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c8af69955a38c7a18cb6","name":"BP-DBN","organization":"University of Toronto","date":"2009-01-01","domains":["Speech"],"country":"Canada","reference":"https://www.cs.utoronto.ca/~gdahl/papers/dbnPhoneRec.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:15867d099fb6c4ac3eae","name":"Semantic Hashing","organization":"University of Toronto","date":"2008-12-10","domains":["Language"],"country":"Canada","reference":"https://www.cs.cmu.edu/~rsalakhu/papers/sdarticle.pdf","confidence":"","accessibility":""},{"id":"epoch:7a5c6656d3323966d0fe","name":"GNN","organization":"University of Siena","date":"2008-12-09","domains":["Other"],"country":"Italy","reference":"https://ieeexplore.ieee.org/document/4700287","confidence":"Confident","accessibility":"Open weights (unrestricted)"},{"id":"epoch:8c0cd59c95285bce7274","name":"HLBL","organization":"University of Toronto","date":"2008-12-08","domains":["Language"],"country":"Canada","reference":"https://www.semanticscholar.org/paper/A-Scalable-Hierarchical-Distributed-Language-Model-Mnih-Hinton/a9fc84f8abe740cdc7ee82e69444d1d00dbe0ceb","confidence":"Confident","accessibility":""},{"id":"epoch:e1bd3b53ed427475e4d4","name":"ADAPTIVE NLPM","organization":"University of Toronto","date":"2008-12-08","domains":["Language"],"country":"Canada","reference":"https://papers.nips.cc/paper_files/paper/2008/file/1e056d2b0ebd5c878c550da6ac5d3724-Paper.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:a8417f431f1e39e415ae","name":"BigChaos 2008","organization":"AT&T","date":"2008-11-25","domains":["Recommendation"],"country":"United States of America","reference":"https://www.asc.ohio-state.edu/statistics/statgen/joul_aut2009/BigChaos.pdf","confidence":"","accessibility":""},{"id":"epoch:35c43e05928193002785","name":"Sparse digit recognition SVM","organization":"University of Lubeck","date":"2008-11-19","domains":["Vision"],"country":"Germany","reference":"https://pubmed.ncbi.nlm.nih.gov/19000969/","confidence":"Unknown","accessibility":""},{"id":"epoch:2963d133f581c689cbe0","name":"Boss (DARPA Urban Challenge)","organization":"Carnegie Mellon University (CMU)","date":"2008-07-23","domains":["Driving"],"country":"United States of America","reference":"https://onlinelibrary.wiley.com/doi/abs/10.1002/rob.20255","confidence":"Unknown","accessibility":""},{"id":"epoch:56cece334b04687a5846","name":"Stacked Semisuperviser Autoencoders","organization":"New York University (NYU),Microsoft","date":"2008-07-15","domains":["Language"],"country":"United States of America,United States of America","reference":"https://dl.acm.org/doi/10.1145/1390156.1390256","confidence":"","accessibility":""},{"id":"epoch:74b50ab1d84194f44ac1","name":"Denoising Autoencoders","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2008-07-05","domains":["Other"],"country":"Canada","reference":"https://dl.acm.org/doi/10.1145/1390156.1390294","confidence":"Unknown","accessibility":""},{"id":"epoch:c154b18f6e35712da8ac","name":"Semi-Supervised Embedding for DL","organization":"Google,NUANCE Communications,IDIAP,University of Illinois Urbana-Champaign (UIUC)","date":"2008-07-05","domains":["Other"],"country":"United States of America,United States of America,Switzerland,United States of America","reference":"https://dl.acm.org/doi/10.1145/1390156.1390303","confidence":"Unknown","accessibility":""},{"id":"epoch:758f186163c9b18fb325","name":"Deep Multitask NLP Network","organization":"NEC Laboratories","date":"2008-07-05","domains":["Language"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.1145/1390156.1390177","confidence":"Speculative","accessibility":""},{"id":"epoch:3b5950c3386992eb7cd6","name":"Multiscale deformable part model","organization":"UC Irvine,University of Chicago,Toyota Technological Institute at Chicago","date":"2008-06-23","domains":["Vision"],"country":"United States of America,United States of America,United States of America","reference":"https://ieeexplore.ieee.org/abstract/document/4587597","confidence":"Unknown","accessibility":""},{"id":"epoch:13477e6c51b8420a0fd1","name":"BLSTM for handwriting (2)","organization":"University of Bern,IDSIA,Technical University of Munich","date":"2007-12-03","domains":["Vision"],"country":"Switzerland,Switzerland,Germany","reference":"https://proceedings.neurips.cc/paper/2007/hash/4b0250793549726d5c1ea3906726ebfe-Abstract.html","confidence":"","accessibility":""},{"id":"epoch:1197e22df44681b18f78","name":"Enhanced Neighborhood-Based Filtering","organization":"AT&T","date":"2007-10-28","domains":["Recommendation"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/abstract/document/4470228","confidence":"Unknown","accessibility":""},{"id":"epoch:6c8e1affaabc31b5a060","name":"BLSTM for handwriting (1)","organization":"University of Bern,IDSIA,Technical University of Munich","date":"2007-09-23","domains":["Vision"],"country":"Switzerland,Switzerland,Germany","reference":"https://people.idsia.ch//~juergen/icdar_2007.pdf","confidence":"Unknown","accessibility":""},{"id":"epoch:31a14606103ec8eeb2d7","name":"Regularized SVD for Collaborative Filtering","organization":"Warsaw University","date":"2007-08-12","domains":["Recommendation"],"country":"Poland","reference":"https://www.semanticscholar.org/paper/Improving-regularized-singular-value-decomposition-Paterek/f732d0f69fe4e84a95c32706b28b9e4ef1753c61","confidence":"","accessibility":""},{"id":"epoch:26efaa8d08ad643899e2","name":"Fisher Kernel GMM","organization":"Xerox","date":"2007-07-16","domains":["Vision"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/4270291","confidence":"Likely","accessibility":""},{"id":"epoch:ee300ff552f927593afb","name":"SB-LM","organization":"Google","date":"2007-06-22","domains":["Language"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Large-Language-Models-in-Machine-Translation-Brants-Popat/ba786c46373892554b98df42df7af6f5da343c9d","confidence":"Likely","accessibility":""},{"id":"epoch:2d0a7016a18ff0ea8623","name":"KN-LM","organization":"Google","date":"2007-06-22","domains":["Language"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Large-Language-Models-in-Machine-Translation-Brants-Popat/ba786c46373892554b98df42df7af6f5da343c9d","confidence":"Likely","accessibility":""},{"id":"epoch:f66e44f4f7ebdf2c04f1","name":"Restricted Bolzmann machines","organization":"University of Toronto","date":"2007-06-20","domains":["Recommendation"],"country":"Canada","reference":"https://dl.acm.org/doi/abs/10.1145/1273496.1273596?casa_token=cfdkH2x12MwAAAAA:sEUzfllIGyPcOfzgUoDPHlpC1ukfCAo8ewocBXWBswIIF9eS5HdFo30nOtfmIV8gm-XpBpQJJ5zYVO8","confidence":"","accessibility":""},{"id":"epoch:88c04e685b90ab3299a5","name":"\u03bb-WASP","organization":"UT Austin","date":"2007-06-01","domains":["Language"],"country":"United States of America","reference":"https://www.aclweb.org/anthology/P07-1121/","confidence":"","accessibility":""},{"id":"epoch:93e65d3e3567748bba29","name":"Empirical evaluation of deep architectures","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2007-06-01","domains":["Other"],"country":"Canada","reference":"https://dl.acm.org/doi/10.1145/1273496.1273556","confidence":"Unknown","accessibility":""},{"id":"epoch:4d559fed9af054cc418f","name":"Greedy layer-wise DNN training","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2006-12-04","domains":["Other"],"country":"Canada","reference":"https://dl.acm.org/doi/10.5555/2976456.2976476","confidence":"Unknown","accessibility":""},{"id":"epoch:92ed40669b77002bb046","name":"Sparse Energy-Based Model","organization":"New York University (NYU)","date":"2006-12-04","domains":["Vision"],"country":"United States of America","reference":"https://papers.nips.cc/paper/2006/hash/87f4d79e36d68c3031ccf6c55e9bbd39-Abstract.html","confidence":"","accessibility":""},{"id":"epoch:d48e6cdf0871a2ca338b","name":"Local Binary Patterns for facial recognition","organization":"University of Oulu,IEEE","date":"2006-12-01","domains":["Vision"],"country":"Finland,Multinational","reference":"http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.456.1094&rep=rep1&type=pdf","confidence":"Unknown","accessibility":""},{"id":"epoch:501c2f63a2858e53274f","name":"Sparse Vision Encoding","organization":"Stanford University","date":"2006-11-01","domains":["Vision"],"country":"United States of America","reference":"https://proceedings.neurips.cc/paper_files/paper/2006/file/2d71b2ae158c7c5912cc0bbde2bb9d95-Paper.pdf","confidence":"Likely","accessibility":"Unreleased"},{"id":"epoch:1972c522bdb35a77f414","name":"Deep Belief Nets","organization":"University of Toronto,National University of Singapore","date":"2006-07-18","domains":["Vision"],"country":"Canada,Singapore","reference":"https://www.cs.toronto.edu/~hinton/absps/fastnc.pdf","confidence":"","accessibility":""},{"id":"epoch:bc3a24ed42ddebf3703d","name":"Dimensionality Reduction","organization":"University of Toronto","date":"2006-07-18","domains":["Vision"],"country":"Canada","reference":"https://www.cs.toronto.edu/~hinton/science.pdf","confidence":"","accessibility":""},{"id":"epoch:6b28d493b46c9ede50bc","name":"Semantic Taxonomy Induction","organization":"Stanford University","date":"2006-07-07","domains":["Language"],"country":"United States of America","reference":"https://www.aclweb.org/anthology/P06-1101/","confidence":"","accessibility":""},{"id":"epoch:4ac58152f97a3a2a9562","name":"CTC-Trained LSTM","organization":"IDSIA,Technical University of Munich","date":"2006-06-25","domains":["Speech"],"country":"Switzerland,Germany","reference":"https://www.cs.toronto.edu/~graves/icml_2006.pdf","confidence":"","accessibility":""},{"id":"epoch:43a9ac46566eb3d76841","name":"DrLIM","organization":"New York University (NYU)","date":"2006-06-17","domains":["Other"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/1640964","confidence":"","accessibility":""},{"id":"epoch:47edd45ce6c6c26d9229","name":"Spatial Pyramid Matching","organization":"INRIA,University of Illinois Urbana-Champaign (UIUC),Ecole Normale Sup\u00e8rieure","date":"2006-06-17","domains":["Vision"],"country":"France,United States of America,France","reference":"https://inc.ucsd.edu/mplab/users/marni/Igert/Lazebnik_06.pdf","confidence":"Unknown","accessibility":""},{"id":"epoch:1d4f8277a1ce025daa71","name":"Hybrid CNN/SVM Object Categorizer","organization":"Courant Institute of Mathematical Sciences","date":"2006-06-17","domains":["Vision"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/1640771/","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:1cc5828a892ec600dbb5","name":"SVM-CNN","organization":"New York University (NYU)","date":"2006-06-17","domains":["Vision"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Large-scale-Learning-with-SVM-and-Convolutional-for-Huang-LeCun/cf03fdf52dd6e4249cbbdbd0bffbbbe5ca389feb","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c1842c2b84ee46a76229","name":"Crazy Stone","organization":"INRIA","date":"2006-05-29","domains":["Games"],"country":"France","reference":"https://inria.hal.science/inria-00116992v1","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:a1c18e87d859b802359f","name":"FAST","organization":"University of Cambridge","date":"2006-05-07","domains":["Video"],"country":"United Kingdom","reference":"https://link.springer.com/chapter/10.1007/11744023_34","confidence":"Unknown","accessibility":""},{"id":"epoch:ee215b23bc8ce857cc16","name":"RL for helicopter flight","organization":"University of California (UC) Berkeley,Stanford University","date":"2006-03-09","domains":["Driving"],"country":"United States of America,United States of America","reference":"https://papers.nips.cc/paper/2003/hash/b427426b8acd2c2e53827970f2c2f526-Abstract.html","confidence":"Unknown","accessibility":""},{"id":"epoch:b0746dd307615bd2963c","name":"TFE SVM","organization":"Centre de Recherche en Automatique de Nancy (CRAN),CENPARMI","date":"2006-02-02","domains":["Vision"],"country":"France,Canada","reference":"https://hal.archives-ouvertes.fr/hal-00018426/en","confidence":"Unknown","accessibility":""},{"id":"epoch:1045dbfc7e00a1a995cc","name":"Stanley (DARPA Grand Challenge 2)","organization":"Stanford University","date":"2006-01-01","domains":["Driving"],"country":"United States of America","reference":"http://robots.stanford.edu/papers/thrun.stanley05.pdf","confidence":"Unknown","accessibility":""},{"id":"epoch:58aabc5f90e48ac05729","name":"Vision-based obstacle avoidance system (2005)","organization":"New York University (NYU),Net-Scale technologies,NEC Laboratories","date":"2005-12-05","domains":["Robotics","Vision"],"country":"United States of America,United States of America,United States of America","reference":"https://papers.nips.cc/paper_files/paper/2005/file/fdf1bc5669e8ff5ba45d02fded729feb-Paper.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:eeb4df4276f1edf266ca","name":"Monocular Depth Prediction","organization":"Stanford University","date":"2005-12-05","domains":["Vision"],"country":"United States of America","reference":"https://papers.nips.cc/paper_files/paper/2005/file/17d8da815fa21c57af9829fb0a869602-Paper.pdf","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:8e0fba30f7ef65750711","name":"RankNet","organization":"Microsoft Research,Microsoft","date":"2005-08-07","domains":["Search"],"country":"United States of America,United States of America","reference":"https://dl.acm.org/doi/abs/10.1145/1102351.1102363","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:598f12fea81bfda84752","name":"BiLSTM for Speech","organization":"IDSIA,Technical University of Munich","date":"2005-08-01","domains":["Speech"],"country":"Switzerland,Germany","reference":"https://www.sciencedirect.com/science/article/abs/pii/S0893608005001206","confidence":"Likely","accessibility":""},{"id":"epoch:276d7807cd88f0922ee8","name":"Histograms of Oriented Gradients","organization":"INRIA","date":"2005-06-25","domains":["Vision"],"country":"France","reference":"https://ieeexplore.ieee.org/document/1467360","confidence":"","accessibility":""},{"id":"epoch:38aa0b8f9d50ae1df0e2","name":"ConvNet similarity metric","organization":"New York University (NYU)","date":"2005-06-20","domains":["Vision"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/1467314","confidence":"","accessibility":""},{"id":"epoch:addd2baa4122cdd12fd4","name":"Hiero","organization":"University of Maryland","date":"2005-06-01","domains":["Language"],"country":"United States of America","reference":"https://aclanthology.org/P05-1033/","confidence":"","accessibility":""},{"id":"epoch:528b0d90f6a6e3cb2cf0","name":"SACHS","organization":"Massachusetts Institute of Technology (MIT),Stanford University","date":"2005-04-22","domains":["Biology"],"country":"United States of America,United States of America","reference":"https://science.sciencemag.org/content/308/5721/523.long","confidence":"","accessibility":""},{"id":"epoch:4f2d7f27d6bafa738ca1","name":"Hierarchical LM","organization":"","date":"2005-01-06","domains":["Language"],"country":"","reference":"https://www.semanticscholar.org/paper/Hierarchical-Probabilistic-Neural-Network-Language-Morin-Bengio/c19fbefdeead6a4154a22a9c8551a18b1530033a","confidence":"Confident","accessibility":""},{"id":"epoch:ba027c3bd984f8a4196b","name":"LMICA","organization":"","date":"2004-12-01","domains":["Vision"],"country":"","reference":"https://www.semanticscholar.org/paper/Linear-Multilayer-Independent-Component-Analysis-Matsuda-Yamaguchi/7061b01572fbff2e223ce3abb59f397895b1ebf1","confidence":"Confident","accessibility":""},{"id":"epoch:156f3c211696e8d58c19","name":"Synergistic Face Detector","organization":"NEC Laboratories,Courant Institute of Mathematical Sciences","date":"2004-12-01","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://proceedings.neurips.cc/paper/2004/hash/06c284d3f757b15c02f47f3ff06dc275-Abstract.html","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:c845fe1beadf86f6f17c","name":"LIRA","organization":"Instituto de Ciencias Aplicadas y Technologia","date":"2004-07-30","domains":["Vision"],"country":"Mexico","reference":"https://www.sciencedirect.com/science/article/abs/pii/S0262885604000721","confidence":"","accessibility":""},{"id":"epoch:d2c3d4e9bf3856933176","name":"Automated WSD via WordNet","organization":"University of Sussex","date":"2004-07-01","domains":["Language"],"country":"United Kingdom","reference":"https://aclanthology.org/P04-1036/","confidence":"","accessibility":""},{"id":"epoch:fe68749ff54a4dfc7777","name":"Invariant CNN","organization":"New York University (NYU)","date":"2004-06-27","domains":["Vision"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Learning-methods-for-generic-object-recognition-to-LeCun-Huang/f354310098e09c1e1dc88758fca36767fd9d084d","confidence":"Confident","accessibility":""},{"id":"epoch:b4d79c9249fba09d41c3","name":"Sandstorm (DARPA Grand Challenge I)","organization":"Carnegie Mellon University (CMU)","date":"2004-06-14","domains":["Driving"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/1336386","confidence":"Unknown","accessibility":""},{"id":"epoch:1faaaf34b09c7a43ed1f","name":"GPU implementation of neural networks","organization":"Soongsil University","date":"2004-06-01","domains":["Vision"],"country":"South Korea","reference":"https://www.sciencedirect.com/science/article/pii/S0031320304000524","confidence":"Unknown","accessibility":""},{"id":"epoch:e412925ba50b887ca0da","name":"Max-Margin Markov Networks","organization":"Stanford University","date":"2004-03-01","domains":["Vision","Language"],"country":"United States of America","reference":"https://papers.nips.cc/paper/2003/file/878d5691c824ee2aaf770f7d36c151d6-Paper.pdf","confidence":"","accessibility":""},{"id":"epoch:2ae4552bd462493421fe","name":"RankBoost (EachMovie)","organization":"Columbia University,Princeton University,Hebrew University of Jerusalem","date":"2003-11-15","domains":["Recommendation"],"country":"United States of America,United States of America,Israel","reference":"https://www.schapire.net/papers/FreundIyScSi98.pdf","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:b7f8028b9d107056210d","name":"RankBoost (meta-search)","organization":"Columbia University,Princeton University,Hebrew University of Jerusalem","date":"2003-11-15","domains":["Recommendation"],"country":"United States of America,United States of America,Israel","reference":"https://www.schapire.net/papers/FreundIyScSi98.pdf","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:c43dd3e0e215f8343864","name":"Bayesian object categorizer","organization":"California Institute of Technology,University of Oxford","date":"2003-10-13","domains":["Vision"],"country":"United States of America,United Kingdom","reference":"http://vision.stanford.edu/documents/Fei-Fei_ICCV03.pdf","confidence":"Speculative","accessibility":""},{"id":"epoch:05bafd6947ba23b03ef3","name":"CNN Best Practices","organization":"Microsoft Research","date":"2003-08-06","domains":["Vision"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/1227801","confidence":"","accessibility":""},{"id":"epoch:23554230414c68ed8e37","name":"Unsupervised Scale-Invariant Learning","organization":"University of Oxford","date":"2003-06-18","domains":["Vision"],"country":"United Kingdom","reference":"https://ieeexplore.ieee.org/document/1211479","confidence":"","accessibility":""},{"id":"epoch:11c91a922dec71922496","name":"Phrase-based translation","organization":"University of Southern California","date":"2003-05-01","domains":["Language"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.3115/1073445.1073462","confidence":"","accessibility":""},{"id":"epoch:f5b4e51003e9e0456150","name":"NPLM (Brown)","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2003-03-15","domains":["Language"],"country":"Canada","reference":"https://dl.acm.org/doi/10.5555/944919.944966","confidence":"Confident","accessibility":""},{"id":"epoch:ff469bda36852d191860","name":"NPLM (AP News)","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2003-03-15","domains":["Language"],"country":"Canada","reference":"https://dl.acm.org/doi/10.5555/944919.944966","confidence":"Confident","accessibility":""},{"id":"epoch:8add13a0bc6dde889dc3","name":"LDA","organization":"Stanford University","date":"2003-02-02","domains":["Language"],"country":"United States of America","reference":"https://jmlr.org/papers/volume3/blei03a/blei03a.pdf","confidence":"Unknown","accessibility":""},{"id":"epoch:933f13c0e2b57edfd87f","name":"Statistical Shape Constellations","organization":"California Institute of Technology","date":"2003-01-01","domains":["Vision"],"country":"United States of America","reference":"https://link.springer.com/content/pdf/10.1007/3-540-45054-8_2.pdf","confidence":"Unknown","accessibility":""},{"id":"epoch:7d29e4d6f8f3b81abd77","name":"Web mining + Decision tree recommender","organization":"Korea Advanced Institute of Science and Technology (KAIST)","date":"2002-10-01","domains":["Recommendation"],"country":"South Korea","reference":"https://reader.elsevier.com/reader/sd/pii/S0957417402000520?token=155B6D1937982D7D0271AFD1CFB034DFD7F3D1DE816B66C025EBC9D0A305BA6DA685DD62989DC05246C794CAC74CDAEF&originRegion=us-east-1&originCreation=20220325235441","confidence":"Unknown","accessibility":""},{"id":"epoch:2abf6f7b506b60e16af7","name":"Maximum Entropy Models for machine translation","organization":"University of Southern California,RWTH Aachen University","date":"2002-07-06","domains":["Language"],"country":"United States of America,Germany","reference":"https://aclanthology.org/P02-1038/","confidence":"","accessibility":""},{"id":"epoch:77d2426c512e63de3300","name":"Tagging via Viterbi Decoding","organization":"AT&T","date":"2002-06-01","domains":["Language"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.3115/1118693.1118694","confidence":"Unknown","accessibility":""},{"id":"epoch:8121ac3a924498d97dc8","name":"NEAT","organization":"UT Austin","date":"2002-06-01","domains":["Other"],"country":"United States of America","reference":"https://direct.mit.edu/evco/article/10/2/99/1123/Evolving-Neural-Networks-through-Augmenting","confidence":"Unknown","accessibility":""},{"id":"epoch:30b86680fa6ea57575cb","name":"Joint Probability Machine Translation","organization":"University of Southern California","date":"2002-06-01","domains":["Language"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.3115/1118693.1118711","confidence":"","accessibility":""},{"id":"epoch:4bd418016005a60d2cca","name":"Thumbs Up?","organization":"Cornell University,IBM","date":"2002-05-28","domains":["Language"],"country":"United States of America,United States of America","reference":"https://arxiv.org/abs/cs/0205070","confidence":"","accessibility":""},{"id":"epoch:683218d594ab9480593c","name":"Decision tree (classification)","organization":"Mitsubishi Electric Research Labs,Compaq CRL","date":"2001-12-08","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://www.cs.cmu.edu/~efros/courses/LBMV07/Papers/viola-cvpr-01.pdf","confidence":"Likely","accessibility":""},{"id":"epoch:4c2fd79b3827ee7e65d5","name":"Gradient Boosting Machine","organization":"Stanford University","date":"2001-10-01","domains":["Mathematics"],"country":"United States of America","reference":"https://projecteuclid.org/journals/annals-of-statistics/volume-29/issue-5/Greedy-function-approximation-A-gradient-boostingmachine/10.1214/aos/1013203451.full","confidence":"Unknown","accessibility":""},{"id":"epoch:b2ad02b4268af9c903b5","name":"Immediate trihead","organization":"Brown University","date":"2001-07-06","domains":["Language"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.3115/1073012.1073029","confidence":"Unknown","accessibility":""},{"id":"epoch:c8146f9536b035511615","name":"Restricted Boltzmann machine for Face Recognition","organization":"University of Toronto,University College London (UCL)","date":"2001-04-01","domains":["Vision"],"country":"Canada,United Kingdom","reference":"https://proceedings.neurips.cc/paper_files/paper/2000/hash/c366c2c97d47b02b24c3ecade4c40a01-Abstract.html","confidence":"Confident","accessibility":""},{"id":"epoch:405fcd55f761542c7271","name":"PoE MNIST","organization":"University College London (UCL)","date":"2000-11-28","domains":["Vision"],"country":"United Kingdom","reference":"https://proceedings.neurips.cc/paper_files/paper/2000/file/1f1baa5b8edac74eb4eaa329f14a0361-Paper.pdf","confidence":"Confident","accessibility":""},{"id":"epoch:e86cc4e831699f173561","name":"Neural LM","organization":"University of Montreal / Universit\u00e9 de Montr\u00e9al","date":"2000-11-28","domains":["Language"],"country":"Canada","reference":"https://papers.nips.cc/paper_files/paper/2000/file/728f206c2a01bf572b5940d7d9a8fa4c-Paper.pdf","confidence":"Confident","accessibility":""},{"id":"epoch:f6977f62ea6ffda7569a","name":"FrameNet role labeling","organization":"University of Rochester","date":"2000-09-01","domains":["Language"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.1162/089120102760275983","confidence":"","accessibility":""},{"id":"epoch:f3017d43eb2c064d5ba4","name":"Peephole LSTM","organization":"IDSIA","date":"2000-07-26","domains":["Mathematics"],"country":"Switzerland","reference":"https://ieeexplore.ieee.org/document/861302","confidence":"","accessibility":""},{"id":"epoch:05cd2537081c6e66e721","name":"SVD in recommender systems","organization":"University of Minnesota","date":"2000-07-14","domains":["Recommendation"],"country":"United States of America","reference":"http://robotics.stanford.edu/~ronnyk/WEBKDD2000/papers/sarwar.pdf","confidence":"Unknown","accessibility":""},{"id":"epoch:83f6a75e7a7a9b4eec7e","name":"Perceptron for Large Margin Classification","organization":"University of California San Diego,Shannon Laboratory,AT&T","date":"1999-12-01","domains":["Vision"],"country":"United States of America,United States of America,United States of America","reference":"https://link.springer.com/article/10.1023/A:1007662407062","confidence":"","accessibility":""},{"id":"epoch:f53a8ef412392816ebe9","name":"IBM Model 4","organization":"University of Southern California,IBM,University of Pennsylvania","date":"1999-07-02","domains":["Language"],"country":"United States of America,United States of America,United States of America","reference":"http://www-i6.informatik.rwth-aachen.de/publications/download/266/al-onaizan--1999.pdf","confidence":"","accessibility":""},{"id":"epoch:85f9f54c080fe0119d34","name":"Credibilty Network","organization":"University College London (UCL),University of Toronto","date":"1999-07-01","domains":["Vision"],"country":"United Kingdom,Canada","reference":"https://proceedings.neurips.cc/paper_files/paper/1999/file/5a142a55461d5fef016acfb927fee0bd-Paper.pdf","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:d8949a527f6b9cce8e9b","name":"RECONTRA-uncategorized","organization":"","date":"1999-06-02","domains":["Language"],"country":"","reference":"https://www.semanticscholar.org/paper/Text-to-Text-Machine-Translation-Using-the-RECONTRA-Casta%C3%B1o-Casacuberta/47ee046b416d9258952cb8f4b0e2b6e65f334fad","confidence":"Likely","accessibility":""},{"id":"epoch:49db556bdd183a0151bc","name":"RECONTRA-categorized","organization":"","date":"1999-06-02","domains":["Language"],"country":"","reference":"https://www.semanticscholar.org/paper/Text-to-Text-Machine-Translation-Using-the-RECONTRA-Casta%C3%B1o-Casacuberta/47ee046b416d9258952cb8f4b0e2b6e65f334fad","confidence":"Likely","accessibility":""},{"id":"epoch:fe2c0e8a66b48ba5596c","name":"Learning to Order Things","organization":"AT&T","date":"1999-05-15","domains":["Recommendation","Search"],"country":"United States of America","reference":"https://arxiv.org/abs/1105.5464","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:730a9fa64f66978aae65","name":"LSTM with forget gates","organization":"IDSIA","date":"1999-01-02","domains":["Language"],"country":"Switzerland","reference":"https://ieeexplore.ieee.org/document/818041","confidence":"","accessibility":""},{"id":"epoch:fb7fe1e9b11e17e76ee6","name":"LeNet-5","organization":"AT&T","date":"1998-11-01","domains":["Vision"],"country":"United States of America","reference":"http://vision.stanford.edu/cs598_spring07/papers/Lecun98.pdf","confidence":"Confident","accessibility":""},{"id":"epoch:03c0561d9f853879f9bb","name":"Social and content-based classification","organization":"AT&T,Bell Laboratories,Rutgers University","date":"1998-07-01","domains":["Recommendation"],"country":"United States of America,United States of America,United States of America","reference":"https://www.aaai.org/Papers/AAAI/1998/AAAI98-101.pdf","confidence":"","accessibility":""},{"id":"epoch:ea45cdb80afdd705faad","name":"Probabilistic modeling for object recognition","organization":"Carnegie Mellon University (CMU)","date":"1998-06-23","domains":["Vision"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/698586","confidence":"","accessibility":""},{"id":"epoch:be1a7d9e820e4f2f3737","name":"RNN for speech","organization":"National Chiao Tung University","date":"1998-05-15","domains":["Speech"],"country":"Taiwan","reference":"https://ieeexplore.ieee.org/abstract/document/668817","confidence":"","accessibility":""},{"id":"epoch:1e142ab146ac162d073b","name":"Sparse coding model for V1 receptive fields","organization":"UC Davis,Cornell University","date":"1997-12-01","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://www.sciencedirect.com/science/article/pii/S0042698997001697","confidence":"","accessibility":""},{"id":"epoch:211374ceef0c86afb27a","name":"LSTM","organization":"Technical University of Munich","date":"1997-11-15","domains":["Language"],"country":"Germany","reference":"https://direct.mit.edu/neco/article-abstract/9/8/1735/6109/Long-Short-Term-Memory?redirectedFrom=fulltext","confidence":"Confident","accessibility":""},{"id":"epoch:a0cedc9029c38afab1ba","name":"Bidirectional RNN","organization":"Advanced Telecommunications Research Institute","date":"1997-11-01","domains":["Speech"],"country":"Japan","reference":"https://ieeexplore.ieee.org/document/650093","confidence":"","accessibility":""},{"id":"epoch:bdc81e5c9b0f9773674b","name":"n-gram LM","organization":"University of Cambridge,Carnegie Mellon University (CMU)","date":"1997-07-01","domains":["Language"],"country":"United Kingdom,United States of America","reference":"https://www.semanticscholar.org/paper/Statistical-language-modeling-using-the-toolkit-Clarkson-Rosenfeld/fdf4aa623e4d5b5edaeb873ed8e8b1cef0b59c87","confidence":"Unknown","accessibility":""},{"id":"epoch:4624169a2ddc4611ffec","name":"SVM for face detection","organization":"Massachusetts Institute of Technology (MIT)","date":"1997-06-17","domains":["Vision"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/609310","confidence":"","accessibility":""},{"id":"epoch:187c7e2cab24560d9a75","name":"Deep Blue","organization":"IBM","date":"1997-05-01","domains":["Games"],"country":"United States of America","reference":"https://www.sciencedirect.com/science/article/pii/S0004370201001291","confidence":"Speculative","accessibility":"Unreleased"},{"id":"epoch:21520b6088b510882d85","name":"SOM-CNN","organization":"","date":"1997-01-31","domains":["Vision"],"country":"","reference":"https://www.semanticscholar.org/paper/Face-recognition%3A-a-convolutional-neural-network-Lawrence-Giles/86890c82b589e24007c56e1f40c5f928a0e04183","confidence":"Confident","accessibility":""},{"id":"epoch:26d4c08d2a5815a4374e","name":"HMM Word Alignment","organization":"University of Erlangen - Nuremburg","date":"1996-08-05","domains":["Language"],"country":"Germany","reference":"https://dl.acm.org/doi/10.3115/993268.993313","confidence":"","accessibility":""},{"id":"epoch:891b834981b31224f7ee","name":"AdaBoost.M2 Digit Recognition","organization":"AT&T","date":"1996-07-03","domains":["Vision"],"country":"United States of America","reference":"https://cseweb.ucsd.edu/~yfreund/papers/boostingexperiments.pdf","confidence":"Confident","accessibility":""},{"id":"epoch:77a0def659decd1a5d2f","name":"System 11","organization":"Carnegie Mellon University (CMU)","date":"1996-06-18","domains":["Vision"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/655647","confidence":"Confident","accessibility":""},{"id":"epoch:755c5fdfd3ca6233745d","name":"MUSIC perceptron","organization":"","date":"1996-06-03","domains":["Vision"],"country":"","reference":"https://ieeexplore.ieee.org/document/549237","confidence":"Confident","accessibility":""},{"id":"epoch:44a0f953eb015ff3a6d7","name":"LISSOM","organization":"University of Texas at Austin","date":"1995-11-27","domains":["Vision"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Laterally-Interconnected-Self-Organizing-Maps-in-Choe-Sirosh/785f5facc76538ceba6f6b9e2d7b641d322e9854","confidence":"Likely","accessibility":""},{"id":"epoch:3c1fbde436524475b0ee","name":"Support Vector Machines","organization":"AT&T,Bell Laboratories","date":"1995-09-01","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://link.springer.com/article/10.1007/BF00994018","confidence":"","accessibility":""},{"id":"epoch:351daef51faa28f4f3f4","name":"Random Decision Forests","organization":"AT&T,Bell Laboratories","date":"1995-08-14","domains":["Other"],"country":"United States of America,United States of America","reference":"https://ieeexplore.ieee.org/document/598994","confidence":"","accessibility":""},{"id":"epoch:a86e62d82aba922ac3f3","name":"Iterative Bootstrapping WSD","organization":"University of Pennsylvania","date":"1995-06-26","domains":["Language"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.3115/981658.981684","confidence":"","accessibility":""},{"id":"epoch:9d03447623123b21e77a","name":"Multi-cause Binary Clustering","organization":"Xerox","date":"1995-01-01","domains":["Other"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/6795568","confidence":"Unknown","accessibility":""},{"id":"epoch:891926eaf183be55e05c","name":"JPMAX","organization":"","date":"1994-12-02","domains":["Vision"],"country":"","reference":"https://proceedings.neurips.cc/paper_files/paper/1994/hash/4b0250793549726d5c1ea3906726ebfe-Abstract.html","confidence":"Speculative","accessibility":""},{"id":"epoch:ca7ef0609e34b56046bf","name":"Predictive Coding NN","organization":"Technical University of Munich","date":"1994-12-02","domains":["Language"],"country":"Germany","reference":"https://proceedings.neurips.cc/paper/1994/file/5705e1164a8394aace6018e27d20d237-Paper.pdf","confidence":"Confident","accessibility":""},{"id":"epoch:060e813bb8c86c63c401","name":"Mixture of linear models","organization":"","date":"1994-12-02","domains":["Vision"],"country":"","reference":"https://www.semanticscholar.org/paper/Recognizing-Handwritten-Digits-Using-Mixtures-of-Hinton-Revow/9dea20c1e5bbb1f543ff08113ffde5380c679f1f","confidence":"Likely","accessibility":""},{"id":"epoch:2a61ae1d032037139c31","name":"NeuroChess","organization":"","date":"1994-12-02","domains":["Games"],"country":"","reference":"https://www.semanticscholar.org/paper/Learning-to-Play-the-Game-of-Chess-Thrun/4bc7a6dcb9e0e6c7a26800532e2a00f5572eea47","confidence":"Speculative","accessibility":""},{"id":"epoch:6bc47b362fc0e4098265","name":"GroupLens","organization":"Massachusetts Institute of Technology (MIT)","date":"1994-10-22","domains":["Recommendation"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.1145/192844.192905","confidence":"","accessibility":""},{"id":"epoch:8d20fcf958be7cb5ced9","name":"Markov-driven POS tagger","organization":"EURECOM","date":"1994-06-01","domains":["Language"],"country":"France","reference":"https://dl.acm.org/doi/10.5555/972525.972526","confidence":"","accessibility":""},{"id":"epoch:bcd42f3caed9b7938b57","name":"Ceramic-MLP","organization":"Sapienza Universit\u00e0 di Roma","date":"1994-01-07","domains":["Materials science"],"country":"Italy","reference":"https://www.sciencedirect.com/science/article/abs/pii/S0921883108605506","confidence":"Likely","accessibility":""},{"id":"epoch:bff8be13eb870434e39e","name":"ANN Eye Tracker","organization":"","date":"1993-11-29","domains":["Vision"],"country":"","reference":"https://www.semanticscholar.org/paper/Non-Intrusive-Gaze-Tracking-Using-Artificial-Neural-Baluja-Pomerleau/574c0cf98825bf09b0ab7bbfe9ba89cd6090745e","confidence":"Confident","accessibility":""},{"id":"epoch:15dd1dab2435bd3a2053","name":"Learning-curve prediction","organization":"AT&T","date":"1993-11-29","domains":["Vision"],"country":"United States of America","reference":"https://proceedings.neurips.cc/paper/1993/file/1aa48fc4880bb0c9b8a3bf979d3b917e-Paper.pdf","confidence":"Confident","accessibility":"Unreleased"},{"id":"epoch:9cc11e2348b13cb3ed97","name":"Siamese-TDNN","organization":"Bell Laboratories","date":"1993-08-01","domains":["Vision"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Signature-Verification-Using-A-%22Siamese%22-Time-Delay-Bromley-Bentz/997dc5d9a058753f034422afe7bd0cc0b8ad808b","confidence":"Likely","accessibility":""},{"id":"epoch:b713e06e9ac62def7487","name":"IBM-5","organization":"IBM","date":"1993-06-15","domains":["Language"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.5555/972470.972474","confidence":"","accessibility":""},{"id":"epoch:b7d58b9b499bfd12472d","name":"Futures trading net","organization":"","date":"1993-01-01","domains":["Other"],"country":"","reference":"https://legacy.farmdoc.illinois.edu/nccc134/conf_1993/pdf/confp25-93.pdf","confidence":"Speculative","accessibility":""},{"id":"epoch:e08af31f3aaf18db257a","name":"Boosting","organization":"Bell Laboratories","date":"1992-11-30","domains":["Vision"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Improving-Performance-in-Neural-Networks-Using-a-Drucker-Schapire/77b5185dafb9e5b884a677a32713e54c253a4e0b","confidence":"Likely","accessibility":""},{"id":"epoch:b03355f52e905bd2b7d4","name":"Cancer drug mechanism prediction","organization":"National Cancer Institute","date":"1992-10-16","domains":["Medicine"],"country":"United States of America","reference":"https://pubmed.ncbi.nlm.nih.gov/1411538/","confidence":"Likely","accessibility":""},{"id":"epoch:0c7cfbf3a1140b4d673e","name":"Golem","organization":"Alan Turing Institute","date":"1992-10-01","domains":["Biology"],"country":"United Kingdom","reference":"https://www.semanticscholar.org/paper/Protein-secondary-structure-prediction-using-Muggleton-King/9f744e48091a24b569435d070920e60db45f4fdc","confidence":"Confident","accessibility":""},{"id":"epoch:8440eba9676a861610ee","name":"Fuzzy NN","organization":"Indian Statistical Institute","date":"1992-09-01","domains":["Speech"],"country":"India","reference":"https://ieeexplore.ieee.org/document/159058","confidence":"Confident","accessibility":""},{"id":"epoch:004650086accdc88bd44","name":"TD-Gammon","organization":"IBM","date":"1992-05-01","domains":["Games"],"country":"United States of America","reference":"https://papers.nips.cc/paper/1991/file/68ce199ec2c5517597ce0a4d89620f55-Paper.pdf","confidence":"Speculative","accessibility":""},{"id":"epoch:c80c181f82de0f9f281d","name":"Weight Decay","organization":"","date":"1991-12-02","domains":["Speech"],"country":"","reference":"https://www.semanticscholar.org/paper/A-Simple-Weight-Decay-Can-Improve-Generalization-Krogh-Hertz/48e1de7d085808004d5f0493d486669a3d2930b5","confidence":"Confident","accessibility":""},{"id":"epoch:ce4ee3b5897bdec19a77","name":"SRN-Encoded Grammatical Structures","organization":"University of California San Diego","date":"1991-09-01","domains":["Language"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.1007/BF00114844","confidence":"","accessibility":""},{"id":"epoch:797482f5033ebe004c88","name":"DIABETES","organization":"Aalborg University,University of London","date":"1991-06-24","domains":["Medicine"],"country":"Denmark,United Kingdom","reference":"https://link.springer.com/chapter/10.1007/978-3-642-48650-0_19","confidence":"","accessibility":""},{"id":"epoch:5f40c00fa9860182cb48","name":"RAAM","organization":"","date":"1990-11-01","domains":["Other"],"country":"","reference":"https://www.sciencedirect.com/science/article/abs/pii/000437029090005K?via%3Dihub","confidence":"Confident","accessibility":""},{"id":"epoch:5ad4866c717d55635942","name":"SexNet compression","organization":"","date":"1990-10-01","domains":["Vision"],"country":"","reference":"https://www.semanticscholar.org/paper/SEXNET%3A-A-Neural-Network-Identifies-Sex-From-Human-Golomb-Lawrence/cbf90aa78fea0c8a1028705d92bc4bc7808ddeeb","confidence":"Confident","accessibility":""},{"id":"epoch:7905e02b99ce5ed781b9","name":"SexNet classification","organization":"","date":"1990-10-01","domains":["Vision"],"country":"","reference":"https://www.semanticscholar.org/paper/SEXNET%3A-A-Neural-Network-Identifies-Sex-From-Human-Golomb-Lawrence/cbf90aa78fea0c8a1028705d92bc4bc7808ddeeb","confidence":"Likely","accessibility":""},{"id":"epoch:3fe19be0c92176b8dc02","name":"ISR network","organization":"Stanford University","date":"1990-10-01","domains":["Vision"],"country":"United States of America","reference":"https://papers.nips.cc/paper_files/paper/1990/hash/e46de7e1bcaaced9a54f1e9d0d2f800d-Abstract.html","confidence":"Confident","accessibility":""},{"id":"epoch:67d32be65acbaff531f9","name":"Bankruptcy-NN","organization":"","date":"1990-06-17","domains":["Other"],"country":"","reference":"https://www.semanticscholar.org/paper/A-neural-network-model-for-bankruptcy-prediction-Odom-Sharda/ead9fa02902850a7418fb5ba720f3d9d8ab2f88b","confidence":"Confident","accessibility":""},{"id":"epoch:1b561e8760a5e49192b3","name":"NETtalk reimplementation","organization":"Oregon State University","date":"1990-06-01","domains":["Speech"],"country":"United States of America","reference":"https://www.sciencedirect.com/science/article/abs/pii/B9781558601413500079","confidence":"Confident","accessibility":""},{"id":"epoch:9a39c89883e76394d93d","name":"Zip CNN","organization":"AT&T,Bell Laboratories","date":"1989-12-01","domains":["Vision"],"country":"United States of America,United States of America","reference":"https://ieeexplore.ieee.org/document/6795724","confidence":"Confident","accessibility":""},{"id":"epoch:65d8dca66553d5290371","name":"Innervator","organization":"Stanford University,California Institute of Technology","date":"1989-12-01","domains":["Mathematics"],"country":"United States of America,United States of America","reference":"https://www.researchgate.net/publication/220885651_Designing_Neural_Networks_using_Genetic_Algorithms","confidence":"Confident","accessibility":""},{"id":"epoch:c0635471f7c1ddeef3c6","name":"ALVINN","organization":"Carnegie Mellon University (CMU)","date":"1989-12-01","domains":["Driving"],"country":"United States of America","reference":"https://proceedings.neurips.cc/paper/1988/hash/812b4ba287f5ee0bc9d43bbf5bbe87fb-Abstract.html","confidence":"Confident","accessibility":""},{"id":"epoch:8707c0cc6281f0b52078","name":"Speaker-independent vowel classification","organization":"University of Washington","date":"1989-11-27","domains":["Speech"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Performance-Comparisons-Between-Backpropagation-and-Atlas-Cole/e42d2b89fcb4a1a3dfa63408f424f76975ed1e1b","confidence":"Confident","accessibility":""},{"id":"epoch:dba570bb969018017a7f","name":"Handwritten digit recognition network","organization":"AT&T","date":"1989-11-27","domains":["Vision"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Handwritten-Digit-Recognition-with-a-Network-LeCun-Boser/86ab4cae682fbd49c5a5bedb630e5a40fa7529f6","confidence":"Confident","accessibility":""},{"id":"epoch:609b82199f6044a36feb","name":"Invariant image recognition","organization":"Complutense University of Madrid","date":"1989-06-18","domains":["Vision"],"country":"Spain","reference":"https://ieeexplore.ieee.org/document/118669","confidence":"Confident","accessibility":""},{"id":"epoch:9d36b07a6442848db515","name":"Truck backer-upper","organization":"Stanford University","date":"1989-06-18","domains":["Driving"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/118723","confidence":"Confident","accessibility":""},{"id":"epoch:e0b049ea2641b4acf073","name":"Time-delay neural networks","organization":"Advanced Telecommunications Research Institute,Carnegie Mellon University (CMU)","date":"1989-03-03","domains":["Speech"],"country":"Japan,United States of America","reference":"https://ieeexplore.ieee.org/abstract/document/21701","confidence":"Unknown","accessibility":""},{"id":"epoch:b437357864011ddc797b","name":"Q-learning","organization":"University of London","date":"1989-01-01","domains":["Robotics","Games"],"country":"United Kingdom","reference":"http://www.cs.rhul.ac.uk/~chrisw/thesis.html","confidence":"Unknown","accessibility":""},{"id":"epoch:31a0230f3f5658a4c8dc","name":"MLP baggage detector","organization":"Science Applications International Corporation / SAIC","date":"1989-01-01","domains":["Vision"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Detection-of-explosives-in-checked-airline-baggage-Shea-Lin/71da4057401f459bc079696a029aee45a0a89728","confidence":"Confident","accessibility":""},{"id":"epoch:7312d9af3bb184294bc5","name":"MLN-ASR","organization":"McGill University","date":"1988-08-01","domains":["Speech"],"country":"Canada","reference":"https://aaai.org/papers/00734-aaai88-130-data-driven-execution-of-multi-layered-networks-for-automatic-speech-recognition/","confidence":"Confident","accessibility":""},{"id":"epoch:782ff82c0796bd1fb3c3","name":"Adaptive Broom Balancer","organization":"Stanford University","date":"1988-07-24","domains":["Games"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/23982","confidence":"","accessibility":""},{"id":"epoch:a6b2eaac65fdb636ee35","name":"MADALINE II","organization":"Stanford University","date":"1988-07-24","domains":["Mathematics"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/23872","confidence":"Unknown","accessibility":""},{"id":"epoch:2e58175aab8460104f85","name":"Latent semantic analysis","organization":"University of Chicago,Bell Laboratories,University of Western Ontario","date":"1988-04-05","domains":["Language"],"country":"United States of America,United States of America,Canada","reference":"https://asistdl.onlinelibrary.wiley.com/doi/10.1002/%28SICI%291097-4571%28199009%2941%3A6%3C391%3A%3AAID-ASI1%3E3.0.CO%3B2-9\n\nhttps://www.cs.csustan.edu/~mmartin/LDS/Deerwester-et-al.pdf","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:64abc2501834e9055a8e","name":"Translation-invariant MLP","organization":"Carnegie Mellon University (CMU)","date":"1987-06-15","domains":[],"country":"United States of America","reference":"https://www.cs.toronto.edu/~hinton/absps/parle.pdf","confidence":"Confident","accessibility":""},{"id":"epoch:fb0943015c7d6dfe7e8e","name":"NetTalk (transcription)","organization":"Princeton University","date":"1987-06-06","domains":["Speech"],"country":"United States of America","reference":"http://citeseerx.ist.psu.edu/viewdoc/download;jsessionid=03A3D3EDF0BAF35405ABCF083411B55E?doi=10.1.1.154.7012&rep=rep1&type=pdf","confidence":"Confident","accessibility":""},{"id":"epoch:924906810a4a0d2bf057","name":"NetTalk (dictionary)","organization":"Princeton University","date":"1987-06-06","domains":["Speech"],"country":"United States of America","reference":"http://citeseerx.ist.psu.edu/viewdoc/download;jsessionid=03A3D3EDF0BAF35405ABCF083411B55E?doi=10.1.1.154.7012&rep=rep1&type=pdf","confidence":"Confident","accessibility":""},{"id":"epoch:7d21994dc5c3ad79909a","name":"Optimized Multi-Scale Edge Detection","organization":"Massachusetts Institute of Technology (MIT)","date":"1986-11-01","domains":["Vision"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=4767851","confidence":"Unknown","accessibility":""},{"id":"epoch:c0541c786a6c6ad8712c","name":"MLP with back-propagation","organization":"University of California San Diego,Carnegie Mellon University (CMU)","date":"1986-10-01","domains":["Mathematics"],"country":"United States of America,United States of America","reference":"https://www.semanticscholar.org/paper/Learning-representations-by-back-propagating-errors-Rumelhart-Hinton/052b1d8ce63b07fec3de9dbb583772d860b7c769","confidence":"Confident","accessibility":""},{"id":"epoch:ef7ebcb1f1b3e315520b","name":"Distributed representation NN","organization":"Carnegie Mellon University (CMU)","date":"1986-08-15","domains":["Other"],"country":"United States of America","reference":"https://www.cs.toronto.edu/~hinton/absps/families.pdf","confidence":"Confident","accessibility":""},{"id":"epoch:7a108aaccb6c298a1860","name":"PDP model for serial order","organization":"University of California San Diego","date":"1986-01-05","domains":["Speech"],"country":"United States of America","reference":"https://www.osti.gov/biblio/6910294","confidence":"Unknown","accessibility":""},{"id":"epoch:f0c4f6aa218be20c2082","name":"Error Propagation","organization":"University of California San Diego,Carnegie Mellon University (CMU)","date":"1986-01-03","domains":["Other"],"country":"United States of America,United States of America","reference":"https://stanford.edu/~jlmcc/papers/PDP/Volume%201/Chap8_PDP86.pdf","confidence":"Unknown","accessibility":""},{"id":"epoch:ccce8cce94377bce7b35","name":"Learning past tenses","organization":"Stanford University","date":"1986-01-03","domains":["Language"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/On-learning-the-past-tenses-of-English-verbs%3A-rules-Rumelhart-McClelland/4fa569625b5ab35e955a8d5be11a4aa9f59ca424","confidence":"","accessibility":""},{"id":"epoch:e41eecc8f42467ffa991","name":"Learnability theory of language development","organization":"Massachusetts Institute of Technology (MIT)","date":"1984-07-01","domains":["Language"],"country":"United States of America","reference":"https://psycnet.apa.org/record/1985-97439-000","confidence":"Unknown","accessibility":""},{"id":"epoch:4947f20ed9ccd473e21f","name":"Hierarchical Cognitron","organization":"NHK Broadcasting Science Research Laboratories","date":"1984-04-01","domains":["Other"],"country":"Japan","reference":"https://link.springer.com/article/10.1007/BF00337157","confidence":"Speculative","accessibility":""},{"id":"epoch:0562abd5b4b7a24b9d83","name":"ASE+ACE","organization":"University of Massachusetts Amherst","date":"1983-09-01","domains":["Robotics"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=6313077","confidence":"Likely","accessibility":""},{"id":"epoch:fd9f0835e832454adb41","name":"Hopfield network","organization":"California Institute of Technology","date":"1982-04-01","domains":["Other"],"country":"United States of America","reference":"https://www.pnas.org/doi/10.1073/pnas.79.8.2554","confidence":"","accessibility":""},{"id":"epoch:0bbdca0d2cda04e7db8c","name":"Kohonen network","organization":"Helsinki University of Technology","date":"1981-07-25","domains":["Mathematics"],"country":"Finland","reference":"https://link.springer.com/article/10.1007/BF00337288","confidence":"","accessibility":""},{"id":"epoch:67afef1fdae327096f0a","name":"Neocognitron","organization":"NHK Broadcasting Science Research Laboratories","date":"1980-04-01","domains":["Vision"],"country":"Japan","reference":"https://link.springer.com/article/10.1007/BF00344251","confidence":"Confident","accessibility":""},{"id":"epoch:0880d4d37939a2efaa55","name":"Transfer Learning","organization":"University of Zagreb","date":"1976-07-01","domains":["Vision"],"country":"Croatia","reference":"https://www.informatica.si/index.php/informatica/article/view/2828","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:c4cb394d4656ba32d7cc","name":"Statistical continuous speech recognizer","organization":"Massachusetts Institute of Technology (MIT)","date":"1976-04-30","domains":["Speech"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/1454428","confidence":"Confident","accessibility":""},{"id":"epoch:d004b0544716aded6412","name":"Cognitron","organization":"Biological Cybernetics","date":"1975-09-01","domains":["Other"],"country":"Japan","reference":"https://link.springer.com/article/10.1007%2FBF00342633","confidence":"Confident","accessibility":""},{"id":"epoch:4e549b3bc7011b2d061d","name":"Piecewise linear model","organization":"University of Kansas","date":"1973-11-01","domains":["Vision"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/4309314","confidence":"Confident","accessibility":""},{"id":"epoch:16510c4adf0790c35ed6","name":"Punish/Reward","organization":"IEEE","date":"1973-09-01","domains":["Games"],"country":"Multinational","reference":"https://ieeexplore.ieee.org/document/4309272","confidence":"","accessibility":""},{"id":"epoch:109cf3a7c51a62879718","name":"Self-Organizing Nets of Threshold Elements","organization":"University of Tokyo","date":"1972-11-30","domains":["Other"],"country":"Japan","reference":"https://ieeexplore.ieee.org/document/1672070/\n\nhttps://people.idsia.ch/~juergen/amari1972hopfield.pdf","confidence":"Unknown","accessibility":"Unreleased"},{"id":"epoch:8a0f5876d1fe0bbec861","name":"Graph-based structural reasoning","organization":"Massachusetts Institute of Technology (MIT)","date":"1970-09-01","domains":["Vision"],"country":"United States of America","reference":"https://dspace.mit.edu/handle/1721.1/6884","confidence":"Unknown","accessibility":""},{"id":"epoch:ca72c0b4034031410d0a","name":"Decision tree adaline","organization":"Tokyo Medical and Dental University","date":"1969-05-01","domains":["Medicine"],"country":"Japan","reference":"https://pubmed.ncbi.nlm.nih.gov/5820353/","confidence":"Confident","accessibility":""},{"id":"epoch:ee93af2722270cf019d6","name":"GLEE","organization":"University of Edinburgh","date":"1968-07-01","domains":["Games"],"country":"United Kingdom","reference":"https://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.474.2430","confidence":"Unknown","accessibility":""},{"id":"epoch:1bb4ad1df4ac3e7612f2","name":"Boxes (pole)","organization":"University of Edinburgh","date":"1968-07-01","domains":["Games"],"country":"United Kingdom","reference":"https://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.474.2430","confidence":"Unknown","accessibility":""},{"id":"epoch:56f4804787076b008f05","name":"Samuel Neural Checkers II","organization":"University of Geneva","date":"1967-11-01","domains":["Games"],"country":"Switzerland","reference":"https://www.cs.virginia.edu/~evans/greatworks/samuel.pdf","confidence":"","accessibility":""},{"id":"epoch:2b8215433df0ed880d33","name":"LTE speaker verification system","organization":"IBM","date":"1966-11-01","domains":["Speech"],"country":"United States of America","reference":"https://pubs.aip.org/asa/jasa/article-abstract/40/5/966/754180/Experimental-Studies-in-Speaker-Verification-Using?redirectedFrom=fulltext","confidence":"Likely","accessibility":""},{"id":"epoch:f40738700c0e6671950f","name":"Heuristic Reinforcement Learning","organization":"Purdue University","date":"1965-10-01","domains":["Robotics"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/1098193","confidence":"Speculative","accessibility":""},{"id":"epoch:08ebd548054613e007b1","name":"MENACE","organization":"University of Edinburgh","date":"1963-11-01","domains":["Games"],"country":"United Kingdom","reference":"https://academic.oup.com/comjnl/article/6/3/232/360077","confidence":"Unknown","accessibility":""},{"id":"epoch:0650ed15cee61f73fd9f","name":"STeLLA","organization":"University of Canterbury","date":"1963-06-01","domains":["Robotics"],"country":"New Zealand","reference":"https://www.sciencedirect.com/science/article/pii/S1474667017696824?via%3Dihub","confidence":"Unknown","accessibility":""},{"id":"epoch:c4c7f32ac73a77579bf6","name":"Print Recognition Logic","organization":"IBM","date":"1963-01-01","domains":["Vision"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/5392331","confidence":"Speculative","accessibility":""},{"id":"epoch:3522adda09050197c416","name":"MADALINE I","organization":"Stanford University","date":"1962-07-01","domains":["Other"],"country":"United States of America","reference":"https://www.proquest.com/openview/7898314db50a218b58052ac91e3bde1e/1?","confidence":"Unknown","accessibility":""},{"id":"epoch:f741ecfda925e91e5042","name":"Linear Decision Functions","organization":"Bell Laboratories","date":"1962-06-01","domains":["Mathematics"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/document/4066882?denied=","confidence":"Speculative","accessibility":""},{"id":"epoch:703b93fab58212f8cd3d","name":"PAPA","organization":"University of Genoa","date":"1961-09-01","domains":["Other"],"country":"Italy","reference":"https://www.semanticscholar.org/paper/Further-experiments-with-PAPA-Gamba-Gamberini/c3a20b9aa86033cec29f08e69f4bc81e8b329ae2","confidence":"Unknown","accessibility":""},{"id":"epoch:b21857b0cb92bd0b1f36","name":"ADALINE","organization":"Stanford University","date":"1960-06-30","domains":["Vision"],"country":"United States of America","reference":"https://isl.stanford.edu/~widrow/papers/c1960adaptiveswitching.pdf","confidence":"Confident","accessibility":""},{"id":"epoch:37a4a6ddc0e3e70ddf92","name":"Perceptron (1960)","organization":"Cornell Aeronautical Laboratory","date":"1960-03-30","domains":["Vision"],"country":"United States of America","reference":"https://www.semanticscholar.org/paper/Perceptron-Simulation-Experiments-Rosenblatt/ae76ce1ba27ac29addce4aab93b927e9bc7f7c67","confidence":"Speculative","accessibility":""},{"id":"epoch:d099f764d70520a28f0e","name":"Pattern recognition and reading by machine","organization":"Sandia Corporation","date":"1959-12-01","domains":["Vision"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.1145/1460299.1460326","confidence":"","accessibility":""},{"id":"epoch:9ce1cf4fb91a61d99b49","name":"Samuel Neural Checkers","organization":"IBM","date":"1959-07-01","domains":["Games"],"country":"United States of America","reference":"https://ieeexplore.ieee.org/abstract/document/5392560","confidence":"Likely","accessibility":""},{"id":"epoch:e7312cc1eea097a2f5a3","name":"Pandemonium (morse)","organization":"Massachusetts Institute of Technology (MIT)","date":"1959-02-01","domains":["Language"],"country":"United States of America","reference":"https://aitopics.org/doc/classics:504E1BAC/","confidence":"Speculative","accessibility":""},{"id":"epoch:787b11189d7462d799b8","name":"Perceptron Mark I","organization":"Cornell Aeronautical Laboratory,Cornell University","date":"1957-01-01","domains":["Other"],"country":"United States of America,United States of America","reference":"https://blogs.umass.edu/brain-wars/files/2016/03/rosenblatt-1957.pdf","confidence":"Likely","accessibility":""},{"id":"epoch:a068e76fff93c0b0c404","name":"Sequence-based pattern recognition","organization":"Massachusetts Institute of Technology (MIT)","date":"1955-03-01","domains":["Vision"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.1145/1455292.1455310","confidence":"Unknown","accessibility":""},{"id":"epoch:dd2787ab41b1431c6286","name":"Self Organizing System","organization":"Massachusetts Institute of Technology (MIT)","date":"1955-03-01","domains":["Other"],"country":"United States of America","reference":"https://dl.acm.org/doi/10.1145/1455292.1455309","confidence":"","accessibility":""},{"id":"epoch:15ad0ab4c212442d9b01","name":"Genetic algorithm","organization":"Institute for Advanced Study","date":"1954-07-02","domains":["Mathematics","Biology"],"country":"United States of America","reference":"https://link.springer.com/article/10.1007/BF01556771","confidence":"Unknown","accessibility":""},{"id":"epoch:2fc4929ecc20cb5f4ee0","name":"SNARC","organization":"Harvard University","date":"1952-01-08","domains":["Robotics"],"country":"United States of America","reference":"https://en.wikipedia.org/wiki/Stochastic_neural_analog_reinforcement_calculator","confidence":"","accessibility":""},{"id":"epoch:1ded709e181fdf0416f8","name":"Theseus","organization":"Bell Laboratories","date":"1950-07-02","domains":["Robotics"],"country":"United States of America","reference":"https://www.technologyreview.com/2018/12/19/138508/mighty-mouse/","confidence":"Confident","accessibility":""}],"observations":[{"id":"epoch:412993d893d566bd5555:parameters","entityId":"epoch:412993d893d566bd5555","metric":"parameters","value":2100000000.0,"unit":"parameters","observationPeriod":"2026-09-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Not confirmed in official release but Elon said \"Grok 4.7 will be the 2.1T model released a few weeks later. This will be better than 4.6 in every way, except slightly slower to serve, albeit with even better token efficiency.\" on July 28, 2026.\n\nOn Sepetember 14 he said \"It\u2019s unequivocally better than the 2.1T that was trained with Jax [...]\", which presumably refers to 4.7's size."},{"id":"epoch:62eb096966556ab919cf:parameters","entityId":"epoch:62eb096966556ab919cf","metric":"parameters","value":600000000000.0,"unit":"parameters","observationPeriod":"2026-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9f8a4ce481b1194dc259:parameters","entityId":"epoch:9f8a4ce481b1194dc259","metric":"parameters","value":1040000000000.0,"unit":"parameters","observationPeriod":"2026-09-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Kimi K2.6 as base"},{"id":"epoch:1f135d59b7429310efa1:parameters","entityId":"epoch:1f135d59b7429310efa1","metric":"parameters","value":2800000000000.0,"unit":"parameters","observationPeriod":"2026-09-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Kimi K3 as base; 2.8T total, 104B active parameters"},{"id":"epoch:6801998507d114a75aab:parameters","entityId":"epoch:6801998507d114a75aab","metric":"parameters","value":748000000000.0,"unit":"parameters","observationPeriod":"2026-09-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"748B total (196B Engram, which is their method which trains what is essentially a lookup table used during inference + 552B \"Backbone\"), 16B active during decode, 8B during prefill."},{"id":"epoch:6801998507d114a75aab:training-compute","entityId":"epoch:6801998507d114a75aab","metric":"training-compute","value":4.32e+24,"unit":"FLOP","observationPeriod":"2026-09-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(Pre-training) - 6 * 16e9 active parameters * 45e12 tokens = 4.32e+24 FLOP\n"},{"id":"epoch:fd50ac4f7952beb4cca1:training-compute","entityId":"epoch:fd50ac4f7952beb4cca1","metric":"training-compute","value":9.4e+26,"unit":"FLOP","observationPeriod":"2026-09-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"per OpenAI and affiliates (e.g. Nvidia), 6 Astra was trained on at least 100,000 GB200s in Abilene, Texas (so 100,000 in \"Hardware quantity\" is a lower bound/underestimate).\n\nThis suggests around 9e26 FLOP (corresponding to ~100k GB200s over 90 days at 25% FP8 MFU). See more detailed estimate in this notebook, yielding a CI of ~[5e26, 2e27] FLOP.\nhttps://colab.research.google.com/drive/1DRlugIQs-_zoV66AJT1gQg6I1-NX7NqN "},{"id":"epoch:d71c85e4f4277bfe69b7:training-compute","entityId":"epoch:d71c85e4f4277bfe69b7","metric":"training-compute","value":2.0725198384e+23,"unit":"FLOP","observationPeriod":"2026-09-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Hardware-based estimate: \"In total, both models required approximately 4.8 chip-years on TPUv4 and 8.9 chip-years on TPU7x chips.\" Assuming FP16 precision and 30% utilization - \n\n1.3e22 FLOP (4.8y on TPUv4) + \n\n2307000000000000 FLOP/s * 86400 seconds/day * 365 days/year * 8.9 years * 0.3 utilization = 1.9425198384e+23 FLOP (8.9y on TPUv7)\n\n= 2.0725198384000003e+23 FLOP"},{"id":"epoch:ee24010e3280faef3362:parameters","entityId":"epoch:ee24010e3280faef3362","metric":"parameters","value":770000000000.0,"unit":"parameters","observationPeriod":"2026-08-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:250ad376b18772157d87:parameters","entityId":"epoch:250ad376b18772157d87","metric":"parameters","value":320000000000.0,"unit":"parameters","observationPeriod":"2026-08-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"320B total, 18B active."},{"id":"epoch:250ad376b18772157d87:training-compute","entityId":"epoch:250ad376b18772157d87","metric":"training-compute","value":3.24e+24,"unit":"FLOP","observationPeriod":"2026-08-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 * 18e9 active parameters * 30e12 tokens = 3.24e24 FLOP. \n\nhttps://docs.z.ai/guides/vlm/glm-5.3-flash says \"Combined with our latest 30T-token multimodal pre-training corpus, these changes let GLM-5.3-Flash produce more intelligence with less compute.\", which is not definitive in saying that all 30T tokens were used exactly once; likely confidence."},{"id":"epoch:4b060905d75d9b191b0c:parameters","entityId":"epoch:4b060905d75d9b191b0c","metric":"parameters","value":744000000000.0,"unit":"parameters","observationPeriod":"2026-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Same base model as GLM-5.2\", which had 744B total and 40B active."},{"id":"epoch:d3f7eb8492cda2e0a810:parameters","entityId":"epoch:d3f7eb8492cda2e0a810","metric":"parameters","value":27000000000.0,"unit":"parameters","observationPeriod":"2026-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"27B dense"},{"id":"epoch:4c6d008b71aa4c1863e6:parameters","entityId":"epoch:4c6d008b71aa4c1863e6","metric":"parameters","value":1500000000000.0,"unit":"parameters","observationPeriod":"2026-08-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Per Elon Musk, this is a 1.5T parameter model, presumably meaning total parameters. Grok 4.6 is almost certainly a sparse mixture of experts model with much fewer active parameters than total.\n\nhttps://x.com/elonmusk/status/2082123925283041545?s=20 "},{"id":"epoch:a917df48e391b8e99a88:parameters","entityId":"epoch:a917df48e391b8e99a88","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2026-08-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"30B total, 3B active"},{"id":"epoch:5bead17b8e1031c76386:parameters","entityId":"epoch:5bead17b8e1031c76386","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2026-08-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ad7146f299aa000780c0:parameters","entityId":"epoch:ad7146f299aa000780c0","metric":"parameters","value":2400000000000.0,"unit":"parameters","observationPeriod":"2026-08-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.4T total, 95B active"},{"id":"epoch:9a12d7383424f34c0c36:parameters","entityId":"epoch:9a12d7383424f34c0c36","metric":"parameters","value":314000000000.0,"unit":"parameters","observationPeriod":"2026-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Fine-grained sparse MoE: approximately 314B total parameters, ~13.2B activated per token. 53 Transformer layers (2 dense + 51 MoE); each MoE layer has 384 routed experts + 1 shared expert, with top-8 routing. Grouped Differential Latent Attention (GDLA), modified manifold-constrained hyper-connections, Expert-Specific PolyNorm, and a 1-layer MTP head. Trained from scratch."},{"id":"epoch:9a12d7383424f34c0c36:training-compute","entityId":"epoch:9a12d7383424f34c0c36","metric":"training-compute","value":1.1e+24,"unit":"FLOP","observationPeriod":"2026-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.1e24 FLOP reported by Motif Technologies in a spreadsheet provided via email (August 2026).\n\nNote an internal inconsistency in the developer-reported hardware figures: 768 B200 GPUs for ~100 days (~2,400 hours) at the claimed 18.3% MFU would imply ~2.7e24 FLOP at BF16 peak throughput; 1.1e24 FLOP over that duration implies ~7% MFU. The token-based cross-check supports the 1.1e24 figure, so the MFU claim was not recorded."},{"id":"epoch:296b0466aba910dd5c79:parameters","entityId":"epoch:296b0466aba910dd5c79","metric":"parameters","value":284000000000.0,"unit":"parameters","observationPeriod":"2026-07-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"284B total, 13B active"},{"id":"epoch:296b0466aba910dd5c79:training-compute","entityId":"epoch:296b0466aba910dd5c79","metric":"training-compute","value":2.496e+24,"unit":"FLOP","observationPeriod":"2026-07-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From https://api-docs.deepseek.com/updates/: \"DeepSeek-V4-Flash-0731 keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained.\"\n\n6 * 13e9 active parameters * 32e12 tokens = 2.496E24 for pre-training"},{"id":"epoch:dc8643a1b4bcc32a8150:parameters","entityId":"epoch:dc8643a1b4bcc32a8150","metric":"parameters","value":750000000000.0,"unit":"parameters","observationPeriod":"2026-07-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"MoE architecture: 750B total parameters, ~37B activated per token. 78 layers (2 dense + 76 MoE); 256 routed experts + 1 shared expert, top-8 routed experts activated per token. Upcycled from K-EXAONE (236B total / 23B active) by depth expansion (48 -> 78 layers) and expert duplication (128 -> 256 experts per layer). Excludes auxiliary speculative-decoding modules (MTP 0.52B, DSpark 2.53B)."},{"id":"epoch:dc8643a1b4bcc32a8150:training-compute","entityId":"epoch:dc8643a1b4bcc32a8150","metric":"training-compute","value":3.55e+24,"unit":"FLOP","observationPeriod":"2026-07-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.55e24 FLOP reported by LG AI Research via email (July 2026), alongside a 16T-token training data figure.\n\nCross-check by operation counting: 6 * 37e9 active parameters * 16e12 tokens = 3.55e24 FLOP, exactly consistent with the reported figure.\n\nNote that K-EXAONE 2.0 is upcycled from K-EXAONE (11T tokens at 23B active parameters, 1.52e24 FLOP). The technical report describes a healing stage followed by ~8T tokens of continual pre-training, 800B tokens of mid-training (2 x 400B), and 350B tokens of SFT on the 750B/37B architecture; the report does not give a total token count, so the relationship between the reported 16T tokens and the predecessor's 11T tokens is not fully specified, and it is unclear whether the reported FLOP figure includes the predecessor's compute."},{"id":"epoch:2d546354af8efa621e76:parameters","entityId":"epoch:2d546354af8efa621e76","metric":"parameters","value":276000000000.0,"unit":"parameters","observationPeriod":"2026-07-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12B active"},{"id":"epoch:b13077cb88dbdc200700:parameters","entityId":"epoch:b13077cb88dbdc200700","metric":"parameters","value":688000000000.0,"unit":"parameters","observationPeriod":"2026-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"MoE architecture, total parameters 688B, active parameters 33B (61 layers: 1 dense + 60 MoE; 256 routed experts + 1 shared, 8 active per token)"},{"id":"epoch:b13077cb88dbdc200700:training-compute","entityId":"epoch:b13077cb88dbdc200700","metric":"training-compute","value":1.8e+24,"unit":"FLOP","observationPeriod":"2026-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"~1.8e24 FLOP reported by SK Telecom via email (July 2026).\n\nCross-check by operation counting: 6 * 33e9 active parameters * 8.5e12 tokens = 1.68e24 FLOP, consistent with the reported figure.\n\nPer the technical report, ~8.2T of the 8.5T tokens were used in pre-training (512 B200 GPUs, ~70 days), with the remainder in post-training."},{"id":"epoch:23c06baf333574e99133:parameters","entityId":"epoch:23c06baf333574e99133","metric":"parameters","value":2400000000000.0,"unit":"parameters","observationPeriod":"2026-07-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.4T total, 95B active"},{"id":"epoch:e344e2c60ab823cd5cc8:parameters","entityId":"epoch:e344e2c60ab823cd5cc8","metric":"parameters","value":2800000000000.0,"unit":"parameters","observationPeriod":"2026-07-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"104 billion active parameters"},{"id":"epoch:e344e2c60ab823cd5cc8:training-compute","entityId":"epoch:e344e2c60ab823cd5cc8","metric":"training-compute","value":2.0001e+25,"unit":"FLOP","observationPeriod":"2026-07-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Based on peer models such as Kimi K2.6 and DeepSeek V4 Pro, plausibly was trained on roughly 30T tokens (likely between 15T and 60T). Using the heuristic that C = 6ND, 30T pretraining tokens would imply ~1.9e25 FLOP. This is a somewhat uncertain estimate (more uncertain than estimates based on disclosed active params and token counts, but comparable to estimates based on disclosed GPU-hours).\n\nK3 was likely pretrained on over 1e25 FLOP. Speculatively, K3's relatively large size compared to other Chinese models may have reduced the training token budget vs smaller models."},{"id":"epoch:cf223bf5394b4dfe83d0:parameters","entityId":"epoch:cf223bf5394b4dfe83d0","metric":"parameters","value":975000000000.0,"unit":"parameters","observationPeriod":"2026-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"975B total, 41B active"},{"id":"epoch:cf223bf5394b4dfe83d0:training-compute","entityId":"epoch:cf223bf5394b4dfe83d0","metric":"training-compute","value":1.1e+25,"unit":"FLOP","observationPeriod":"2026-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 * 41e9 active parameters * 45e12 tokens = 1.1e25 FLOP"},{"id":"epoch:bedf505291d712b2edb4:parameters","entityId":"epoch:bedf505291d712b2edb4","metric":"parameters","value":1500000000000.0,"unit":"parameters","observationPeriod":"2026-07-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"SpaceXAI CEO Elon Musk stated on X that Grok 4.5 would have 1.5 trillion parameters, presumably meaning total parameters. Grok 4.5 is very likely a mixture of experts model.\n\nhttps://x.com/elonmusk/status/2071184354756477041?s=20"},{"id":"epoch:974f4354fa22ffad802e:parameters","entityId":"epoch:974f4354fa22ffad802e","metric":"parameters","value":295000000000.0,"unit":"parameters","observationPeriod":"2026-07-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Hy3 is a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters and 3.8B MTP layer parameters\""},{"id":"epoch:d151aead00e5b4b05a4c:parameters","entityId":"epoch:d151aead00e5b4b05a4c","metric":"parameters","value":1600000000000.0,"unit":"parameters","observationPeriod":"2026-06-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.6T total, 48B active"},{"id":"epoch:d151aead00e5b4b05a4c:training-compute","entityId":"epoch:d151aead00e5b4b05a4c","metric":"training-compute","value":1.008e+25,"unit":"FLOP","observationPeriod":"2026-06-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 * 48e9 active parameters * 35e12 tokens = 1.008e25 FLOP. \n\nConfidence between Likely and Speculative because the 35T is a lower bound and 48B is an aggregate metric since expert size isn't consistently 48B.\n\nHardware quantity of 50k is also a lower bound - \"LongCat-2.0 is pre-trained on over 50K AI ASICs, \""},{"id":"epoch:fbaa821000be951f4440:parameters","entityId":"epoch:fbaa821000be951f4440","metric":"parameters","value":250300000000.0,"unit":"parameters","observationPeriod":"2026-06-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Total: 250.3B, Active: 14.7B per token. Mixture-of-Experts (320 routed experts + 1 shared expert)."},{"id":"epoch:fbaa821000be951f4440:training-compute","entityId":"epoch:fbaa821000be951f4440","metric":"training-compute","value":1.05e+24,"unit":"FLOP","observationPeriod":"2026-06-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Per Upstage in direct correspondence, Solar Open2 250B was trained on 11T tokens. At 14.7B parameters, this implies 9.7e23 FLOP. Solar Open2 250B was trained using 1,656\u00d7 B200 and 720\u00d7 H200 for 1900 hours, costing $32 million (implying a cost of $7/GPU-hour)\n\nTech report clarifies 11.9T pretraining tokens, for 1.05e24 FLOP, with additional post-training.\n\n"},{"id":"epoch:00035b648ca396941276:parameters","entityId":"epoch:00035b648ca396941276","metric":"parameters","value":744000000000.0,"unit":"parameters","observationPeriod":"2026-06-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"40 billion active"},{"id":"epoch:31068fc70d9d3f02c986:parameters","entityId":"epoch:31068fc70d9d3f02c986","metric":"parameters","value":1040000000000.0,"unit":"parameters","observationPeriod":"2026-06-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c2f0c099ee697b271f4c:parameters","entityId":"epoch:c2f0c099ee697b271f4c","metric":"parameters","value":25200000000.0,"unit":"parameters","observationPeriod":"2026-06-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"25.2B total, 3,8B active"},{"id":"epoch:a2dc63a706d89d0dbf5e:parameters","entityId":"epoch:a2dc63a706d89d0dbf5e","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2026-06-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"30B total, 3B active"},{"id":"epoch:65b6f4dca987e7c3880c:parameters","entityId":"epoch:65b6f4dca987e7c3880c","metric":"parameters","value":550000000000.0,"unit":"parameters","observationPeriod":"2026-06-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"550B total parameters with up to 55B active per token\" from https://docs.nvidia.com/nemotron/nightly/usage-cookbook/Nemotron-3-Ultra-Base/README.html "},{"id":"epoch:65b6f4dca987e7c3880c:training-compute","entityId":"epoch:65b6f4dca987e7c3880c","metric":"training-compute","value":6.6e+24,"unit":"FLOP","observationPeriod":"2026-06-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"55 billion active parameters, ~20T tokens"},{"id":"epoch:ecc1e95d828e6c43f840:parameters","entityId":"epoch:ecc1e95d828e6c43f840","metric":"parameters","value":11950000000.0,"unit":"parameters","observationPeriod":"2026-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:229c620dd807bd5eb9b7:parameters","entityId":"epoch:229c620dd807bd5eb9b7","metric":"parameters","value":962000000000.0,"unit":"parameters","observationPeriod":"2026-06-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"34.7B active, 962B total from Table 1"},{"id":"epoch:229c620dd807bd5eb9b7:training-compute","entityId":"epoch:229c620dd807bd5eb9b7","metric":"training-compute","value":6.98511e+24,"unit":"FLOP","observationPeriod":"2026-06-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 * 34.7e9 parameters * 33.55e12 tokens = 6.98511e24 FLOP"},{"id":"epoch:5cdd3bd86a742e20f2ce:parameters","entityId":"epoch:5cdd3bd86a742e20f2ce","metric":"parameters","value":428000000000.0,"unit":"parameters","observationPeriod":"2026-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"428B total, 22B active"},{"id":"epoch:b59a6c306a3e8cd879b2:parameters","entityId":"epoch:b59a6c306a3e8cd879b2","metric":"parameters","value":198000000000.0,"unit":"parameters","observationPeriod":"2026-05-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"198B total, 196B language + 1.8B vision encoder; 11B active. "},{"id":"epoch:74c75a0ddc8db9cc09e3:parameters","entityId":"epoch:74c75a0ddc8db9cc09e3","metric":"parameters","value":218000000000.0,"unit":"parameters","observationPeriod":"2026-05-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"218B total; 25B active\""},{"id":"epoch:ab3acbbdcab183211d9e:parameters","entityId":"epoch:ab3acbbdcab183211d9e","metric":"parameters","value":1040000000000.0,"unit":"parameters","observationPeriod":"2026-05-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Based on Kimi K2.5; 1.04T total, 32B active parameters"},{"id":"epoch:ab3acbbdcab183211d9e:training-compute","entityId":"epoch:ab3acbbdcab183211d9e","metric":"training-compute","value":3.87e+25,"unit":"FLOP","observationPeriod":"2026-05-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Said by Cursor that 85% of the total compute was performed by their team (from figure following \"Composer 2.5 is built on the same open-source checkpoint as Composer 2, Moonshot's Kimi K2.5.\" here https://cursor.com/blog/composer-2-5) - 5.8e24 / 0.15 = 3.8666666667E25 FLOP"},{"id":"epoch:1755cacd27b8a18f0ec5:parameters","entityId":"epoch:1755cacd27b8a18f0ec5","metric":"parameters","value":1000000000000.0,"unit":"parameters","observationPeriod":"2026-05-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1B total, 63B active"},{"id":"epoch:f3f55d7c9260508249d6:parameters","entityId":"epoch:f3f55d7c9260508249d6","metric":"parameters","value":276000000000.0,"unit":"parameters","observationPeriod":"2026-05-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The current TML-Interaction-Small is a 276B parameter MoE with 12B active.\""},{"id":"epoch:ae76e8bae6f942ab5a7b:parameters","entityId":"epoch:ae76e8bae6f942ab5a7b","metric":"parameters","value":800000000000.0,"unit":"parameters","observationPeriod":"2026-05-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"They say it's 1/3rd of the total parameter count of ERNIE 5.0, which was 2.4T (https://ernie.baidu.com/blog/posts/ernie5.0/) "},{"id":"epoch:7ffd46da978d65483d1b:parameters","entityId":"epoch:7ffd46da978d65483d1b","metric":"parameters","value":128000000000.0,"unit":"parameters","observationPeriod":"2026-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:375c245aa976d2a1f6a0:parameters","entityId":"epoch:375c245aa976d2a1f6a0","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2026-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Dense 30B"},{"id":"epoch:375c245aa976d2a1f6a0:training-compute","entityId":"epoch:375c245aa976d2a1f6a0","metric":"training-compute","value":2.7e+24,"unit":"FLOP","observationPeriod":"2026-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 * 30e9 parameters * 15e12 tokens = 2.7e24 FLOP"},{"id":"epoch:cbf60930da30f0c2edd5:parameters","entityId":"epoch:cbf60930da30f0c2edd5","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2026-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Dense  8B"},{"id":"epoch:cbf60930da30f0c2edd5:training-compute","entityId":"epoch:cbf60930da30f0c2edd5","metric":"training-compute","value":7.2e+23,"unit":"FLOP","observationPeriod":"2026-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 * 8e9 parameters * 15e12 tokens = 2.7e24 FLOP"},{"id":"epoch:d13f4ddd46a63f0ef33d:parameters","entityId":"epoch:d13f4ddd46a63f0ef33d","metric":"parameters","value":1600000000000.0,"unit":"parameters","observationPeriod":"2026-04-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.6T total, 49B active"},{"id":"epoch:d13f4ddd46a63f0ef33d:training-compute","entityId":"epoch:d13f4ddd46a63f0ef33d","metric":"training-compute","value":9.702e+24,"unit":"FLOP","observationPeriod":"2026-04-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 * 49e9 active parameters * 33e12 tokens = 9.702E24 for pre-training"},{"id":"epoch:610d3c8957bbd565ab8a:parameters","entityId":"epoch:610d3c8957bbd565ab8a","metric":"parameters","value":284000000000.0,"unit":"parameters","observationPeriod":"2026-04-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"284B total, 13B active"},{"id":"epoch:610d3c8957bbd565ab8a:training-compute","entityId":"epoch:610d3c8957bbd565ab8a","metric":"training-compute","value":2.496e+24,"unit":"FLOP","observationPeriod":"2026-04-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 * 13e9 active parameters * 32e12 tokens = 2.496E24 for pre-training"},{"id":"epoch:b2ce4aa48bedb1dea6c1:parameters","entityId":"epoch:b2ce4aa48bedb1dea6c1","metric":"parameters","value":1020000000000.0,"unit":"parameters","observationPeriod":"2026-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"MiMo-V2.5-Pro is a 1.02T-parameter Mixture-of-Experts model with 42B active parameters, \""},{"id":"epoch:b2ce4aa48bedb1dea6c1:training-compute","entityId":"epoch:b2ce4aa48bedb1dea6c1","metric":"training-compute","value":6.804e+24,"unit":"FLOP","observationPeriod":"2026-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOPS/token/active parameter * 27e12 tokens * 42e9 active parameters = 6.804E24 FLOP"},{"id":"epoch:1c5940505868ae1d5446:parameters","entityId":"epoch:1c5940505868ae1d5446","metric":"parameters","value":295000000000.0,"unit":"parameters","observationPeriod":"2026-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Today, we are releasing and open-sourcing Hy3 preview. It is a 295B-parameter Mixture-of-Experts (MoE) model, with 21B active parameters, \""},{"id":"epoch:2673a798543ca20759d3:parameters","entityId":"epoch:2673a798543ca20759d3","metric":"parameters","value":28000000000.0,"unit":"parameters","observationPeriod":"2026-04-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"27B language parameters, higher number listed on HF safetensors metadata, presumably because of the vision encoder."},{"id":"epoch:ffa72da0a4d9afce4efe:parameters","entityId":"epoch:ffa72da0a4d9afce4efe","metric":"parameters","value":104000000000.0,"unit":"parameters","observationPeriod":"2026-04-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"104B total parameters and 7.4B active parameters."},{"id":"epoch:61087fa55644676c60f4:parameters","entityId":"epoch:61087fa55644676c60f4","metric":"parameters","value":27000000000.0,"unit":"parameters","observationPeriod":"2026-04-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Named"},{"id":"epoch:478f875317e225cf29ed:parameters","entityId":"epoch:478f875317e225cf29ed","metric":"parameters","value":1040000000000.0,"unit":"parameters","observationPeriod":"2026-04-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5a6cd58d35b4e91d8e22:parameters","entityId":"epoch:5a6cd58d35b4e91d8e22","metric":"parameters","value":500000000000.0,"unit":"parameters","observationPeriod":"2026-04-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Supplemental training has been added to 4.3. \n\nGrok 4.4 will be twice the size (1T) with training data through early April. Probably ready for release in early May. \" from Elon (https://x.com/elonmusk/status/2045590599206875216?s=20) implies 500B parameters, assumed to be total #. "},{"id":"epoch:16b92b1531480635224a:parameters","entityId":"epoch:16b92b1531480635224a","metric":"parameters","value":5260000000.0,"unit":"parameters","observationPeriod":"2026-04-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"nitialized from the Gemma3 4B-parameter VLM [106] [...] (including a 400M-parameter vision encoder), and a flow\nmatching action expert with 860M parameters. The model\nhas about 5B total parameters.\""},{"id":"epoch:58f722d5d09e787cd11c:parameters","entityId":"epoch:58f722d5d09e787cd11c","metric":"parameters","value":33200000000.0,"unit":"parameters","observationPeriod":"2026-04-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Architecturally, EXAONE 4.5 integrates a 1.2B parameter vision encoder into the robust EXAONE 4.0 32B base model,\""},{"id":"epoch:58f722d5d09e787cd11c:training-compute","entityId":"epoch:58f722d5d09e787cd11c","metric":"training-compute","value":3.85e+24,"unit":"FLOP","observationPeriod":"2026-04-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From Table 1 in paper - 1.57e23 FLOP + 6.43e22 FLOP = 2.213E23 FLOP\nBut also add the compute from base model EXAONE 4.0"},{"id":"epoch:5b69fde1cce50388aea6:parameters","entityId":"epoch:5b69fde1cce50388aea6","metric":"parameters","value":754000000000.0,"unit":"parameters","observationPeriod":"2026-04-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Based on HF page: https://huggingface.co/zai-org/GLM-5.1"},{"id":"epoch:e72cc9df044c662c4f2c:parameters","entityId":"epoch:e72cc9df044c662c4f2c","metric":"parameters","value":5485309424.0,"unit":"parameters","observationPeriod":"2026-04-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From https://huggingface.co/allenai/MolmoAct2/blob/main/model.safetensors.index.json "},{"id":"epoch:e72cc9df044c662c4f2c:training-compute","entityId":"epoch:e72cc9df044c662c4f2c","metric":"training-compute","value":1.1e+22,"unit":"FLOP","observationPeriod":"2026-04-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Estimate from hardware details: 5760 + 2304 + 2304 = GPU hours across pre-training, post-training and fune-tuning. 16-bit tensor precision (BF16 for \"most operations\" from the paper), 30% utilization assumed."},{"id":"epoch:eae9ecca8b217527600a:parameters","entityId":"epoch:eae9ecca8b217527600a","metric":"parameters","value":31000000000.0,"unit":"parameters","observationPeriod":"2026-04-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:106734caf2c98b05d6aa:parameters","entityId":"epoch:106734caf2c98b05d6aa","metric":"parameters","value":25200000000.0,"unit":"parameters","observationPeriod":"2026-04-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"25.2B total, 3.8B active"},{"id":"epoch:88838f9aff9cf31d5e18:parameters","entityId":"epoch:88838f9aff9cf31d5e18","metric":"parameters","value":1040000000000.0,"unit":"parameters","observationPeriod":"2026-03-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Based on Kimi K2.5; 1.04T total, 32B active parameters"},{"id":"epoch:88838f9aff9cf31d5e18:training-compute","entityId":"epoch:88838f9aff9cf31d5e18","metric":"training-compute","value":2.32e+25,"unit":"FLOP","observationPeriod":"2026-03-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Said by Lee Robinson on X - https://x.com/leerob/status/2035035355364081694 - to be 4x the training compute of the base model, Kimi K2.5; 4 x 5.8e24 = 2.32e25.\n\nB300 speculated based on Colossus 2 confirmation (https://x.com/elonmusk/status/2056422097237283295)."},{"id":"epoch:ddd3b4509a777bf741dc:parameters","entityId":"epoch:ddd3b4509a777bf741dc","metric":"parameters","value":229000000000.0,"unit":"parameters","observationPeriod":"2026-03-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Per HuggingFace page: https://huggingface.co/MiniMaxAI/MiniMax-M2.7"},{"id":"epoch:84a3eecbf78681c14f14:parameters","entityId":"epoch:84a3eecbf78681c14f14","metric":"parameters","value":1020000000000.0,"unit":"parameters","observationPeriod":"2026-03-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Total parameters surpass 1T with 42B active, roughly 3x larger than MiMo-V2-Flash.\" Listed on https://mimo.xiaomi.com/mimo-v2-5-pro/ specifically as 1.02T in the \"Full benchmark results\" table"},{"id":"epoch:e908826740ea38fefd7d:parameters","entityId":"epoch:e908826740ea38fefd7d","metric":"parameters","value":119000000000.0,"unit":"parameters","observationPeriod":"2026-03-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"119B total parameters, with 6B active parameters per token (8B including embedding and output layers).\""},{"id":"epoch:7942801966ec19f644d4:parameters","entityId":"epoch:7942801966ec19f644d4","metric":"parameters","value":120000000000.0,"unit":"parameters","observationPeriod":"2026-03-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The new Super model is a 120B total, 12B active-parameter\""},{"id":"epoch:ac00fefd227a2a020881:parameters","entityId":"epoch:ac00fefd227a2a020881","metric":"parameters","value":35000000000.0,"unit":"parameters","observationPeriod":"2026-02-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"35B total, 3B active"},{"id":"epoch:c75e9f16e73d0b767fb4:parameters","entityId":"epoch:c75e9f16e73d0b767fb4","metric":"parameters","value":122000000000.0,"unit":"parameters","observationPeriod":"2026-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e8a6abf56830297b492f:parameters","entityId":"epoch:e8a6abf56830297b492f","metric":"parameters","value":27000000000.0,"unit":"parameters","observationPeriod":"2026-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2c870bab654dcaa1f35c:parameters","entityId":"epoch:2c870bab654dcaa1f35c","metric":"parameters","value":9000000000.0,"unit":"parameters","observationPeriod":"2026-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ba764ad003e7d977e884:parameters","entityId":"epoch:ba764ad003e7d977e884","metric":"parameters","value":4000000000.0,"unit":"parameters","observationPeriod":"2026-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f904ace180d6e7cbd6e3:parameters","entityId":"epoch:f904ace180d6e7cbd6e3","metric":"parameters","value":2000000000.0,"unit":"parameters","observationPeriod":"2026-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e2f5c8e46181f95fe2bb:parameters","entityId":"epoch:e2f5c8e46181f95fe2bb","metric":"parameters","value":800000000.0,"unit":"parameters","observationPeriod":"2026-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:65a29bb3983edd3a8f17:parameters","entityId":"epoch:65a29bb3983edd3a8f17","metric":"parameters","value":35000000000.0,"unit":"parameters","observationPeriod":"2026-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"35B total, 3B active"},{"id":"epoch:4a4949c193d182667168:parameters","entityId":"epoch:4a4949c193d182667168","metric":"parameters","value":398000000000.0,"unit":"parameters","observationPeriod":"2026-02-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We present the technical report for Arcee Trinity Large, a sparse Mixture-of-Experts model with 400B total parameters and 13B activated per token\""},{"id":"epoch:4a4949c193d182667168:training-compute","entityId":"epoch:4a4949c193d182667168","metric":"training-compute","value":1.326e+24,"unit":"FLOP","observationPeriod":"2026-02-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13e9 active parameters * 17e12 tokens * 6 = 1.326e24 FLOPs estimated"},{"id":"epoch:1f8df1f8d34ae8e608ac:parameters","entityId":"epoch:1f8df1f8d34ae8e608ac","metric":"parameters","value":500000000000.0,"unit":"parameters","observationPeriod":"2026-02-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"This is just our V8 small foundation model, so 500B params\" from Elon Musk on X (https://x.com/elonmusk/status/2023840935285760142?s=20)\n\nLikely sparse"},{"id":"epoch:0393e6c54fad1cd3650a:parameters","entityId":"epoch:0393e6c54fad1cd3650a","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2026-02-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In this paper, we present DreamZero, a 14B robot foundation model\""},{"id":"epoch:dad8d348286df9176cd5:parameters","entityId":"epoch:dad8d348286df9176cd5","metric":"parameters","value":397000000000.0,"unit":"parameters","observationPeriod":"2026-02-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"it comprises 397 billion total parameters\""},{"id":"epoch:6d6a7d85b7337be2fdb9:parameters","entityId":"epoch:6d6a7d85b7337be2fdb9","metric":"parameters","value":744000000000.0,"unit":"parameters","observationPeriod":"2026-02-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"GLM-5 scales to 256 experts and reduces its layer count to 80 to minimize expert parallelism communication overhead. This results in a 744B parameter model\""},{"id":"epoch:6d6a7d85b7337be2fdb9:training-compute","entityId":"epoch:6d6a7d85b7337be2fdb9","metric":"training-compute","value":6.84e+24,"unit":"FLOP","observationPeriod":"2026-02-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 * 40B parameters * 28.5e12 tokens = 6.84e24 FLOP estimated"},{"id":"epoch:22a595cce118707f7898:parameters","entityId":"epoch:22a595cce118707f7898","metric":"parameters","value":229000000000.0,"unit":"parameters","observationPeriod":"2026-02-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"HF lists 229B params\n10B active parameters"},{"id":"epoch:0d782a9fd8f1847707cb:parameters","entityId":"epoch:0d782a9fd8f1847707cb","metric":"parameters","value":1040000000000.0,"unit":"parameters","observationPeriod":"2026-02-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The model comprises 1.04 trillion total parameters\""},{"id":"epoch:0d782a9fd8f1847707cb:training-compute","entityId":"epoch:0d782a9fd8f1847707cb","metric":"training-compute","value":5.8e+24,"unit":"FLOP","observationPeriod":"2026-02-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Per section 3.2, \"The joint pre-training stage continues from a near-end Kimi K2 checkpoint over additional 15T vision-text tokens at 4K sequence length\"\n\nKimi K2 was pretrained on 15T tokens. A near-end checkpoint of K2 was trained into K2.5 over another 15T tokens (or 15.2 to 15.5T tokens). This makes for a total of ~30T.\n\nThere are an additional 1T tokens in \"Stage 1\" used to train the ViT encoder, which is likely much smaller than the full LLM, so we omit those tokens to calculate training compute.\n\n30e12 tokens * 32B parameters * 6 = 5.76e24 FLOP\n\nPer Kimi talk (timestamp 26:00), the 15 trillion tokens of K2.5 were additive, for 30T training tokens overall: https://x.com/Kimi_Moonshot/status/2035011542115393973?s=20"},{"id":"epoch:1c90fdfd5611642e8868:parameters","entityId":"epoch:1c90fdfd5611642e8868","metric":"parameters","value":80000000000.0,"unit":"parameters","observationPeriod":"2026-02-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Finetune of Qwen3-Next-80B-A3B, which has 80 billion parameters"},{"id":"epoch:8752ece372ac9dccdebc:parameters","entityId":"epoch:8752ece372ac9dccdebc","metric":"parameters","value":102000000000.0,"unit":"parameters","observationPeriod":"2026-01-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Solar Pro 3 is Upstage's 102B Mixture-of-Experts large language model, activating only 12B parameters per token\""},{"id":"epoch:79735fc52bf59d0d6c15:parameters","entityId":"epoch:79735fc52bf59d0d6c15","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2026-01-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"GLM-4.7-Flash is a 30B-A3B MoE model.\""},{"id":"epoch:59c94aaee2a0bb196cf7:parameters","entityId":"epoch:59c94aaee2a0bb196cf7","metric":"parameters","value":102000000000.0,"unit":"parameters","observationPeriod":"2025-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"102B, of which 12B active\n"},{"id":"epoch:59c94aaee2a0bb196cf7:training-compute","entityId":"epoch:59c94aaee2a0bb196cf7","metric":"training-compute","value":1.4e+24,"unit":"FLOP","observationPeriod":"2025-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"In direct correspondence with Epoch AI, Upstage confirmed that the training cost of Solar Open 100B was ~$8.4 million.\n\nusing 6 * active parameters * training tokens, training FLOP is 6 * 12 billion * 19 trillion = ~1.4e24 FLOP"},{"id":"epoch:e522d8577d61c7974eac:parameters","entityId":"epoch:e522d8577d61c7974eac","metric":"parameters","value":236000000000.0,"unit":"parameters","observationPeriod":"2025-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"236B"},{"id":"epoch:e522d8577d61c7974eac:training-compute","entityId":"epoch:e522d8577d61c7974eac","metric":"training-compute","value":1.52e+24,"unit":"FLOP","observationPeriod":"2025-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:19250504a73e36311588:parameters","entityId":"epoch:19250504a73e36311588","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2025-12-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Total parameters: 100B (Sparse), Active parameters: 10B per token. It uses a Mixture-of-Experts architecture with 1 shared expert and 128 total experts(top-8 routing)"},{"id":"epoch:5adb9034eed4fe41582b:parameters","entityId":"epoch:5adb9034eed4fe41582b","metric":"parameters","value":519000000000.0,"unit":"parameters","observationPeriod":"2025-12-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"MoE Architecture, Total parameters 519B, Active parameters 33B"},{"id":"epoch:502a6c60bc0e7b37f6e7:parameters","entityId":"epoch:502a6c60bc0e7b37f6e7","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-12-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:af25685532c535bb6df5:parameters","entityId":"epoch:af25685532c535bb6df5","metric":"parameters","value":229000000000.0,"unit":"parameters","observationPeriod":"2025-12-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"HF lists 229B params\n10B active parameters"},{"id":"epoch:86b359dbdb31d629f765:parameters","entityId":"epoch:86b359dbdb31d629f765","metric":"parameters","value":358000000000.0,"unit":"parameters","observationPeriod":"2025-12-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"mixture of experts (MoE) with 358B total, 32B active"},{"id":"epoch:86b359dbdb31d629f765:training-compute","entityId":"epoch:86b359dbdb31d629f765","metric":"training-compute","value":4.42e+24,"unit":"FLOP","observationPeriod":"2025-12-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training tokens not disclosed.\n\nPresuming this is a post-training update from GLM-4.5 (23T tokens) with minor additional compute, training compute is at least:\n\n6 FLOP/parameter/token * 32B active parameters * 23T tokens = 4.42e24 FLOP"},{"id":"epoch:90562e2b3e6e1882347c:parameters","entityId":"epoch:90562e2b3e6e1882347c","metric":"parameters","value":309000000000.0,"unit":"parameters","observationPeriod":"2025-12-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We present MiMo-V2-Flash, a Mixture-of-Experts (MoE) model with 309B total parameters and 15B active parameters\""},{"id":"epoch:90562e2b3e6e1882347c:training-compute","entityId":"epoch:90562e2b3e6e1882347c","metric":"training-compute","value":2.43e+24,"unit":"FLOP","observationPeriod":"2025-12-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 * 15e9 active parameters * 27e12 tokens = 2.43e+24"},{"id":"epoch:0cb37aa18a94add5217a:parameters","entityId":"epoch:0cb37aa18a94add5217a","metric":"parameters","value":31600000000.0,"unit":"parameters","observationPeriod":"2025-12-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Nano 30B-A3B Base contains 31.6B total parameters out of which 3.2B are activ\""},{"id":"epoch:0cb37aa18a94add5217a:training-compute","entityId":"epoch:0cb37aa18a94add5217a","metric":"training-compute","value":4.8e+23,"unit":"FLOP","observationPeriod":"2025-12-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.2e9 active parameters * 25e12 tokens * 6 = 4.8e23 FLOP"},{"id":"epoch:3f4d25858c08a69e4b8b:parameters","entityId":"epoch:3f4d25858c08a69e4b8b","metric":"parameters","value":14800000000.0,"unit":"parameters","observationPeriod":"2025-12-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"using the pretrained [...] Qwen3-14B-Base\""},{"id":"epoch:7f50014373ac0f4a90b2:parameters","entityId":"epoch:7f50014373ac0f4a90b2","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2025-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"MoE specialization of Qwen3-30B-A3B-Thinking-2507, reported ~3B active parameters. Designed to be used with the open-source Nomos reasoning harness."},{"id":"epoch:041bd113b20556507f1c:parameters","entityId":"epoch:041bd113b20556507f1c","metric":"parameters","value":123000000000.0,"unit":"parameters","observationPeriod":"2025-12-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:3f5b59d4408e24610fbe:parameters","entityId":"epoch:3f5b59d4408e24610fbe","metric":"parameters","value":24000000000.0,"unit":"parameters","observationPeriod":"2025-12-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:7a20dacb8e19b70d2d10:parameters","entityId":"epoch:7a20dacb8e19b70d2d10","metric":"parameters","value":24000000000.0,"unit":"parameters","observationPeriod":"2025-12-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"24B dense"},{"id":"epoch:4a79fa47584a43d31fb5:parameters","entityId":"epoch:4a79fa47584a43d31fb5","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2025-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e1ded62ea1482b528879:parameters","entityId":"epoch:e1ded62ea1482b528879","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ba1dcf990da763c6edd4:parameters","entityId":"epoch:ba1dcf990da763c6edd4","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2025-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f047eb7d38b35f0bf091:parameters","entityId":"epoch:f047eb7d38b35f0bf091","metric":"parameters","value":675000000000.0,"unit":"parameters","observationPeriod":"2025-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"a sparse mixture-of-experts trained with 41B active and 675B total parameters.\""},{"id":"epoch:2f3ee310859dc8e22678:training-compute","entityId":"epoch:2f3ee310859dc8e22678","metric":"training-compute","value":4.2e+24,"unit":"FLOP","observationPeriod":"2025-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Pre-trained base was DeepSeek-V3.1-Terminus\n3.594058e+24 FLOP [base model] + 2.095014e+23 FLOP = 3.8035594e+24 FLOP for pre-training plus continued pre-training\n\nThen, \"this framework allocates a\npost-training computational budget exceeding 10% of the pre-training cost\", so 4.18e24"},{"id":"epoch:79de87c2bbcc3ccf5061:parameters","entityId":"epoch:79de87c2bbcc3ccf5061","metric":"parameters","value":685000000000.0,"unit":"parameters","observationPeriod":"2025-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"HF lists 685B params. Built on top of DeepSeek-V3.2-Exp-Base"},{"id":"epoch:91761965cca4e08edfc0:parameters","entityId":"epoch:91761965cca4e08edfc0","metric":"parameters","value":106000000000.0,"unit":"parameters","observationPeriod":"2025-11-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We present INTELLECT-3, a 106B-parameter Mixture-of-Experts model (12B active)\""},{"id":"epoch:33000eaebd8ff22e003c:parameters","entityId":"epoch:33000eaebd8ff22e003c","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-11-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"32B parameter flow matching transformer\" (https://github.com/black-forest-labs/flux2, README; https://bfl.ai/blog/flux-2)"},{"id":"epoch:8d89b09d4179957d70ba:parameters","entityId":"epoch:8d89b09d4179957d70ba","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-11-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B with Base/Instruct/Think variants. Base here"},{"id":"epoch:8d89b09d4179957d70ba:training-compute","entityId":"epoch:8d89b09d4179957d70ba","metric":"training-compute","value":1.1e+24,"unit":"FLOP","observationPeriod":"2025-11-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*3.2e10*5.5e12=1.1e24"},{"id":"epoch:b43df9a83f83be6b5c4f:parameters","entityId":"epoch:b43df9a83f83be6b5c4f","metric":"parameters","value":848000000.0,"unit":"parameters","observationPeriod":"2025-11-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The SAM 3 model has 848M parameters\" (https://www.marktechpost.com/2025/11/20/meta-ai-releases-segment-anything-model-3-sam-3-for-promptable-concept-segmentation-in-images-and-videos/). Uses Perception Encoder (PE) vision backbone shared between detector and tracker."},{"id":"epoch:5691e80707a7c847df44:parameters","entityId":"epoch:5691e80707a7c847df44","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-11-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B and 32B variants. Dense transformer architecture. (https://arxiv.org/abs/2512.13961)"},{"id":"epoch:1f1c7757591daa34a5d1:parameters","entityId":"epoch:1f1c7757591daa34a5d1","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-11-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B and 32B variants. Dense transformer architecture. (https://arxiv.org/abs/2512.13961)"},{"id":"epoch:9421f210cc0f386fb009:parameters","entityId":"epoch:9421f210cc0f386fb009","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-11-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B and 32B variants. Dense transformer architecture. (https://arxiv.org/abs/2512.13961)"},{"id":"epoch:a544fad233f2d0df42e0:parameters","entityId":"epoch:a544fad233f2d0df42e0","metric":"parameters","value":671000000000.0,"unit":"parameters","observationPeriod":"2025-11-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"671B total, 37B active"},{"id":"epoch:4e7f8efdbb96194e95c4:parameters","entityId":"epoch:4e7f8efdbb96194e95c4","metric":"parameters","value":5260000000.0,"unit":"parameters","observationPeriod":"2025-11-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"(ii) The base VLM is Gemma 3 [78] 4B model. (iii) The size of the action expert is increased to 860M parameters.\" Model card (https://website.pi-asset.com/pi06star/PI06_model_card.pdf) notes in Figure 1 that the VLM also includes SigLIP (400M). "},{"id":"epoch:0195d2bfb8f382f1166a:parameters","entityId":"epoch:0195d2bfb8f382f1166a","metric":"parameters","value":235000000000.0,"unit":"parameters","observationPeriod":"2025-11-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"MoE: P1-235B-A22B has 235B total, 22B active parameters. Lightweight variant P1-30B-A3B has 30B total, 3B active. (https://arxiv.org/abs/2511.13612)"},{"id":"epoch:9b593af386e3e85f66a5:parameters","entityId":"epoch:9b593af386e3e85f66a5","metric":"parameters","value":110000000000.0,"unit":"parameters","observationPeriod":"2025-11-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The Kaiju family of models comes in 3 production variants: Small (13B), Medium (34B), and Large (110B).\""},{"id":"epoch:9ac88e88662c3708b719:parameters","entityId":"epoch:9ac88e88662c3708b719","metric":"parameters","value":34000000000.0,"unit":"parameters","observationPeriod":"2025-11-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The Kaiju family of models comes in 3 production variants: Small (13B), Medium (34B), and Large (110B).\""},{"id":"epoch:f684f7af5403a2dcb89b:parameters","entityId":"epoch:f684f7af5403a2dcb89b","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2025-11-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The Kaiju family of models comes in 3 production variants: Small (13B), Medium (34B), and Large (110B).\""},{"id":"epoch:7b8547d6a3308103b345:parameters","entityId":"epoch:7b8547d6a3308103b345","metric":"parameters","value":1000000000000.0,"unit":"parameters","observationPeriod":"2025-11-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Total Parameters\t1T\nActivated Parameters\t32B"},{"id":"epoch:7b8547d6a3308103b345:training-compute","entityId":"epoch:7b8547d6a3308103b345","metric":"training-compute","value":4.2e+24,"unit":"FLOP","observationPeriod":"2025-11-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming the additional post-training contributed between 1% and 100% of Kimi K2's training compute (which we confidently estimate at 2.976e+24), we get a range of 3.0e24 to 6.0e24, and a geometric mean of 4.2e24."},{"id":"epoch:a67654651def79c59d77:parameters","entityId":"epoch:a67654651def79c59d77","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2025-11-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We\u2019ve since scaled GEN-0 to 10B+ model sizes\""},{"id":"epoch:6d1a3d66aa9202513ef4:parameters","entityId":"epoch:6d1a3d66aa9202513ef4","metric":"parameters","value":34100000000.0,"unit":"parameters","observationPeriod":"2025-10-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Overall, the model contains 34.1 billion(B) parameters, including 31.2 B in the transformer layers and 2.9 B in the embedding layers. \""},{"id":"epoch:6d1a3d66aa9202513ef4:training-compute","entityId":"epoch:6d1a3d66aa9202513ef4","metric":"training-compute","value":9.86736e+24,"unit":"FLOP","observationPeriod":"2025-10-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.084799999999999e+24 FLOP [base model compute] + 2.78256e+24 FLOP = 9.86736e+24 FLOP"},{"id":"epoch:b4af9537942bc841be6d:parameters","entityId":"epoch:b4af9537942bc841be6d","metric":"parameters","value":48000000000.0,"unit":"parameters","observationPeriod":"2025-10-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3B activated parameters and 48B total parameters"},{"id":"epoch:b4af9537942bc841be6d:training-compute","entityId":"epoch:b4af9537942bc841be6d","metric":"training-compute","value":1.026e+23,"unit":"FLOP","observationPeriod":"2025-10-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 3000000000 activated parameters * 5700000000000 tokens = 1.026e+23 FLOP"},{"id":"epoch:9bf964cbb93a2d750e4e:parameters","entityId":"epoch:9bf964cbb93a2d750e4e","metric":"parameters","value":300000000000.0,"unit":"parameters","observationPeriod":"2025-10-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"hundreds of billions of parameters\"\nAssuming ~300B with \"Likely\" confidence"},{"id":"epoch:dac367ca21d03903903f:parameters","entityId":"epoch:dac367ca21d03903903f","metric":"parameters","value":30500000000.0,"unit":"parameters","observationPeriod":"2025-10-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"30.5 billion total parameters,\nwith only 3.3 billion activated per token"},{"id":"epoch:b2a8afa9d867574e99de:parameters","entityId":"epoch:b2a8afa9d867574e99de","metric":"parameters","value":229000000000.0,"unit":"parameters","observationPeriod":"2025-10-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"maintaining activations around 10B\"\nsafetensors: 229B params"},{"id":"epoch:ff2eb268e9efa9d5bf47:parameters","entityId":"epoch:ff2eb268e9efa9d5bf47","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-10-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:386109e71cdc80a56570:parameters","entityId":"epoch:386109e71cdc80a56570","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2025-10-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B"},{"id":"epoch:b38f679ebcf125bee813:parameters","entityId":"epoch:b38f679ebcf125bee813","metric":"parameters","value":12000000000.0,"unit":"parameters","observationPeriod":"2025-10-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12B"},{"id":"epoch:b38f679ebcf125bee813:training-compute","entityId":"epoch:b38f679ebcf125bee813","metric":"training-compute","value":8.64e+23,"unit":"FLOP","observationPeriod":"2025-10-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Not including finetuning compute\n"},{"id":"epoch:670df5d18df4ed2e2720:parameters","entityId":"epoch:670df5d18df4ed2e2720","metric":"parameters","value":16400000000.0,"unit":"parameters","observationPeriod":"2025-10-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Ring-mini-linear-2.0 comprises 16B parameters and 957M activations"},{"id":"epoch:670df5d18df4ed2e2720:training-compute","entityId":"epoch:670df5d18df4ed2e2720","metric":"training-compute","value":1.714452e+23,"unit":"FLOP","observationPeriod":"2025-10-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.68e+23 FLOP [base model compute] + 3.4452e+21 FLOP = 1.714452e+23 FLOP"},{"id":"epoch:59b9c7faeded38b756dc:parameters","entityId":"epoch:59b9c7faeded38b756dc","metric":"parameters","value":104200000000.0,"unit":"parameters","observationPeriod":"2025-10-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" Ring-flash-linear-2.0 contains 104B parameters and 6.1B activations"},{"id":"epoch:59b9c7faeded38b756dc:training-compute","entityId":"epoch:59b9c7faeded38b756dc","metric":"training-compute","value":7.686e+23,"unit":"FLOP","observationPeriod":"2025-10-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.32e+23 FLOP [base model compute] + 3.66e+22 FLOP = 7.686e+23 FLOP"},{"id":"epoch:1c52f8d948795ae26e03:parameters","entityId":"epoch:1c52f8d948795ae26e03","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2025-10-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"DeepEncoder is approximately 380M in parameters, mainly composed of an 80M SAM-base [17] and a 300M CLIP-large [29] connected in series. The decoder adopts a 3B MoE [19, 20] architecture with 570M activated parameters.\""},{"id":"epoch:e438575ad2dbda08c43d:parameters","entityId":"epoch:e438575ad2dbda08c43d","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-10-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:b23fc5e1ee932d41ee4b:parameters","entityId":"epoch:b23fc5e1ee932d41ee4b","metric":"parameters","value":102000000000.0,"unit":"parameters","observationPeriod":"2025-10-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"102B"},{"id":"epoch:b23fc5e1ee932d41ee4b:training-compute","entityId":"epoch:b23fc5e1ee932d41ee4b","metric":"training-compute","value":1.1e+23,"unit":"FLOP","observationPeriod":"2025-10-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"trained over 1.1 \u00d7 10^23 FLOPs\" from the abstract"},{"id":"epoch:bf94d0a6aae0d037b431:parameters","entityId":"epoch:bf94d0a6aae0d037b431","metric":"parameters","value":12000000000.0,"unit":"parameters","observationPeriod":"2025-10-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12B"},{"id":"epoch:c63852c76ab3f11c3a9b:parameters","entityId":"epoch:c63852c76ab3f11c3a9b","metric":"parameters","value":1200000000.0,"unit":"parameters","observationPeriod":"2025-10-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.2B"},{"id":"epoch:9eeb546ac898582099b0:parameters","entityId":"epoch:9eeb546ac898582099b0","metric":"parameters","value":109000000000.0,"unit":"parameters","observationPeriod":"2025-10-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"17Bx16 MoE (Scout)\n109B total parameters"},{"id":"epoch:30b8fca504eff135530a:parameters","entityId":"epoch:30b8fca504eff135530a","metric":"parameters","value":27000000000.0,"unit":"parameters","observationPeriod":"2025-10-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"27B"},{"id":"epoch:b292c509592906e8200d:parameters","entityId":"epoch:b292c509592906e8200d","metric":"parameters","value":1000000000000.0,"unit":"parameters","observationPeriod":"2025-10-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1 trillion total parameters with 50 billion activated parameters"},{"id":"epoch:5d5e21c9354e4536eec6:parameters","entityId":"epoch:5d5e21c9354e4536eec6","metric":"parameters","value":1000000000000.0,"unit":"parameters","observationPeriod":"2025-10-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1 trillion total parameters with 50 billion activated parameters"},{"id":"epoch:5d5e21c9354e4536eec6:training-compute","entityId":"epoch:5d5e21c9354e4536eec6","metric":"training-compute","value":6.000001e+24,"unit":"FLOP","observationPeriod":"2025-10-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 50000000000 active parameters * 20000000000000 tokens = 6e+24 FLOP"},{"id":"epoch:de022b5763435a23e563:parameters","entityId":"epoch:de022b5763435a23e563","metric":"parameters","value":7000000.0,"unit":"parameters","observationPeriod":"2025-10-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7M"},{"id":"epoch:de022b5763435a23e563:training-compute","entityId":"epoch:de022b5763435a23e563","metric":"training-compute","value":3.07742976e+20,"unit":"FLOP","observationPeriod":"2025-10-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"989400000000000 FLOP/GPU/sec * 72 hours * 3600 sec / hour * 4 GPUs * 0.3 [assumed utilization] = 3.07742976e20 FLOP"},{"id":"epoch:43d2037f33c42f14373f:parameters","entityId":"epoch:43d2037f33c42f14373f","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-10-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"a hybrid MoE with 7B total parameters (1B active)"},{"id":"epoch:43d2037f33c42f14373f:training-compute","entityId":"epoch:43d2037f33c42f14373f","metric":"training-compute","value":1.35e+23,"unit":"FLOP","observationPeriod":"2025-10-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 1000000000 active parameters * 22500000000000 tokens = 1.35e+23 FLOP"},{"id":"epoch:e0e790b0e05a3b7d061a:parameters","entityId":"epoch:e0e790b0e05a3b7d061a","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2025-10-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" dense hybrid model with 3B parameters."},{"id":"epoch:e0e790b0e05a3b7d061a:training-compute","entityId":"epoch:e0e790b0e05a3b7d061a","metric":"training-compute","value":3.15e+23,"unit":"FLOP","observationPeriod":"2025-10-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 3000000000 parameters * 17500000000000 tokens = 3.15e+23 FLOP"},{"id":"epoch:dcb5f85853791d6212e8:parameters","entityId":"epoch:dcb5f85853791d6212e8","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-10-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"a hybrid mixture of experts (MoE) model with 32B total parameters (9B active)"},{"id":"epoch:dcb5f85853791d6212e8:training-compute","entityId":"epoch:dcb5f85853791d6212e8","metric":"training-compute","value":1.215e+24,"unit":"FLOP","observationPeriod":"2025-10-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 9000000000 active parameters * 22500000000000 tokens = 1.215e+24 FLOP"},{"id":"epoch:e757efa786032276c02b:parameters","entityId":"epoch:e757efa786032276c02b","metric":"parameters","value":357000000000.0,"unit":"parameters","observationPeriod":"2025-09-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Similarly to GLM 4.5: 355 billion total parameters (reported) with 32 billion active parameters (assumed)"},{"id":"epoch:e757efa786032276c02b:training-compute","entityId":"epoch:e757efa786032276c02b","metric":"training-compute","value":4.42e+24,"unit":"FLOP","observationPeriod":"2025-09-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 32000000000 active parameters [very likely assumption - everything else is reported to be same as at GLM 4.5] * 23000000000000 tokens = 4.42e24 FLOP"},{"id":"epoch:62c7cc06f36f17afd291:parameters","entityId":"epoch:62c7cc06f36f17afd291","metric":"parameters","value":2000000000.0,"unit":"parameters","observationPeriod":"2025-09-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2B"},{"id":"epoch:ce2b4f27cbe29c1552f2:parameters","entityId":"epoch:ce2b4f27cbe29c1552f2","metric":"parameters","value":2000000000.0,"unit":"parameters","observationPeriod":"2025-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:4258e3c4faa2a2c18348:parameters","entityId":"epoch:4258e3c4faa2a2c18348","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2025-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:93d8290871eea0671e7a:parameters","entityId":"epoch:93d8290871eea0671e7a","metric":"parameters","value":2000000000.0,"unit":"parameters","observationPeriod":"2025-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:cfb2217fc520033ecfd6:parameters","entityId":"epoch:cfb2217fc520033ecfd6","metric":"parameters","value":671000000000.0,"unit":"parameters","observationPeriod":"2025-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"671B total, 37B active (assuming same as in the base model)"},{"id":"epoch:cfb2217fc520033ecfd6:training-compute","entityId":"epoch:cfb2217fc520033ecfd6","metric":"training-compute","value":4.18e+24,"unit":"FLOP","observationPeriod":"2025-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.594058e+24 FLOP [base model] + 2.095014e+23 FLOP = 3.8035594e+24 FLOP for pre-training plus continued pre-training\n\nThen, \"this framework allocates a\npost-training computational budget exceeding 10% of the pre-training cost\", so 4.18e24"},{"id":"epoch:e4ca585901f1cab801fa:parameters","entityId":"epoch:e4ca585901f1cab801fa","metric":"parameters","value":1200000000.0,"unit":"parameters","observationPeriod":"2025-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5b9dbd29b67a420d9cc5:parameters","entityId":"epoch:5b9dbd29b67a420d9cc5","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2025-09-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3B"},{"id":"epoch:63b3122d44f845c08b46:parameters","entityId":"epoch:63b3122d44f845c08b46","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2025-09-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3B"},{"id":"epoch:63b3122d44f845c08b46:training-compute","entityId":"epoch:63b3122d44f845c08b46","metric":"training-compute","value":2e+21,"unit":"FLOP","observationPeriod":"2025-09-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 3000000000 parameters * 2227200000 amino acids [see dataset size notes] * 141 epochs = 5.6526336e+21 FLOP\n\nfrom figure 4 ~2e+12 GFLOPs = 2e+21 FLOP"},{"id":"epoch:e9c33511272b902137d2:parameters","entityId":"epoch:e9c33511272b902137d2","metric":"parameters","value":671000000000.0,"unit":"parameters","observationPeriod":"2025-09-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"671B total, 37B active"},{"id":"epoch:e9c33511272b902137d2:training-compute","entityId":"epoch:e9c33511272b902137d2","metric":"training-compute","value":3.594058e+24,"unit":"FLOP","observationPeriod":"2025-09-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"This was based on V3 which was trained on 14.8T tokens for a total of 14.8T + 840B = 15.64T pretraining tokens. Adding the FLOPs from continued pretraining onto our estimate for V3, which is 3.4e24 FLOPs, we get 3.4e24 + 1.9e23 = 3.59e24"},{"id":"epoch:9599cb54300ad3418fa9:parameters","entityId":"epoch:9599cb54300ad3418fa9","metric":"parameters","value":35300000000.0,"unit":"parameters","observationPeriod":"2025-09-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Audio Encoder AuT 650M \nVision Encoder SigLIP2-So400M 540M \nThinker MoE Transformer 30B-A3B\nTalker MoE Transformer 3B-A0.3B\nMTP Dense Transformer 80M\nCode2wav ConvNet 200M"},{"id":"epoch:9599cb54300ad3418fa9:training-compute","entityId":"epoch:9599cb54300ad3418fa9","metric":"training-compute","value":3.6e+22,"unit":"FLOP","observationPeriod":"2025-09-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 3000000000 active parameters * 2000000000000 tokens = 3.6e+22 FLOP"},{"id":"epoch:167dd39626e060acda7c:parameters","entityId":"epoch:167dd39626e060acda7c","metric":"parameters","value":35300000000.0,"unit":"parameters","observationPeriod":"2025-09-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Audio Encoder AuT 650M \nVision Encoder SigLIP2-So400M 540M \nThinker MoE Transformer 30B-A3B\nTalker MoE Transformer 3B-A0.3B\nMTP Dense Transformer 80M\nCode2wav ConvNet 200M"},{"id":"epoch:167dd39626e060acda7c:training-compute","entityId":"epoch:167dd39626e060acda7c","metric":"training-compute","value":3.6e+22,"unit":"FLOP","observationPeriod":"2025-09-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 3000000000 active parameters * 2000000000000 tokens = 3.6e+22 FLOP"},{"id":"epoch:50f4973a4fef719479ff:parameters","entityId":"epoch:50f4973a4fef719479ff","metric":"parameters","value":24000000000.0,"unit":"parameters","observationPeriod":"2025-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"24B"},{"id":"epoch:a6b56a032c09298adfc7:parameters","entityId":"epoch:a6b56a032c09298adfc7","metric":"parameters","value":258000000.0,"unit":"parameters","observationPeriod":"2025-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"258M"},{"id":"epoch:f1047f616e51b6111bfa:parameters","entityId":"epoch:f1047f616e51b6111bfa","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2025-09-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"30B-A3B"},{"id":"epoch:f1047f616e51b6111bfa:training-compute","entityId":"epoch:f1047f616e51b6111bfa","metric":"training-compute","value":6.5367e+23,"unit":"FLOP","observationPeriod":"2025-09-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6.48e+23 FLOP [base model compute] + 5.67e+21 FLOP = 6.5367e+23 FLOP"},{"id":"epoch:f7b24f31855253b648ff:parameters","entityId":"epoch:f7b24f31855253b648ff","metric":"parameters","value":80000000000.0,"unit":"parameters","observationPeriod":"2025-09-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"80B - A3B"},{"id":"epoch:f7b24f31855253b648ff:training-compute","entityId":"epoch:f7b24f31855253b648ff","metric":"training-compute","value":2.7e+23,"unit":"FLOP","observationPeriod":"2025-09-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 3 * 10^9 active parameters * 15 * 10^12 pre-training tokens = 2.7e+23 FLOP\n\n\"It uses less than 80% of the GPU hours needed by Qwen3-30A-3B, and only 9.3% of the compute cost of Qwen3-32B \u2014 while achieving better performance.\"\n\n(Qwen3-30A-3B compute estimation: 6.48e+23 FLOP, Qwen3-32B - 7.0848e+24 FLOP)"},{"id":"epoch:bba229caf04b81395135:parameters","entityId":"epoch:bba229caf04b81395135","metric":"parameters","value":16000000000.0,"unit":"parameters","observationPeriod":"2025-09-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"It has 16B total parameters, but only 1.4B are activated per input token (non-embedding 789M)."},{"id":"epoch:bba229caf04b81395135:training-compute","entityId":"epoch:bba229caf04b81395135","metric":"training-compute","value":1.68e+23,"unit":"FLOP","observationPeriod":"2025-09-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 1400000000 active parameters * 20000000000000 tokens = 1.68e+23 FLOP\n\n"},{"id":"epoch:21de0ab39830c91166d0:parameters","entityId":"epoch:21de0ab39830c91166d0","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2025-09-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"100B total parameters and 6.1B activated parameters (4.8B non-embedding). "},{"id":"epoch:21de0ab39830c91166d0:training-compute","entityId":"epoch:21de0ab39830c91166d0","metric":"training-compute","value":7.32e+23,"unit":"FLOP","observationPeriod":"2025-09-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 6100000000 active parameters * 20000000000000 tokens = 7.32e+23 FLOP\n\n"},{"id":"epoch:d69d34f4b2915384ac4a:parameters","entityId":"epoch:d69d34f4b2915384ac4a","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-09-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:b7a78e7c8b88087cc748:parameters","entityId":"epoch:b7a78e7c8b88087cc748","metric":"parameters","value":1000000000000.0,"unit":"parameters","observationPeriod":"2025-09-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"MoE architecture"},{"id":"epoch:b7a78e7c8b88087cc748:training-compute","entityId":"epoch:b7a78e7c8b88087cc748","metric":"training-compute","value":1.512e+25,"unit":"FLOP","observationPeriod":"2025-09-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND with:\n36T tokens is taken from the qwen3 technical report\n70B active params is based on it having >1T params, and the architectures of Qwen3-235B-A22B and Qwen3-Coder-480B-A35B"},{"id":"epoch:cbd9223c0b4c9a5086cc:parameters","entityId":"epoch:cbd9223c0b4c9a5086cc","metric":"parameters","value":308000000.0,"unit":"parameters","observationPeriod":"2025-09-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"300M"},{"id":"epoch:ed22cb6c71d9038e986e:parameters","entityId":"epoch:ed22cb6c71d9038e986e","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-09-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:b79ea51b073432776b2e:parameters","entityId":"epoch:b79ea51b073432776b2e","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2025-09-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:b79ea51b073432776b2e:training-compute","entityId":"epoch:b79ea51b073432776b2e","metric":"training-compute","value":6.74e+24,"unit":"FLOP","observationPeriod":"2025-09-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6.74 \u00b7 10^24 FLOPs - reported\n\n6 FLOP / parameter / token * 70 * 10^9 parameters * 15 * 10^12 tokens = 6.3e+24 FLOP \n\n989500000000000 FLOP / GPU / sec [GH200, bf16 reported] * 6*10^6 GPU-hours * 3600 sec / hour * 0.3 [assumed utilization] = 6.41196e+24 FLOP"},{"id":"epoch:3ac59eaa9c1b367eb1e3:parameters","entityId":"epoch:3ac59eaa9c1b367eb1e3","metric":"parameters","value":560000000000.0,"unit":"parameters","observationPeriod":"2025-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"560 billion total parameters, featuring an innovative Mixture-of-Experts (MoE) architecture. The model incorporates a dynamic computation mechanism that activates 18.6B\u223c31.3B parameters (averaging\u223c27B)\""},{"id":"epoch:3ac59eaa9c1b367eb1e3:training-compute","entityId":"epoch:3ac59eaa9c1b367eb1e3","metric":"training-compute","value":3.726e+24,"unit":"FLOP","observationPeriod":"2025-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 27000000000 active parameters * 23000000000000 tokens [\"Likely\" confidence, see dataset size notes] = 3.726e+24 FLOP"},{"id":"epoch:a548a7536fd64102902f:parameters","entityId":"epoch:a548a7536fd64102902f","metric":"parameters","value":27000000000.0,"unit":"parameters","observationPeriod":"2025-08-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Each expert model has about 14B parameters, resulting in a total of 27B parameters but only 14B active parameters per step"},{"id":"epoch:a9c9ca9a2c572228497d:parameters","entityId":"epoch:a9c9ca9a2c572228497d","metric":"parameters","value":671000000000.0,"unit":"parameters","observationPeriod":"2025-08-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"671B total, 37B active"},{"id":"epoch:a9c9ca9a2c572228497d:training-compute","entityId":"epoch:a9c9ca9a2c572228497d","metric":"training-compute","value":3.594058e+24,"unit":"FLOP","observationPeriod":"2025-08-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"This was based on V3 which was trained on 14.8T tokens for a total of 14.8T + 840B = 15.64T pretraining tokens. Adding the FLOPs from continued pretraining onto our estimate for V3, which is 3.4e24 FLOPs, we get 3.4e24 + 1.9e23 = 3.59e24"},{"id":"epoch:c9010661ce113486201d:training-compute","entityId":"epoch:c9010661ce113486201d","metric":"training-compute","value":2.592e+24,"unit":"FLOP","observationPeriod":"2025-08-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 36 * 10^9 parameters * 12 * 10^12 tokens = 2.592e+24 FLOP"},{"id":"epoch:91dae98968ba4f30dde3:parameters","entityId":"epoch:91dae98968ba4f30dde3","metric":"parameters","value":111000000000.0,"unit":"parameters","observationPeriod":"2025-08-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9618fcf4c166f022b886:parameters","entityId":"epoch:9618fcf4c166f022b886","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-08-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:9618fcf4c166f022b886:training-compute","entityId":"epoch:9618fcf4c166f022b886","metric":"training-compute","value":2.1444092e+23,"unit":"FLOP","observationPeriod":"2025-08-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 7000000000 parameters * 4000000000000 tokens = 1.68e+23 FLOP\n\n312000000000000 FLOP/GPU/sec * 812321 GPU-hours * 3600 sec / hour * 0.3 [assumed utilization] = 2.7371968415999997e+23 FLOP\n\nsqrt(1.68e+23 * 2.7371968415999997e+23) = 2.1444092e+23 FLOP"},{"id":"epoch:172cb9c19ab266274bb9:parameters","entityId":"epoch:172cb9c19ab266274bb9","metric":"parameters","value":366000000.0,"unit":"parameters","observationPeriod":"2025-08-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"366M"},{"id":"epoch:172cb9c19ab266274bb9:training-compute","entityId":"epoch:172cb9c19ab266274bb9","metric":"training-compute","value":2.9474214e+21,"unit":"FLOP","observationPeriod":"2025-08-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 366 * 10^6 parameters * 1342177300000 tokens [see dataset size notes] = 2.9474214e+21 FLOP"},{"id":"epoch:a3a156d63aaaec19950f:parameters","entityId":"epoch:a3a156d63aaaec19950f","metric":"parameters","value":27000000000.0,"unit":"parameters","observationPeriod":"2025-08-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a3a156d63aaaec19950f:training-compute","entityId":"epoch:a3a156d63aaaec19950f","metric":"training-compute","value":2.349e+24,"unit":"FLOP","observationPeriod":"2025-08-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.268e+24 FLOP [base model] + 8.1e+22 FLOP = 2.349e+24 FLOP"},{"id":"epoch:66d8ce96cfbb58df4488:parameters","entityId":"epoch:66d8ce96cfbb58df4488","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-08-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7M"},{"id":"epoch:66d8ce96cfbb58df4488:training-compute","entityId":"epoch:66d8ce96cfbb58df4488","metric":"training-compute","value":5.88e+16,"unit":"FLOP","observationPeriod":"2025-08-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 7000000000 parameters * 1400000 tokens = 58800000000000000 FLOP"},{"id":"epoch:631f430a021e19d56bed:parameters","entityId":"epoch:631f430a021e19d56bed","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-08-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:a81159aaa0b7e0bf8bb2:parameters","entityId":"epoch:a81159aaa0b7e0bf8bb2","metric":"parameters","value":9000000000.0,"unit":"parameters","observationPeriod":"2025-08-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9B"},{"id":"epoch:a81159aaa0b7e0bf8bb2:training-compute","entityId":"epoch:a81159aaa0b7e0bf8bb2","metric":"training-compute","value":1.53e+24,"unit":"FLOP","observationPeriod":"2025-08-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Cumulative compute : 1.53E+24 FLOPS [reported]\n\n6 FLOP/parameter/token * 9000000000 parameters * 21100000000000 tokens = 1.1394e+24 FLOP"},{"id":"epoch:dd834757df0271fb6f82:parameters","entityId":"epoch:dd834757df0271fb6f82","metric":"parameters","value":12000000000.0,"unit":"parameters","observationPeriod":"2025-08-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12B"},{"id":"epoch:dd834757df0271fb6f82:training-compute","entityId":"epoch:dd834757df0271fb6f82","metric":"training-compute","value":1.5192e+24,"unit":"FLOP","observationPeriod":"2025-08-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 12000000000 parameters * 21100000000000 tokens = 1.5192e+24 FLOP"},{"id":"epoch:2abfa80cd8afc6d4cb90:parameters","entityId":"epoch:2abfa80cd8afc6d4cb90","metric":"parameters","value":27000000000.0,"unit":"parameters","observationPeriod":"2025-08-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Built upon our 20B Qwen-Image model\"\nQwen Image has 27B parameters though.  20B is just the size of the main module (MMDiT)"},{"id":"epoch:33b9484926673363a0c2:parameters","entityId":"epoch:33b9484926673363a0c2","metric":"parameters","value":9000000000.0,"unit":"parameters","observationPeriod":"2025-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9B"},{"id":"epoch:c621b08818663c7530ea:parameters","entityId":"epoch:c621b08818663c7530ea","metric":"parameters","value":2000000000.0,"unit":"parameters","observationPeriod":"2025-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2B"},{"id":"epoch:dbc71839149acaebe641:parameters","entityId":"epoch:dbc71839149acaebe641","metric":"parameters","value":108000000000.0,"unit":"parameters","observationPeriod":"2025-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"108 total, 12 B active"},{"id":"epoch:dbc71839149acaebe641:training-compute","entityId":"epoch:dbc71839149acaebe641","metric":"training-compute","value":1.8e+24,"unit":"FLOP","observationPeriod":"2025-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.6560000000000004e+24 FLOP [backbone model compute] + 1.44e+23 FLOP = 1.8e+24 FLOP"},{"id":"epoch:3ab31778ebbad6d7ee6c:parameters","entityId":"epoch:3ab31778ebbad6d7ee6c","metric":"parameters","value":9000000000.0,"unit":"parameters","observationPeriod":"2025-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9B"},{"id":"epoch:3ab31778ebbad6d7ee6c:training-compute","entityId":"epoch:3ab31778ebbad6d7ee6c","metric":"training-compute","value":9.18e+23,"unit":"FLOP","observationPeriod":"2025-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8.1e+23 FLOP [backbone model compute, \"Likely\" confidence] + 1.08e+23 FLOP [\"Confident\"] = 9.18e+23 FLOP"},{"id":"epoch:50eb697237ff8f90ccb9:parameters","entityId":"epoch:50eb697237ff8f90ccb9","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2025-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1B"},{"id":"epoch:53fc556c240d35af99db:parameters","entityId":"epoch:53fc556c240d35af99db","metric":"parameters","value":600000.0,"unit":"parameters","observationPeriod":"2025-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"0.6B"},{"id":"epoch:6761af6bf1cef2ba3649:parameters","entityId":"epoch:6761af6bf1cef2ba3649","metric":"parameters","value":270000000.0,"unit":"parameters","observationPeriod":"2025-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"270M\nCompact and capable architecture: Our new model has a total of 270 million parameters: 170 million embedding parameters due to a large vocabulary size and 100 million for our transformer blocks."},{"id":"epoch:6761af6bf1cef2ba3649:training-compute","entityId":"epoch:6761af6bf1cef2ba3649","metric":"training-compute","value":9.72e+21,"unit":"FLOP","observationPeriod":"2025-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 270 * 10^9 parameters * 6 * 10^12 tokens = 9.72e+21 FLOP"},{"id":"epoch:092a1011beca69a8240d:parameters","entityId":"epoch:092a1011beca69a8240d","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-08-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we successfully train a DINO model with 7B parameters\" (https://arxiv.org/abs/2508.10104, Section 1). Family includes distilled ViT-S/B/L/H+ and ConvNeXt variants."},{"id":"epoch:9dc79002375c3984662a:training-compute","entityId":"epoch:9dc79002375c3984662a","metric":"training-compute","value":6.6e+25,"unit":"FLOP","observationPeriod":"2025-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Fairly speculative, possibly around 6e25 [CI: 2e25 to 2e26] FLOP. See document below for details\n\nhttps://docs.google.com/document/d/1V2jIk365LnhH4WDoCw5dYJjZr1Htw8IHaK1noMf5Y48/edit?tab=t.z871imftkus"},{"id":"epoch:cb6481e6c1c0164f34fd:parameters","entityId":"epoch:cb6481e6c1c0164f34fd","metric":"parameters","value":116830000000.0,"unit":"parameters","observationPeriod":"2025-08-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Total parameters: 116.83B"},{"id":"epoch:cb6481e6c1c0164f34fd:training-compute","entityId":"epoch:cb6481e6c1c0164f34fd","metric":"training-compute","value":4.94e+24,"unit":"FLOP","observationPeriod":"2025-08-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The training run for gpt-oss-120b required 2.1 million H100-hours to complete\"\n(2.1e6 hours)*(1,979 H100 FLOP/s)*(30% utilization)*(60*60) = 4.49e24\nThey also do post training similar to o3, which we assume adds at least 10% as much compute, so we multiply this estimate by 1.1 to get 4.94e24\n\n"},{"id":"epoch:73edcf2c54f4d6411f96:parameters","entityId":"epoch:73edcf2c54f4d6411f96","metric":"parameters","value":20910000000.0,"unit":"parameters","observationPeriod":"2025-08-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Total parameters: 20.91B"},{"id":"epoch:73edcf2c54f4d6411f96:training-compute","entityId":"epoch:73edcf2c54f4d6411f96","metric":"training-compute","value":5.49e+23,"unit":"FLOP","observationPeriod":"2025-08-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The training run for gpt-oss-120b required 2.1\nmillion H100-hours to complete, with gpt-oss-20b needing almost 10x fewer\"\nassuming \"almost 10x fewer\" means ~9x fewer: 4.94e24/9 = 5.49e23"},{"id":"epoch:77d80db11953a2a70586:parameters","entityId":"epoch:77d80db11953a2a70586","metric":"parameters","value":27000000000.0,"unit":"parameters","observationPeriod":"2025-08-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:e84540b76fe6639b6c78:parameters","entityId":"epoch:e84540b76fe6639b6c78","metric":"parameters","value":27000000.0,"unit":"parameters","observationPeriod":"2025-08-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"27M"},{"id":"epoch:e728095ecd4e8d89d146:parameters","entityId":"epoch:e728095ecd4e8d89d146","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:fe7cd60c377e4f9ac553:parameters","entityId":"epoch:fe7cd60c377e4f9ac553","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5b5aea282185703720a2:parameters","entityId":"epoch:5b5aea282185703720a2","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2025-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"72B"},{"id":"epoch:0d4fe2d01b38b168c9ea:parameters","entityId":"epoch:0d4fe2d01b38b168c9ea","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:3f86ff76b79cdfe3a0c6:parameters","entityId":"epoch:3f86ff76b79cdfe3a0c6","metric":"parameters","value":20730000000.0,"unit":"parameters","observationPeriod":"2025-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"20.73B\n"},{"id":"epoch:3f86ff76b79cdfe3a0c6:training-compute","entityId":"epoch:3f86ff76b79cdfe3a0c6","metric":"training-compute","value":2.95e+23,"unit":"FLOP","observationPeriod":"2025-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.95E+23 FLOPs (reported)\n\n6 FLOP/parameter/token * 20730000000 parameters * 2300000000000 tokens = 2.86074e+23 FLOP"},{"id":"epoch:b9caee3821b5bca9c7f8:parameters","entityId":"epoch:b9caee3821b5bca9c7f8","metric":"parameters","value":112000000000.0,"unit":"parameters","observationPeriod":"2025-07-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"112B\nThis is a vision-language model that uses a language model based on Command A paired with the SigLIP2-patch16-512 vision encoder through a multimodal adapter for vision-language understanding."},{"id":"epoch:7f71fed75d8e1a533f78:parameters","entityId":"epoch:7f71fed75d8e1a533f78","metric":"parameters","value":480000000.0,"unit":"parameters","observationPeriod":"2025-07-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We trained \u223c1B and \u223c480M parameter variants of AEF, and ultimately proceeded with the smaller variant for improved inference efficiency.\""},{"id":"epoch:7f71fed75d8e1a533f78:training-compute","entityId":"epoch:7f71fed75d8e1a533f78","metric":"training-compute","value":8.515584e+21,"unit":"FLOP","observationPeriod":"2025-07-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"275000000000000 FLOP / sec / TPU v4 * 512 TPUs * 56 hours * 3600 sec / hour * 0.3 [assumed utilization] = 8.515584E21 FLOP"},{"id":"epoch:3ac3df690ca7b84d945c:parameters","entityId":"epoch:3ac3df690ca7b84d945c","metric":"parameters","value":49000000000.0,"unit":"parameters","observationPeriod":"2025-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9397530c23042a7a1e58:parameters","entityId":"epoch:9397530c23042a7a1e58","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2025-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"30 billion total parameters and 3 billion activated parameters\n\nNumber of Layers: 48\nNumber of Attention Heads (GQA): 32 for Q and 4 for KV\nNumber of Experts: 128\nNumber of Activated Experts: 8\nContext Length: 32,768"},{"id":"epoch:9397530c23042a7a1e58:training-compute","entityId":"epoch:9397530c23042a7a1e58","metric":"training-compute","value":6.48e+23,"unit":"FLOP","observationPeriod":"2025-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 3*10^9 active parameters * 36000000000000 tokens = 6.48e+23 FLOP"},{"id":"epoch:3f5556a01570b7c2101a:parameters","entityId":"epoch:3f5556a01570b7c2101a","metric":"parameters","value":355000000000.0,"unit":"parameters","observationPeriod":"2025-07-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"355 billion total parameters with 32 billion active parameters"},{"id":"epoch:3f5556a01570b7c2101a:training-compute","entityId":"epoch:3f5556a01570b7c2101a","metric":"training-compute","value":4.42e+24,"unit":"FLOP","observationPeriod":"2025-07-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(6 FLOP/token/parameter) * (23T tokens) * (32B active parameters) = 4.42e24 FLOP"},{"id":"epoch:5d19d4a0aea61ebe1e05:parameters","entityId":"epoch:5d19d4a0aea61ebe1e05","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2025-07-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B"},{"id":"epoch:5d19d4a0aea61ebe1e05:training-compute","entityId":"epoch:5d19d4a0aea61ebe1e05","metric":"training-compute","value":4.2e+23,"unit":"FLOP","observationPeriod":"2025-07-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 14 * 10^9 active parameters * 5 * 10^12 tokens [assumed, see dataset size notes] = 4.2e+23 FLOP"},{"id":"epoch:17e9da64ac3152545d37:parameters","entityId":"epoch:17e9da64ac3152545d37","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2025-07-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B"},{"id":"epoch:33a3c8e573f59cae5573:parameters","entityId":"epoch:33a3c8e573f59cae5573","metric":"parameters","value":106000000000.0,"unit":"parameters","observationPeriod":"2025-07-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"106B parameters, 12B active"},{"id":"epoch:33a3c8e573f59cae5573:training-compute","entityId":"epoch:33a3c8e573f59cae5573","metric":"training-compute","value":1.656e+24,"unit":"FLOP","observationPeriod":"2025-07-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 12 * 10^9 active parameters * 23 * 10^12 tokens = 1.656e+24 FLOP"},{"id":"epoch:86cfc26ae60788dd6fe1:parameters","entityId":"epoch:86cfc26ae60788dd6fe1","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-07-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:897db880ed75e877411d:parameters","entityId":"epoch:897db880ed75e877411d","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-07-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:aa4abb71e847f49243d0:parameters","entityId":"epoch:aa4abb71e847f49243d0","metric":"parameters","value":235000000000.0,"unit":"parameters","observationPeriod":"2025-07-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"235 billion total parameters and 22 billion activated parameters\n\nNumber of Layers: 94\nNumber of Attention Heads (GQA): 64 for Q and 4 for KV\nNumber of Experts: 128\nNumber of Activated Experts: 8\nContext Length: 32,768 natively and 131,072 tokens with YaRN."},{"id":"epoch:aa4abb71e847f49243d0:training-compute","entityId":"epoch:aa4abb71e847f49243d0","metric":"training-compute","value":4.752e+24,"unit":"FLOP","observationPeriod":"2025-07-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 22*10^9 active parameters * 36000000000000 tokens = 4.752e+24 FLOP"},{"id":"epoch:0a302eaaff7a5c0c4ea0:parameters","entityId":"epoch:0a302eaaff7a5c0c4ea0","metric":"parameters","value":235000000000.0,"unit":"parameters","observationPeriod":"2025-07-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"235 billion total parameters and 22 billion activated parameters\n\nNumber of Layers: 94\nNumber of Attention Heads (GQA): 64 for Q and 4 for KV\nNumber of Experts: 128\nNumber of Activated Experts: 8\nContext Length: 32,768 natively and 131,072 tokens with YaRN."},{"id":"epoch:0a302eaaff7a5c0c4ea0:training-compute","entityId":"epoch:0a302eaaff7a5c0c4ea0","metric":"training-compute","value":4.752e+24,"unit":"FLOP","observationPeriod":"2025-07-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 22*10^9 active parameters * 36000000000000 tokens = 4.752e+24 FLOP"},{"id":"epoch:fe82a5d3b9350c3ead35:parameters","entityId":"epoch:fe82a5d3b9350c3ead35","metric":"parameters","value":24000000000.0,"unit":"parameters","observationPeriod":"2025-07-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"24B\n\n\"We finetuned Mistral Small 3 Instruct (a 24-billion parameter model) for 4 epochs\""},{"id":"epoch:dc1fe5a0dae754e8866b:training-compute","entityId":"epoch:dc1fe5a0dae754e8866b","metric":"training-compute","value":2.2875955e+21,"unit":"FLOP","observationPeriod":"2025-07-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"197000000000000 FLOP / chip / sec * 64 chips * 168 hours * 3600 sec / hour * 0.3 [assumed utilization] = 2.2875955e+21 FLOP"},{"id":"epoch:0f68009569bcdd5d651a:parameters","entityId":"epoch:0f68009569bcdd5d651a","metric":"parameters","value":480000000000.0,"unit":"parameters","observationPeriod":"2025-07-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"a 480B-parameter Mixture-of-Experts model with 35B active parameters which supports the context length of 256K tokens natively and 1M tokens with extrapolation methods\""},{"id":"epoch:0f68009569bcdd5d651a:training-compute","entityId":"epoch:0f68009569bcdd5d651a","metric":"training-compute","value":1.575e+24,"unit":"FLOP","observationPeriod":"2025-07-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 35 * 10^9 active parameters * 7.5 * 10^12 tokens = 1.575e+24 FLOP`"},{"id":"epoch:18682002f407858f58c3:parameters","entityId":"epoch:18682002f407858f58c3","metric":"parameters","value":4000000000.0,"unit":"parameters","observationPeriod":"2025-07-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6ea6cac0e4c2f8d301f6:parameters","entityId":"epoch:6ea6cac0e4c2f8d301f6","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-07-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:e06547b1f4a6268525e1:parameters","entityId":"epoch:e06547b1f4a6268525e1","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-07-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:845698026b414428bec9:parameters","entityId":"epoch:845698026b414428bec9","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:845698026b414428bec9:training-compute","entityId":"epoch:845698026b414428bec9","metric":"training-compute","value":2.69000000000001e+24,"unit":"FLOP","observationPeriod":"2025-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Reported in Table 2.\n\nfrom communication with the authors:\n\"EXAONE 4.0 32B: NVIDIA H200 GPUs x 512 EA for 19 weeks (FP8 mode training)\""},{"id":"epoch:7dc656257a4e67e18f08:parameters","entityId":"epoch:7dc656257a4e67e18f08","metric":"parameters","value":1200000000.0,"unit":"parameters","observationPeriod":"2025-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.2B"},{"id":"epoch:7dc656257a4e67e18f08:training-compute","entityId":"epoch:7dc656257a4e67e18f08","metric":"training-compute","value":8.65e+22,"unit":"FLOP","observationPeriod":"2025-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Reported in Table 2.\n\nfrom communication with the authors:\n\"EXAONE 4.0 1.2B: NVIDIA H200 GPUs x 512 EA for 2 weeks (FP8 mode training)\""},{"id":"epoch:20d69cce4a48c14d7db9:parameters","entityId":"epoch:20d69cce4a48c14d7db9","metric":"parameters","value":24300000000.0,"unit":"parameters","observationPeriod":"2025-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:774d1a4ce51b256a28b6:parameters","entityId":"epoch:774d1a4ce51b256a28b6","metric":"parameters","value":4700000000.0,"unit":"parameters","observationPeriod":"2025-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:9d6edd63babf878faa6f:parameters","entityId":"epoch:9d6edd63babf878faa6f","metric":"parameters","value":1000000000000.0,"unit":"parameters","observationPeriod":"2025-07-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"MoE with 1T total parameters and 32B parameters active per forward pass"},{"id":"epoch:9d6edd63babf878faa6f:training-compute","entityId":"epoch:9d6edd63babf878faa6f","metric":"training-compute","value":2.976e+24,"unit":"FLOP","observationPeriod":"2025-07-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 32 * 10^9 activated parameters * 15.5 * 10^12 tokens = 2.976e+24 FLOP"},{"id":"epoch:d5f1e9c16d3f096b439b:parameters","entityId":"epoch:d5f1e9c16d3f096b439b","metric":"parameters","value":3000000000000.0,"unit":"parameters","observationPeriod":"2025-07-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Rumored to be 2.4T params (https://x.com/kalomaze/status/1942996555088134592)"},{"id":"epoch:d5f1e9c16d3f096b439b:training-compute","entityId":"epoch:d5f1e9c16d3f096b439b","metric":"training-compute","value":5.0000000000001e+26,"unit":"FLOP","observationPeriod":"2025-07-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"We think that RL relative to pre-compute is between our estimate for o3 (10% of pre-training) and the 100% implied by this slide in the launch ( https://archive.is/f0vJU ). Assuming the same pre-training as Grok 3 (also implied by that slide, and much more consistent) and that Grok 3 used a tenth as much RL, we get:\n\n2 * (grok3/1.1) in the high case (rl is 10% of grok 3, so grok3/1.1 is grok3 precompute, and in this case twice that is grok 4)\n1.1 * (grok3/1.01) in the low case\nThe geometric mean is (rounded to one sig fig): 5e26\n"},{"id":"epoch:8486a659eb702b2565ef:parameters","entityId":"epoch:8486a659eb702b2565ef","metric":"parameters","value":175000000.0,"unit":"parameters","observationPeriod":"2025-07-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"~175M parameters based on Figure 1"},{"id":"epoch:efbf341f81c6d81227a1:parameters","entityId":"epoch:efbf341f81c6d81227a1","metric":"parameters","value":800000000.0,"unit":"parameters","observationPeriod":"2025-07-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"MedSigLIP contains a 400M parameter vision encoder and 400M parameter text encoder,\""},{"id":"epoch:af8ac44aba8ad40b119a:parameters","entityId":"epoch:af8ac44aba8ad40b119a","metric":"parameters","value":142000000000.0,"unit":"parameters","observationPeriod":"2025-07-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"a large-scale MoE model that activates 14\nbillion parameters out of a total of 142 billion parameters"},{"id":"epoch:af8ac44aba8ad40b119a:training-compute","entityId":"epoch:af8ac44aba8ad40b119a","metric":"training-compute","value":1.2164856e+24,"unit":"FLOP","observationPeriod":"2025-07-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 14000000000 active parameters * 11328000000000 tokens = 9.51552e+23 FLOP\n\n989000000000000 FLOP/GPU/sec [H800 assumed] * 1456000 GPU-hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.55518272e+24 FLOP\n\nsqrt(9.51552e+23*1.55518272e+24) = 1.2164856e+24"},{"id":"epoch:9a1b15246fe0e67a171a:parameters","entityId":"epoch:9a1b15246fe0e67a171a","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:df4abc174600019f86c5:parameters","entityId":"epoch:df4abc174600019f86c5","metric":"parameters","value":28000000000.0,"unit":"parameters","observationPeriod":"2025-06-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"MoE\ntotal parameters: 28B\nactive parameters: 3B"},{"id":"epoch:a947cd51f1dc448192c9:parameters","entityId":"epoch:a947cd51f1dc448192c9","metric":"parameters","value":300000000000.0,"unit":"parameters","observationPeriod":"2025-06-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"MoE\ntotal parameters: 300B\nactive parameters: 47B"},{"id":"epoch:a947cd51f1dc448192c9:training-compute","entityId":"epoch:a947cd51f1dc448192c9","metric":"training-compute","value":2.82e+24,"unit":"FLOP","observationPeriod":"2025-06-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"They say the model was trained on \"trillions of tokens\" \nSpeculatively assuming 10T tokens:\n\n6 FLOP / token / parameter * 47 * 10^9 active parameters * 10 * 10^12 assumed training tokens = 2.82e+24 FLOP"},{"id":"epoch:5c00c2f166e2ed93bdea:parameters","entityId":"epoch:5c00c2f166e2ed93bdea","metric":"parameters","value":21000000000.0,"unit":"parameters","observationPeriod":"2025-06-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"MoE\ntotal parameters: 21B\nactive parameters: 3B"},{"id":"epoch:5c00c2f166e2ed93bdea:training-compute","entityId":"epoch:5c00c2f166e2ed93bdea","metric":"training-compute","value":1.8e+23,"unit":"FLOP","observationPeriod":"2025-06-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"They say the model was trained on \"\"trillions of tokens\"\nSpeculatively assuming 10T tokens:\n\n6 FLOP / token / parameter * 3 * 10^9 active parameters * 10 * 10^12 assumed training tokens = 1.8e+23 FLOP"},{"id":"epoch:681ad7e718f8c0ff8449:parameters","entityId":"epoch:681ad7e718f8c0ff8449","metric":"parameters","value":360000000.0,"unit":"parameters","observationPeriod":"2025-06-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"0,36B dense"},{"id":"epoch:9d1b4f2c8932c78bca0a:parameters","entityId":"epoch:9d1b4f2c8932c78bca0a","metric":"parameters","value":12000000000.0,"unit":"parameters","observationPeriod":"2025-06-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12B"},{"id":"epoch:99faa010e98cdd431419:parameters","entityId":"epoch:99faa010e98cdd431419","metric":"parameters","value":107000000000.0,"unit":"parameters","observationPeriod":"2025-06-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:72713e68a8cf72338b86:parameters","entityId":"epoch:72713e68a8cf72338b86","metric":"parameters","value":450000000.0,"unit":"parameters","observationPeriod":"2025-06-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\" AlphaGenome has approximately 450 million trainable parameters (20% in the encoder, 28% in the sequence transformer, 15% in the pairwise blocks, 25% in the decoder, and 12% in the output embedding and prediction heads)\""},{"id":"epoch:72713e68a8cf72338b86:training-compute","entityId":"epoch:72713e68a8cf72338b86","metric":"training-compute","value":1.362969e+22,"unit":"FLOP","observationPeriod":"2025-06-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Pre-training:\n\"Each gradient step processed a batch size of 64 samples using\n8-way sequence parallelism, requiring 512 TPUv3 cores, with pre-training runs typically completing in approximately 4 hours.\"\n\n\"Distillation using many teacher models (e.g., 64; orange crosses)\"\n\n123000000000000 FLOP / TPUv3 chip / sec * (512 TPUv3 cores / 2) * 4 hours * 3600 sec / hour * 0.3 [assumed utilization] * 64 training runs [\"Likely\" confidence] = 8.7058022e+21 FLOP\n\nDistillation:\n\"Distillation training was performed without sequence parallelism across 64 NVIDIA H100 GPUs, with a batch size of 64 (effectively one sample per GPU). Each GPU loaded a different frozen teacher model from the pool of 64 pretrained all-folds models. <..> The distillation process ran for 250,000 steps, taking approximately 3 days\"\n\n989400000000000 FLOP / GPU / sec * 64 GPUs * 3 days * 24 hours / day * 3600 sec / hour * 0.3 [assumed utilization] = 4.9238876e+21 FLOP\n\n8.7058022e+21 FLOP + 4.9238876e+21 FLOP = 1.362969e+22 FLOP \n\n\"Likely\" confidence,  because I am not sure how many pre-training runs were there. "},{"id":"epoch:aa4b4aa79eb4fa2b30f5:parameters","entityId":"epoch:aa4b4aa79eb4fa2b30f5","metric":"parameters","value":16000000000.0,"unit":"parameters","observationPeriod":"2025-06-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" Structurally, Kimi-VL\nconsists of our Moonlight (J. Liu et al. 2025a) MoE language model with only 2.8B activated (16B total) parameters,\npaired with a 400M native-resolution MoonViT vision encoder."},{"id":"epoch:007702d9d1a7fa3bdef6:parameters","entityId":"epoch:007702d9d1a7fa3bdef6","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2025-06-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"72B"},{"id":"epoch:1aa95aa1b5e3f30532da:parameters","entityId":"epoch:1aa95aa1b5e3f30532da","metric":"parameters","value":24000000000.0,"unit":"parameters","observationPeriod":"2025-06-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"24B"},{"id":"epoch:bec13afd78de9d51ab87:parameters","entityId":"epoch:bec13afd78de9d51ab87","metric":"parameters","value":456000000000.0,"unit":"parameters","observationPeriod":"2025-06-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from the base model:\nTotal Parameters: 456B\nActivated Parameters per Token: 45.9B"},{"id":"epoch:bec13afd78de9d51ab87:training-compute","entityId":"epoch:bec13afd78de9d51ab87","metric":"training-compute","value":4.3240062e+24,"unit":"FLOP","observationPeriod":"2025-06-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.9828799999999997e+24 FLOP [base model compute] + 2.3411262e+24 FLOP [see finetune compute notes] = 4.3240062e+24 FLOP"},{"id":"epoch:961af46f4d6e6146681f:parameters","entityId":"epoch:961af46f4d6e6146681f","metric":"parameters","value":456000000000.0,"unit":"parameters","observationPeriod":"2025-06-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from the base model:\nTotal Parameters: 456B\nActivated Parameters per Token: 45.9B\n\n(matches safetensors)"},{"id":"epoch:961af46f4d6e6146681f:training-compute","entityId":"epoch:961af46f4d6e6146681f","metric":"training-compute","value":4.1861931e+24,"unit":"FLOP","observationPeriod":"2025-06-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.9828799999999997e+24 FLOP [base model compute] + 2.2033131e+24 FLOP [see finetune compute notes] = 4.1861931e+24 FLOP"},{"id":"epoch:1052689ece242983c0ac:parameters","entityId":"epoch:1052689ece242983c0ac","metric":"parameters","value":720000000.0,"unit":"parameters","observationPeriod":"2025-06-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In this work we train \ud835\udc3d = 4 models. To ensemble their predictions, we generate an equal number of ensemble member trajectories from each model\"\n\"FGN is a larger model, with \u223c180m parameters per model seed\"\n\nThey train a 4 model ensemble with 180M parameters per model. Total parameters: 4*180=720M"},{"id":"epoch:1052689ece242983c0ac:training-compute","entityId":"epoch:1052689ece242983c0ac","metric":"training-compute","value":9.618950880000001e+21,"unit":"FLOP","observationPeriod":"2025-06-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"TPU v5 FLOPs: 459000000000000\nTPU v6 FLOPs: 918000000000000\n\nThey train for 490 TPU days, with an unspecified mix of v5 and v6. Assuming same proportion of both TPU and a utilization of 0.33. \n\nMean TPU FLOPs: 688500000000000\n\nCompute: 490*24*60*60*688500000000000*0.33=9618950880000001000000"},{"id":"epoch:eb1fd3dee2d72a4d7394:parameters","entityId":"epoch:eb1fd3dee2d72a4d7394","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-06-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"same as the base model"},{"id":"epoch:c3c4e1eb3787069e2138:parameters","entityId":"epoch:c3c4e1eb3787069e2138","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2025-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 4"},{"id":"epoch:c3c4e1eb3787069e2138:training-compute","entityId":"epoch:c3c4e1eb3787069e2138","metric":"training-compute","value":9.05969664e+21,"unit":"FLOP","observationPeriod":"2025-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The number of tokens per video seems to be ~2048: 16x16 patches on 256x256 videos = 256 patches per frame, 2 tubelets across 16 frames = 8 temporal dimensions (not sure about this). Total tokens per clip: 8*16*16=2048\n\nBatch size = 3072\nIterations = 240k (\"Our training process begins with a warmup phase where we train on 16-frame, 256 \u00d7 256-resolution videos with linear learning rate warmup over 12K iterations, followed by a main training phase with a constant learning rate for 228K iterations.\")\n\nTotal pretraining compute using 6ND: 6*1e9 parameters *240000 steps * 3072 batch * 2048 tokens = 9059696640000000000000"},{"id":"epoch:009d61cd2b11fad55433:parameters","entityId":"epoch:009d61cd2b11fad55433","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2025-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:de3a49072721b71b31e8:parameters","entityId":"epoch:de3a49072721b71b31e8","metric":"parameters","value":2000000000.0,"unit":"parameters","observationPeriod":"2025-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:bafd6b6a3cc00e109e7f:parameters","entityId":"epoch:bafd6b6a3cc00e109e7f","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2025-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:41141bda368d8b4505ba:parameters","entityId":"epoch:41141bda368d8b4505ba","metric":"parameters","value":2000000000.0,"unit":"parameters","observationPeriod":"2025-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d3345bb2e14256fdd645:parameters","entityId":"epoch:d3345bb2e14256fdd645","metric":"parameters","value":230000000000.0,"unit":"parameters","observationPeriod":"2025-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Model has 230B parameters total, with 23B active parameters (sparse MoE setting)."},{"id":"epoch:c688ce49b96ca1468705:parameters","entityId":"epoch:c688ce49b96ca1468705","metric":"parameters","value":230000000000.0,"unit":"parameters","observationPeriod":"2025-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"MoE"},{"id":"epoch:b3e8d7aa52dcb525bcf3:parameters","entityId":"epoch:b3e8d7aa52dcb525bcf3","metric":"parameters","value":38000000000.0,"unit":"parameters","observationPeriod":"2025-06-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"38B"},{"id":"epoch:b4d3f23b41eb27b19694:parameters","entityId":"epoch:b4d3f23b41eb27b19694","metric":"parameters","value":24000000000.0,"unit":"parameters","observationPeriod":"2025-06-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"24B\n\n\"We finetuned Mistral Small 3 Instruct (a 24-billion parameter model) for 4 epochs\""},{"id":"epoch:5c7c24d6da0c05d24532:parameters","entityId":"epoch:5c7c24d6da0c05d24532","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-06-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8b"},{"id":"epoch:fba773cedcd7e4c17e8e:parameters","entityId":"epoch:fba773cedcd7e4c17e8e","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-06-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We set the number of Transformer layers to 36 and the hidden\ndimension to 4,096. The intermediate hidden dimension of FFN is set to 11,008. The number of attention heads is 32 and there are 8 key-value groups\""},{"id":"epoch:fba773cedcd7e4c17e8e:training-compute","entityId":"epoch:fba773cedcd7e4c17e8e","metric":"training-compute","value":1.05e+24,"unit":"FLOP","observationPeriod":"2025-06-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 7 * 10^9 parameters * 25 * 10^12 tokens = 1.05e+24 FLOP"},{"id":"epoch:ed3c0f8924637842e582:parameters","entityId":"epoch:ed3c0f8924637842e582","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-06-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:9006bc210701949a5c96:parameters","entityId":"epoch:9006bc210701949a5c96","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-06-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:852d55f589acb2dedae0:parameters","entityId":"epoch:852d55f589acb2dedae0","metric":"parameters","value":24000000000.0,"unit":"parameters","observationPeriod":"2025-06-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"24B"},{"id":"epoch:f594100b20c8d4478f26:parameters","entityId":"epoch:f594100b20c8d4478f26","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-06-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f594100b20c8d4478f26:training-compute","entityId":"epoch:f594100b20c8d4478f26","metric":"training-compute","value":1.1508e+24,"unit":"FLOP","observationPeriod":"2025-06-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.05e+24 FLOP [base model compute] + 1.008e+23 FLOP [finetune compute] = 1.1508e+24 FLOP"},{"id":"epoch:c0afa5429dbebcd3e6d0:parameters","entityId":"epoch:c0afa5429dbebcd3e6d0","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-06-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1b4aa336e057f1125e08:parameters","entityId":"epoch:1b4aa336e057f1125e08","metric":"parameters","value":8200000000.0,"unit":"parameters","observationPeriod":"2025-06-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1b4aa336e057f1125e08:training-compute","entityId":"epoch:1b4aa336e057f1125e08","metric":"training-compute","value":2.952e+23,"unit":"FLOP","observationPeriod":"2025-06-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 * 8.2e9 parameters * 6e12 tokens = 295200000000000000000000 FLOP. "},{"id":"epoch:3c4f72476838d9211fa6:parameters","entityId":"epoch:3c4f72476838d9211fa6","metric":"parameters","value":4000000000.0,"unit":"parameters","observationPeriod":"2025-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"the 4-billion parameter S1 model\""},{"id":"epoch:3c4f72476838d9211fa6:training-compute","entityId":"epoch:3c4f72476838d9211fa6","metric":"training-compute","value":3.36e+22,"unit":"FLOP","observationPeriod":"2025-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 4 * 10^9  parameters * 14 * 10^11 token [see training dataset size note] = 3.36e+22 FLOP"},{"id":"epoch:1fa0afe084467171672a:parameters","entityId":"epoch:1fa0afe084467171672a","metric":"parameters","value":500000000.0,"unit":"parameters","observationPeriod":"2025-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"S1-mini (0.5B): A distilled version of S1."},{"id":"epoch:1906d8cfa1a4554eb3bd:parameters","entityId":"epoch:1906d8cfa1a4554eb3bd","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-05-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:ff8df82825c27d83321a:parameters","entityId":"epoch:ff8df82825c27d83321a","metric":"parameters","value":71990000000.0,"unit":"parameters","observationPeriod":"2025-05-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"a total of 71.99B parameters and 16.50 billion active parameters"},{"id":"epoch:ff8df82825c27d83321a:training-compute","entityId":"epoch:ff8df82825c27d83321a","metric":"training-compute","value":1.287e+24,"unit":"FLOP","observationPeriod":"2025-05-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 16.5 * 10^9 active parameters * 13 * 10^12 tokens = 1.287e+24 FLOP"},{"id":"epoch:728b5b604bddad11da1e:parameters","entityId":"epoch:728b5b604bddad11da1e","metric":"parameters","value":671000000000.0,"unit":"parameters","observationPeriod":"2025-05-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"671B total\n37B activated\nhttps://github.com/deepseek-ai/DeepSeek-R1/tree/main"},{"id":"epoch:728b5b604bddad11da1e:training-compute","entityId":"epoch:728b5b604bddad11da1e","metric":"training-compute","value":4.020010000000001e+24,"unit":"FLOP","observationPeriod":"2025-05-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Estimates by Ege Erdil in Gradient Updates:\nhttps://epoch.ai/gradient-updates/what-went-into-training-deepseek-r1\n\"A dataset size of 14.8 trillion tokens is reasonable and in line with other models of this scale. Assuming that\u2019s valid, the pretraining of this model would have required 6 * (37 billion) * (14.8 trillion) = 3e24 FLOP. If we assume DeepSeek\u2019s training cluster consists of H800s with the PCIe form factor, then each should be capable of 1.5e15 FP8 per second, and the implied model FLOP utilization (MFU) of DeepSeek v3\u2019s 55 day training run ends up being around 23%.\"\n\n6 FLOP/token/param * 14.8T tokens * 37B active params = 3.29e24 FLOP (pretraining)\n1.2e23 FLOP (post-training)\n6.1e23 FLOP (fine-tuning)\n\nTotal compute: 3.29e24 + 1.2e23 + 6.1e23 = 4.02e24"},{"id":"epoch:681e558490fde1b7e00e:parameters","entityId":"epoch:681e558490fde1b7e00e","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-05-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B\n\"We design the architecture following LLaVA series [17, 18], where the omnimodal large language model consists of four key components: an LLM (Qwen2.5-7B-Instruct [2]) for next token prediction, an image encoder (CLIP-ViT-L [74]) for extracting visual features, a speech encoder (Whisper-large-v3 [75]) for extracting audio features and a streaming speech decoder (Qwen2.5-0.5B-Instruct [2]) for generating vivid speech in real-time.\""},{"id":"epoch:9a1efdadba4173fb8669:parameters","entityId":"epoch:9a1efdadba4173fb8669","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2025-05-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"A baseline 1B model is trained on the unfiltered dataset, while a second 1B model is trained analogously on the same dataset, but filtered by a DataRater.\""},{"id":"epoch:9a1efdadba4173fb8669:training-compute","entityId":"epoch:9a1efdadba4173fb8669","metric":"training-compute","value":1.2e+20,"unit":"FLOP","observationPeriod":"2025-05-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"c = 6nd where n = 1B, d = 20B\nc = 6 * 1B * 20B = 1.2e20 FLOPs\n\n\"All inner language models and DataRater models are based on the Chinchilla [Hoffmann et al., 2022] transformer architectures [Vaswani et al., 2017].\"\n\"Meta-training a DataRater requires approximately 58.4% of the FLOPS needed to train a single 1B LLM.\""},{"id":"epoch:6f1cf7069da696cba849:parameters","entityId":"epoch:6f1cf7069da696cba849","metric":"parameters","value":150000000.0,"unit":"parameters","observationPeriod":"2025-05-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"150M"},{"id":"epoch:6f1cf7069da696cba849:training-compute","entityId":"epoch:6f1cf7069da696cba849","metric":"training-compute","value":3.6832407e+21,"unit":"FLOP","observationPeriod":"2025-05-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"fine-tune compute is negligible"},{"id":"epoch:ec5a0dddc72d99ef3760:parameters","entityId":"epoch:ec5a0dddc72d99ef3760","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-05-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:38075516310fff3e4d98:parameters","entityId":"epoch:38075516310fff3e4d98","metric":"parameters","value":34000000000.0,"unit":"parameters","observationPeriod":"2025-05-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In this release, we feature six open-weight models: 0.5B, 1.5B, 1.5B-Deep, 3B, 7B, and 34B, along with their instruct versions\" (https://huggingface.co/blog/tiiuae/falcon-h1)."},{"id":"epoch:38075516310fff3e4d98:training-compute","entityId":"epoch:38075516310fff3e4d98","metric":"training-compute","value":3.672e+24,"unit":"FLOP","observationPeriod":"2025-05-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 34 * 10^9 parameters * 18 * 10^12 tokens = 3.672e+24 FLOP"},{"id":"epoch:e24d551e49ca89fbfdc1:parameters","entityId":"epoch:e24d551e49ca89fbfdc1","metric":"parameters","value":24000000000.0,"unit":"parameters","observationPeriod":"2025-05-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"24B"},{"id":"epoch:95bda1a3b67b1d843bfd:parameters","entityId":"epoch:95bda1a3b67b1d843bfd","metric":"parameters","value":27000000000.0,"unit":"parameters","observationPeriod":"2025-05-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"27B"},{"id":"epoch:979bc9e691dcd5a6371c:parameters","entityId":"epoch:979bc9e691dcd5a6371c","metric":"parameters","value":7850000000.0,"unit":"parameters","observationPeriod":"2025-05-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.85B (Safetensors)"},{"id":"epoch:979bc9e691dcd5a6371c:training-compute","entityId":"epoch:979bc9e691dcd5a6371c","metric":"training-compute","value":5.181e+23,"unit":"FLOP","observationPeriod":"2025-05-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 7.85 * 10^9 parameters * 11 * 10^12 tokens = 5.181e+23 FLOP"},{"id":"epoch:1fd6286517d3f71c29c6:parameters","entityId":"epoch:1fd6286517d3f71c29c6","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-05-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B\n\nArchitecture Details\nArchitecture: Llama 3 8B\nHidden size: 4096\nFeedforward size: 14336\nNumber of layers: 32\nNumber of attention heads: 32\nNumber of KV heads: 8"},{"id":"epoch:1fd6286517d3f71c29c6:training-compute","entityId":"epoch:1fd6286517d3f71c29c6","metric":"training-compute","value":6.12e+23,"unit":"FLOP","observationPeriod":"2025-05-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 8 *10^9 parameters * 12.75 * 10^12 tokens = 6.12e+23 FLOP"},{"id":"epoch:17dcb099743ba1b576b5:parameters","entityId":"epoch:17dcb099743ba1b576b5","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-05-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:256efebcad774541deae:parameters","entityId":"epoch:256efebcad774541deae","metric":"parameters","value":56000000000.0,"unit":"parameters","observationPeriod":"2025-05-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"56B"},{"id":"epoch:c4367aedb5d1a565545f:parameters","entityId":"epoch:c4367aedb5d1a565545f","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2025-05-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3B"},{"id":"epoch:02dd85d1a14073fff95c:parameters","entityId":"epoch:02dd85d1a14073fff95c","metric":"parameters","value":4800000000.0,"unit":"parameters","observationPeriod":"2025-05-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4.8B\n\nOur final model (SANA-4.8B)\nscales to 60 layers while maintaining the same channel\ndimension (2240 per layer) and FFN dimension (5600)\nas SANA-1.6B. The architecture, training data, and other\nhyperparameters remain consistent with SANA-1.6B"},{"id":"epoch:9789c855af7a50a00cd6:parameters","entityId":"epoch:9789c855af7a50a00cd6","metric":"parameters","value":133000000.0,"unit":"parameters","observationPeriod":"2025-05-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"133M"},{"id":"epoch:20298f0dd6ddea59a95d:parameters","entityId":"epoch:20298f0dd6ddea59a95d","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-05-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:41632e83bcad2e2d9534:parameters","entityId":"epoch:41632e83bcad2e2d9534","metric":"parameters","value":32800000000.0,"unit":"parameters","observationPeriod":"2025-05-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1559549554cee1bb11a9:training-compute","entityId":"epoch:1559549554cee1bb11a9","metric":"training-compute","value":1.388556e+24,"unit":"FLOP","observationPeriod":"2025-05-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"989000000000000 FLOP / GPU / sec [H800, bf16 assumed] * 1300000 GPU-hours [see training time notes] * 3600 sec / hour * 0.3 [assumed utilization] = 1.388556e+24 FLOP"},{"id":"epoch:277f2a0ccf74d1cfa392:parameters","entityId":"epoch:277f2a0ccf74d1cfa392","metric":"parameters","value":330000000.0,"unit":"parameters","observationPeriod":"2025-05-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"cBottle-SR has 330M parameters."},{"id":"epoch:277f2a0ccf74d1cfa392:training-compute","entityId":"epoch:277f2a0ccf74d1cfa392","metric":"training-compute","value":1.6014864e+21,"unit":"FLOP","observationPeriod":"2025-05-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"989400000000000 FLOP / GPU / sec * 5395480 GPU-seconds [see training time notes] * 0.3 [assumed utilization] = 1.6014864e+21 FLOP"},{"id":"epoch:93444fcfb2befe7bba10:parameters","entityId":"epoch:93444fcfb2befe7bba10","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-05-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:8cb0d277ccceb2d13728:parameters","entityId":"epoch:8cb0d277ccceb2d13728","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2025-05-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:86da392f6a33e3a96b38:parameters","entityId":"epoch:86da392f6a33e3a96b38","metric":"parameters","value":718000000000.0,"unit":"parameters","observationPeriod":"2025-05-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"a sparse LLM with 718 billion parameters\n\n39B activated parameters"},{"id":"epoch:86da392f6a33e3a96b38:training-compute","entityId":"epoch:86da392f6a33e3a96b38","metric":"training-compute","value":3.0888e+24,"unit":"FLOP","observationPeriod":"2025-05-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Speculatively assuming they trained MoE model on the same amount of tokens as Pangu Ultra (13,2 T): \n\n6 FLOP / parameter / token * 39 * 10^9 activated parameters * 13,2 * 10^12 tokens = 3.0888e+24 FLOP"},{"id":"epoch:10319c16702a051199cd:parameters","entityId":"epoch:10319c16702a051199cd","metric":"parameters","value":15000000000.0,"unit":"parameters","observationPeriod":"2025-05-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"15B"},{"id":"epoch:10319c16702a051199cd:training-compute","entityId":"epoch:10319c16702a051199cd","metric":"training-compute","value":9.00001e+21,"unit":"FLOP","observationPeriod":"2025-05-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 15 * 10^9 parameters * 100 * 10^9 tokens = 9e+21 FLOP\n\n[1 epoch assumed] -> \"Likely\" confidence"},{"id":"epoch:85cb23d1efb1115ae339:parameters","entityId":"epoch:85cb23d1efb1115ae339","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-05-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:2b91950248b498239ea9:parameters","entityId":"epoch:2b91950248b498239ea9","metric":"parameters","value":4000000000.0,"unit":"parameters","observationPeriod":"2025-05-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4B\nModel type: A 4B instruct decoder-only model based on Gemma3 architecture."},{"id":"epoch:d03adb44eb834030afcc:parameters","entityId":"epoch:d03adb44eb834030afcc","metric":"parameters","value":12000000000.0,"unit":"parameters","observationPeriod":"2025-05-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12B"},{"id":"epoch:824d1b3fdc72a4a73a68:parameters","entityId":"epoch:824d1b3fdc72a4a73a68","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2025-05-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:0cd069e5dd243aea2ffd:parameters","entityId":"epoch:0cd069e5dd243aea2ffd","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2025-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B"},{"id":"epoch:0cd069e5dd243aea2ffd:training-compute","entityId":"epoch:0cd069e5dd243aea2ffd","metric":"training-compute","value":9.3368077e+23,"unit":"FLOP","observationPeriod":"2025-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9.3202015e+23 FLOP [base model compute] + 1.6606191e+21 FLOP [finetune compute] = 9.3368077e+23 FLOP"},{"id":"epoch:b8caf7b9643ad7ce28f6:parameters","entityId":"epoch:b8caf7b9643ad7ce28f6","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2025-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B"},{"id":"epoch:319f4b6d9f7c355da3e1:parameters","entityId":"epoch:319f4b6d9f7c355da3e1","metric":"parameters","value":671000000000.0,"unit":"parameters","observationPeriod":"2025-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"671B"},{"id":"epoch:5d638eaf00697a6a3794:parameters","entityId":"epoch:5d638eaf00697a6a3794","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:2460017028c8b422baa7:parameters","entityId":"epoch:2460017028c8b422baa7","metric":"parameters","value":149000000.0,"unit":"parameters","observationPeriod":"2025-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"149M"},{"id":"epoch:2460017028c8b422baa7:training-compute","entityId":"epoch:2460017028c8b422baa7","metric":"training-compute","value":3.6832407e+21,"unit":"FLOP","observationPeriod":"2025-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.676128e+21 FLOP [base model compute] + 7.112664e+18 FLOP [ fine-tune compute] = 3.6832407e+21 FLOP"},{"id":"epoch:75f5a3b7b72e77603e1d:parameters","entityId":"epoch:75f5a3b7b72e77603e1d","metric":"parameters","value":235000000000.0,"unit":"parameters","observationPeriod":"2025-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"235 billion total parameters and 22 billion activated parameters\n\nNumber of Layers: 94\nNumber of Attention Heads (GQA): 64 for Q and 4 for KV\nNumber of Experts: 128\nNumber of Activated Experts: 8\nContext Length: 32,768 natively and 131,072 tokens with YaRN."},{"id":"epoch:75f5a3b7b72e77603e1d:training-compute","entityId":"epoch:75f5a3b7b72e77603e1d","metric":"training-compute","value":4.752e+24,"unit":"FLOP","observationPeriod":"2025-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 22*10^9 active parameters * 36000000000000 tokens = 4.752e+24 FLOP"},{"id":"epoch:fba54ae106dfd5a70ce0:parameters","entityId":"epoch:fba54ae106dfd5a70ce0","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2025-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"30 billion total parameters and 3 billion activated parameters\n\nNumber of Layers: 48\nNumber of Attention Heads (GQA): 32 for Q and 4 for KV\nNumber of Experts: 128\nNumber of Activated Experts: 8\nContext Length: 32,768"},{"id":"epoch:fba54ae106dfd5a70ce0:training-compute","entityId":"epoch:fba54ae106dfd5a70ce0","metric":"training-compute","value":6.48e+23,"unit":"FLOP","observationPeriod":"2025-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 3*10^9 active parameters * 36000000000000 tokens = 6.48e+23 FLOP"},{"id":"epoch:6e5a075da0f25059fa02:parameters","entityId":"epoch:6e5a075da0f25059fa02","metric":"parameters","value":32800000000.0,"unit":"parameters","observationPeriod":"2025-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Number of Parameters: 32.8B\nNumber of Paramaters (Non-Embedding): 31.2B\nNumber of Layers: 64\nNumber of Attention Heads (GQA): 64 for Q and 8 for KV\nContext Length: 32,768 natively and 131,072 tokens with YaRN."},{"id":"epoch:6e5a075da0f25059fa02:training-compute","entityId":"epoch:6e5a075da0f25059fa02","metric":"training-compute","value":7.0848e+24,"unit":"FLOP","observationPeriod":"2025-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 36 * 10^12 tokens * 32.8  * 10^9 parameters = 7.0848e+24 FLOP"},{"id":"epoch:da9f3ecd73e073c1c6f6:parameters","entityId":"epoch:da9f3ecd73e073c1c6f6","metric":"parameters","value":14800000000.0,"unit":"parameters","observationPeriod":"2025-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Number of Parameters: 14.8B\nNumber of Paramaters (Non-Embedding): 13.2B\nNumber of Layers: 40\nNumber of Attention Heads (GQA): 40 for Q and 8 for KV\nContext Length: 32,768 natively and 131,072 tokens with YaRN."},{"id":"epoch:da9f3ecd73e073c1c6f6:training-compute","entityId":"epoch:da9f3ecd73e073c1c6f6","metric":"training-compute","value":3.1968e+24,"unit":"FLOP","observationPeriod":"2025-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 36 * 10^12 tokens * 14.8  * 10^9 parameters = 3.1968e+24 FLOP"},{"id":"epoch:9b6b853ef9e9df019d76:parameters","entityId":"epoch:9b6b853ef9e9df019d76","metric":"parameters","value":8200000000.0,"unit":"parameters","observationPeriod":"2025-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Number of Parameters: 8.2B\nNumber of Paramaters (Non-Embedding): 6.95B\nNumber of Layers: 36\nNumber of Attention Heads (GQA): 32 for Q and 8 for KV\nContext Length: 32,768 natively and 131,072 tokens with YaRN."},{"id":"epoch:9b6b853ef9e9df019d76:training-compute","entityId":"epoch:9b6b853ef9e9df019d76","metric":"training-compute","value":1.7712e+24,"unit":"FLOP","observationPeriod":"2025-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 36 * 10^12 tokens * 8.2  * 10^9 parameters = 1.7712e+24 FLOP"},{"id":"epoch:54885e7ae9e3e59d211c:parameters","entityId":"epoch:54885e7ae9e3e59d211c","metric":"parameters","value":4000000000.0,"unit":"parameters","observationPeriod":"2025-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Number of Parameters: 4.0B\nNumber of Paramaters (Non-Embedding): 3.6B\nNumber of Layers: 36\nNumber of Attention Heads (GQA): 32 for Q and 8 for KV\nContext Length: 32,768 natively and 131,072 tokens with YaRN."},{"id":"epoch:54885e7ae9e3e59d211c:training-compute","entityId":"epoch:54885e7ae9e3e59d211c","metric":"training-compute","value":8.64e+23,"unit":"FLOP","observationPeriod":"2025-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 36 * 10^12 tokens * 4  * 10^9 parameters = 8.64e+23 FLOP"},{"id":"epoch:3ab49504665065ad260f:parameters","entityId":"epoch:3ab49504665065ad260f","metric":"parameters","value":1700000000.0,"unit":"parameters","observationPeriod":"2025-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Number of Parameters: 1.7B\nNumber of Paramaters (Non-Embedding): 1.4B\nNumber of Layers: 28\nNumber of Attention Heads (GQA): 16 for Q and 8 for KV\nContext Length: 32,768\n"},{"id":"epoch:3ab49504665065ad260f:training-compute","entityId":"epoch:3ab49504665065ad260f","metric":"training-compute","value":3.672e+23,"unit":"FLOP","observationPeriod":"2025-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 36 * 10^12 tokens * 1.7  * 10^9 parameters = 3.672e+23 FLOP"},{"id":"epoch:13549f3d8308f054c74e:parameters","entityId":"epoch:13549f3d8308f054c74e","metric":"parameters","value":600000000.0,"unit":"parameters","observationPeriod":"2025-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Number of Parameters: 0.6B\nNumber of Paramaters (Non-Embedding): 0.44B\nNumber of Layers: 28\nNumber of Attention Heads (GQA): 16 for Q and 8 for KV\nContext Length: 32,768"},{"id":"epoch:13549f3d8308f054c74e:training-compute","entityId":"epoch:13549f3d8308f054c74e","metric":"training-compute","value":1.296e+23,"unit":"FLOP","observationPeriod":"2025-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 36 * 10^12 tokens * 0.6 * 10^9 parameters = 1.296e+23 FLOP"},{"id":"epoch:c6bd651709a25b3c0904:parameters","entityId":"epoch:c6bd651709a25b3c0904","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-04-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our newly launched Foundation AI group\u2014formed to bring world-class, domain-specific AI infrastructure to the cybersecurity space\u2014is proud to introduce its first release: Llama-3.1-FoundationAI-SecurityLLM-base-8B (Foundation-sec-8b), an 8-billion parameter, open-weight Large Language Model (LLM) purpose-built for security.\""},{"id":"epoch:c6bd651709a25b3c0904:training-compute","entityId":"epoch:c6bd651709a25b3c0904","metric":"training-compute","value":1.2242448e+24,"unit":"FLOP","observationPeriod":"2025-04-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.224e+24 FLOP [base model compute] + 2.448e+20 FLOP [finetune compute] = 1.2242448e24 FLOP"},{"id":"epoch:c1219fedb6a19bbd2857:parameters","entityId":"epoch:c1219fedb6a19bbd2857","metric":"parameters","value":350000000.0,"unit":"parameters","observationPeriod":"2025-04-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"350M\n"},{"id":"epoch:c1219fedb6a19bbd2857:training-compute","entityId":"epoch:c1219fedb6a19bbd2857","metric":"training-compute","value":2.7186806e+21,"unit":"FLOP","observationPeriod":"2025-04-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"base model compute 2.6788982e+21 FLOP + finetune compute  3.9782379e+19 FLOP = 2.7186806e+21 FLOP"},{"id":"epoch:1c55bc13a81788040c58:parameters","entityId":"epoch:1c55bc13a81788040c58","metric":"parameters","value":1200000000.0,"unit":"parameters","observationPeriod":"2025-04-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.2B"},{"id":"epoch:1c55bc13a81788040c58:training-compute","entityId":"epoch:1c55bc13a81788040c58","metric":"training-compute","value":2.9907184e+22,"unit":"FLOP","observationPeriod":"2025-04-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"base model compute 2.9770787e+22 FLOP + finetune compute 1.3639673e+20 FLOP = 2.9907184e+22 FLOP"},{"id":"epoch:06d08f9382a05a3ff00c:parameters","entityId":"epoch:06d08f9382a05a3ff00c","metric":"parameters","value":18000000000.0,"unit":"parameters","observationPeriod":"2025-04-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"18B"},{"id":"epoch:cd389f76a6f60940e4f9:parameters","entityId":"epoch:cd389f76a6f60940e4f9","metric":"parameters","value":9000000000.0,"unit":"parameters","observationPeriod":"2025-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9B"},{"id":"epoch:cd389f76a6f60940e4f9:training-compute","entityId":"epoch:cd389f76a6f60940e4f9","metric":"training-compute","value":4.36e+23,"unit":"FLOP","observationPeriod":"2025-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Gemma 2 9B pretrain on 8T tokens + MamayLM continual pretrain on 75B tokens. Unclear about SFT compute"},{"id":"epoch:78467f8fcf3c244348ce:parameters","entityId":"epoch:78467f8fcf3c244348ce","metric":"parameters","value":3300000000.0,"unit":"parameters","observationPeriod":"2025-04-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Fig. 3: Model overview\npre-training:\n- PaliGemma (3B): SigLIP (400M) + Gemma (2.6B)\n- FAST tokenizer\npost-training:\n- pretrained VLM (3B)\n- action expert (300M)\ninference: same as post-training\n\nthis is in line with \u03c00's 3.3B parameters\n"},{"id":"epoch:5b837a512c5e32ecaa0d:parameters","entityId":"epoch:5b837a512c5e32ecaa0d","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-04-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 8. \"Full Model, 8B\"\nCould be Qwen2.5-7B + MLP Connector (40M)\n\nhttps://nvlabs.github.io/EAGLE/\nas of 2025-05-29: \"Eagle-2.5 Weights (Coming Soon)\" "},{"id":"epoch:1bd1eecb204fcaa6c109:parameters","entityId":"epoch:1bd1eecb204fcaa6c109","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2025-04-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B"},{"id":"epoch:eb1226825de424adbc37:parameters","entityId":"epoch:eb1226825de424adbc37","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2025-04-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70b"},{"id":"epoch:ac1bed91d2c8f9dffe1f:parameters","entityId":"epoch:ac1bed91d2c8f9dffe1f","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-04-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:ac1bed91d2c8f9dffe1f:training-compute","entityId":"epoch:ac1bed91d2c8f9dffe1f","metric":"training-compute","value":9.3e+22,"unit":"FLOP","observationPeriod":"2025-04-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"~9.3\u00d710\u00b2\u00b2 FLOPs (reported)\n\n6 FLOP/parameter/token * 7000000000 parameters * 2000000000000 tokens = 8.4e+22 FLOP\n\n989400000000000 FLOP/GPU/sec * 59400 GPU-hours * 3600 sec / hour * 0.4250 [reported utilization] = 8.9918651e+22 FLOP"},{"id":"epoch:5ee6e4f7c31ea1cce113:parameters","entityId":"epoch:5ee6e4f7c31ea1cce113","metric":"parameters","value":4000000000.0,"unit":"parameters","observationPeriod":"2025-04-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4B"},{"id":"epoch:0bf72f477c6f945e20fd:parameters","entityId":"epoch:0bf72f477c6f945e20fd","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2025-04-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1B"},{"id":"epoch:455cf217660d1aa4207a:parameters","entityId":"epoch:455cf217660d1aa4207a","metric":"parameters","value":12000000000.0,"unit":"parameters","observationPeriod":"2025-04-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12B"},{"id":"epoch:448fcfd8fa9ae2d677cf:parameters","entityId":"epoch:448fcfd8fa9ae2d677cf","metric":"parameters","value":27000000000.0,"unit":"parameters","observationPeriod":"2025-04-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"27B"},{"id":"epoch:0fd82c216348da635af4:parameters","entityId":"epoch:0fd82c216348da635af4","metric":"parameters","value":95000000.0,"unit":"parameters","observationPeriod":"2025-04-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"95M"},{"id":"epoch:0fd82c216348da635af4:training-compute","entityId":"epoch:0fd82c216348da635af4","metric":"training-compute","value":4.579186e+20,"unit":"FLOP","observationPeriod":"2025-04-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 95 * 10^6 parameters * 351 * 10^9 tokens [see dataset size notes] *1.8 epochs = 3.60126e+20 FLOP\n\n312000000000000 FLOP / GPU / sec [A100 reported] * 8 GPUs * 216 hours [9 days reported] * 3600 sec / hour * 0.3 [assumed utilization] = 5.8226688e+20 FLOP\n\nsqrt(3.60126e+20 * 5.8226688e+20) = 4.579186e+20 FLOP"},{"id":"epoch:283983eedf274ecdac11:parameters","entityId":"epoch:283983eedf274ecdac11","metric":"parameters","value":671000000000.0,"unit":"parameters","observationPeriod":"2025-04-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"671B\nMoE"},{"id":"epoch:2907e7680fe044744bb5:training-compute","entityId":"epoch:2907e7680fe044744bb5","metric":"training-compute","value":3.10542336e+21,"unit":"FLOP","observationPeriod":"2025-04-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000 FLOP/GPU/sec * 9216 GPU-hours * 3600 sec / hour * 0.3 [assumed utilization] = 3.10542336e+21 FLOP"},{"id":"epoch:80556a95ce695a246bd3:parameters","entityId":"epoch:80556a95ce695a246bd3","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Model Architecture\nArchitecture Type: Hybrid Mamba-Transformer\nNetwork Architecture: Nemotron-H\nThis model has 8B model parameters."},{"id":"epoch:80556a95ce695a246bd3:training-compute","entityId":"epoch:80556a95ce695a246bd3","metric":"training-compute","value":7.2e+23,"unit":"FLOP","observationPeriod":"2025-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 8 * 10^9 parameters * 15 * 10^12 tokens = 7.2e+23 FLOP"},{"id":"epoch:aed06a4e00ea3a359324:parameters","entityId":"epoch:aed06a4e00ea3a359324","metric":"parameters","value":47000000000.0,"unit":"parameters","observationPeriod":"2025-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use MiniPuzzle to distill Nemotron-H-56B-Base to Nemotron-H-47B-Base, using only 63 billion training tokens and FP8 training\""},{"id":"epoch:19980ca14e437dfe8634:parameters","entityId":"epoch:19980ca14e437dfe8634","metric":"parameters","value":56000000000.0,"unit":"parameters","observationPeriod":"2025-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Model Architecture\nArchitecture Type: Hybrid Mamba-Transformer\nNetwork Architecture: Nemotron-H\nThis model has 56B model parameters."},{"id":"epoch:19980ca14e437dfe8634:training-compute","entityId":"epoch:19980ca14e437dfe8634","metric":"training-compute","value":6.72e+24,"unit":"FLOP","observationPeriod":"2025-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 56 * 10^9 parameters * 20 * 10^12 tokens = 6.72e+24 FLOP"},{"id":"epoch:d047ba114026ff4e0102:parameters","entityId":"epoch:d047ba114026ff4e0102","metric":"parameters","value":400000000.0,"unit":"parameters","observationPeriod":"2025-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"400M"},{"id":"epoch:36dba52f1c1f1ea7bed7:parameters","entityId":"epoch:36dba52f1c1f1ea7bed7","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:36dba52f1c1f1ea7bed7:training-compute","entityId":"epoch:36dba52f1c1f1ea7bed7","metric":"training-compute","value":2.88e+24,"unit":"FLOP","observationPeriod":"2025-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 32 * 10^9 parameters * 15 * 10^12 tokens = 2.88e+24 FLOP"},{"id":"epoch:763aeaec15e5360dacf2:parameters","entityId":"epoch:763aeaec15e5360dacf2","metric":"parameters","value":9000000000.0,"unit":"parameters","observationPeriod":"2025-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9B"},{"id":"epoch:763aeaec15e5360dacf2:training-compute","entityId":"epoch:763aeaec15e5360dacf2","metric":"training-compute","value":8.1e+23,"unit":"FLOP","observationPeriod":"2025-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming it was trained on the same 15T dataset as 32B model:\n\n6 FLOP / parameter / token * 9 * 10^9 parameters * 15 * 10^12 tokens = 8.1e+23 FLOP \n\n\"Likely\" confidence due to the uncertain dataset size"},{"id":"epoch:d0949d91e59cadc80e89:parameters","entityId":"epoch:d0949d91e59cadc80e89","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b659715b7f5b4c2f6ebc:parameters","entityId":"epoch:b659715b7f5b4c2f6ebc","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:b659715b7f5b4c2f6ebc:training-compute","entityId":"epoch:b659715b7f5b4c2f6ebc","metric":"training-compute","value":2.88e+24,"unit":"FLOP","observationPeriod":"2025-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 32 * 10^9 parameters * 15 * 10^12 tokens = 2.88e+24 FLOP"},{"id":"epoch:f398f828e37354fc8626:parameters","entityId":"epoch:f398f828e37354fc8626","metric":"parameters","value":600000000000.0,"unit":"parameters","observationPeriod":"2025-04-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Mixture of Experts (MoE)-based multimodal general foundation model with over 600 billion parameters,\""},{"id":"epoch:21cfdb9e3947d0cf21c1:parameters","entityId":"epoch:21cfdb9e3947d0cf21c1","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-04-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:21cfdb9e3947d0cf21c1:training-compute","entityId":"epoch:21cfdb9e3947d0cf21c1","metric":"training-compute","value":9.0007697e+23,"unit":"FLOP","observationPeriod":"2025-04-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"989400000000000 FLOP / GPU / sec [H100 reported, bf16 assumed] * 665000 GPU-hours * 3600 sec / hour * 0.38 [reported utilization] = 9.0007697e+23 FLOP"},{"id":"epoch:2aa56cf25aea2c4bb003:parameters","entityId":"epoch:2aa56cf25aea2c4bb003","metric":"parameters","value":135000000000.0,"unit":"parameters","observationPeriod":"2025-04-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2aa56cf25aea2c4bb003:training-compute","entityId":"epoch:2aa56cf25aea2c4bb003","metric":"training-compute","value":1.0692e+25,"unit":"FLOP","observationPeriod":"2025-04-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"When compared to Llama 3.1 405B, Pangu Ultra achieves better scores on most of the challenging benchmarks, while utilizing only about 29% of the training FLOPs required by Llama 405B.\n\nCompute = 6 FLOP/token/param *  135e9 params *13.2e12 tokens = 1.069200e+25 FLOP\nThis is consistent with 29% of Llama 405B's compute: 3.8e25*0.29=1.1e25."},{"id":"epoch:1769c24fb34c72a42763:parameters","entityId":"epoch:1769c24fb34c72a42763","metric":"parameters","value":340000000000.0,"unit":"parameters","observationPeriod":"2025-04-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Supposedly same amount of parameters as Palm 2 Large"},{"id":"epoch:1769c24fb34c72a42763:training-compute","entityId":"epoch:1769c24fb34c72a42763","metric":"training-compute","value":7.34e+24,"unit":"FLOP","observationPeriod":"2025-04-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming negligible fine-tune compute\nbase model (Palm 2): 7.34e+24 FLOP\n"},{"id":"epoch:8ff0fb577fe719c90c2c:parameters","entityId":"epoch:8ff0fb577fe719c90c2c","metric":"parameters","value":27000000000.0,"unit":"parameters","observationPeriod":"2025-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"27B"},{"id":"epoch:8ff0fb577fe719c90c2c:training-compute","entityId":"epoch:8ff0fb577fe719c90c2c","metric":"training-compute","value":2.116854e+24,"unit":"FLOP","observationPeriod":"2025-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"base model compute 2.106e+24 FLOP + finetune compute 1.0854e+22 FLOP = 2.116854e+24 FLOP"},{"id":"epoch:2d2f4328e07b79f01457:parameters","entityId":"epoch:2d2f4328e07b79f01457","metric":"parameters","value":9000000000.0,"unit":"parameters","observationPeriod":"2025-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9B"},{"id":"epoch:2d2f4328e07b79f01457:training-compute","entityId":"epoch:2d2f4328e07b79f01457","metric":"training-compute","value":4.35618e+23,"unit":"FLOP","observationPeriod":"2025-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"base model compute 4.32e+23 FLOP + finetune compute 3.618e+21 FLOP = 4.35618e+23 FLOP"},{"id":"epoch:b679a53aed7d5c619a3f:parameters","entityId":"epoch:b679a53aed7d5c619a3f","metric":"parameters","value":2600000000.0,"unit":"parameters","observationPeriod":"2025-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.6B"},{"id":"epoch:b679a53aed7d5c619a3f:training-compute","entityId":"epoch:b679a53aed7d5c619a3f","metric":"training-compute","value":3.22452e+22,"unit":"FLOP","observationPeriod":"2025-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"base model compute 3.12e+22 FLOP + finetune compute 1.0452e+21 FLOP = 3.22452e+22 FLOP"},{"id":"epoch:ab11d85204bb1d611c39:parameters","entityId":"epoch:ab11d85204bb1d611c39","metric":"parameters","value":16700000000.0,"unit":"parameters","observationPeriod":"2025-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"16.7B"},{"id":"epoch:3472cefebea381936e3b:parameters","entityId":"epoch:3472cefebea381936e3b","metric":"parameters","value":109000000000.0,"unit":"parameters","observationPeriod":"2025-04-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our smaller model, Llama 4 Scout, is a general purpose model with 17 billion active parameters, 16 experts, and 109 billion total parameters that delivers state-of-the-art performance for its class.\""},{"id":"epoch:3472cefebea381936e3b:training-compute","entityId":"epoch:3472cefebea381936e3b","metric":"training-compute","value":4.08e+24,"unit":"FLOP","observationPeriod":"2025-04-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"40T training tokens per model card:\n\nhttps://github.com/meta-llama/llama-models/blob/main/models/llama4/MODEL_CARD.md  \n\nEstimating training compute from parameters and tokens:\n6 FLOP per token per parameter * 17B active parameters * 40T tokens = 4.08e24 FLOP\n(Implying mean throughput was 227 TFLOPS/GPU, or 11.5% MFU in FP8)\n\n\nThe model card also states that Llama 4 Scout used 5.0M H100-hours.\nThe blog post gives a figure of 390 TFLOPS/GPU, but this may have been the utilization rate for Behemoth, or all of the models together. Using this utilization, we have:\nCompute = 390 TFLOP/s * 5 million hours = 7.02e24 FLOP\n(This value is higher than the compute implied by parameters and tokens, and suggests utilization may have been lower for Scout than for Behemoth.)"},{"id":"epoch:66c02b081abde74e371c:parameters","entityId":"epoch:66c02b081abde74e371c","metric":"parameters","value":400000000000.0,"unit":"parameters","observationPeriod":"2025-04-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Llama 4 Maverick models have 17B active parameters and 400B total parameters.\"\n\nhttps://ai.meta.com/blog/llama-4-multimodal-intelligence/"},{"id":"epoch:66c02b081abde74e371c:training-compute","entityId":"epoch:66c02b081abde74e371c","metric":"training-compute","value":2.244000000001e+24,"unit":"FLOP","observationPeriod":"2025-04-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"22T training tokens per model card:\n\nhttps://github.com/meta-llama/llama-models/blob/main/models/llama4/MODEL_CARD.md   \n\nMaverick was trained using co-distillation from Llama 4 Behemoth. It isn't 100% clear that all 22T tokens used distillation, but we assume this for the time being.\n\nEstimating training compute from parameters and tokens:\nCompute = 6 FLOP per token per parameter * 17B active parameters * 22T tokens = 2.244e24 FLOP\n(Implying mean throughput was 262 TFLOPS/GPU, or 13.2% MFU in FP8)\n\n\nThe model card also states that Llama 4 Maverick used 2.38M H100-hours.\nThe blog post gives a figure of 390 TFLOPS/GPU, but this may have been the utilization rate for Behemoth, or all of the models together. Using this utilization, we have:\nCompute = 390 TFLOP/s * 2.38 million hours = 3.342e24 FLOP\n(This value is higher than the compute implied by parameters and tokens, and suggests utilization may have been lower for Maverick than for Behemoth.)"},{"id":"epoch:22f4f3edc6b38db24d62:parameters","entityId":"epoch:22f4f3edc6b38db24d62","metric":"parameters","value":2000000000000.0,"unit":"parameters","observationPeriod":"2025-04-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Llama 4 Behemoth, a 288 billion active parameter model with 16 experts that is our most powerful yet and among the world\u2019s smartest LLMs.\""},{"id":"epoch:22f4f3edc6b38db24d62:training-compute","entityId":"epoch:22f4f3edc6b38db24d62","metric":"training-compute","value":5.18400000000001e+25,"unit":"FLOP","observationPeriod":"2025-04-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Behemoth's training dataset is at least 30T tokens:\nhttps://ai.meta.com/blog/llama-4-multimodal-intelligence/ \n\n6 FLOP / parameter / token * 288 * 10^9 activated parameters * 30 * 10^12 tokens = 5.184e+25 FLOP"},{"id":"epoch:894558f1a1eb84fbac37:parameters","entityId":"epoch:894558f1a1eb84fbac37","metric":"parameters","value":335300000.0,"unit":"parameters","observationPeriod":"2025-04-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"335.3M"},{"id":"epoch:59d65a351c3abe7da509:parameters","entityId":"epoch:59d65a351c3abe7da509","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2025-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B"},{"id":"epoch:2410f19ba962dcb0323e:parameters","entityId":"epoch:2410f19ba962dcb0323e","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2025-04-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"72b"},{"id":"epoch:be0d696a7d4e169f175e:parameters","entityId":"epoch:be0d696a7d4e169f175e","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-04-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32b"},{"id":"epoch:0d33a7682201582f7b2c:parameters","entityId":"epoch:0d33a7682201582f7b2c","metric":"parameters","value":24000000000.0,"unit":"parameters","observationPeriod":"2025-03-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"24B"},{"id":"epoch:12985b291a8d4610ee8f:parameters","entityId":"epoch:12985b291a8d4610ee8f","metric":"parameters","value":2600000000.0,"unit":"parameters","observationPeriod":"2025-03-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.6B"},{"id":"epoch:12985b291a8d4610ee8f:training-compute","entityId":"epoch:12985b291a8d4610ee8f","metric":"training-compute","value":4.7794406e+21,"unit":"FLOP","observationPeriod":"2025-03-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000 FLOP / GPU / sec [A100 reported, bf16 assumed] * 14184 GPU-hours [see training time notes] * 3600 sec / hour * 0.3 [assumed utilization] = 4.7794406e+21 FLOP"},{"id":"epoch:2da27fd03744a3d73a8d:parameters","entityId":"epoch:2da27fd03744a3d73a8d","metric":"parameters","value":8685000000.0,"unit":"parameters","observationPeriod":"2025-03-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"video tokenizer: 285M parameters\nworkd model:  8.4B parameters\n\ntotal: 8.685B parameters"},{"id":"epoch:2da27fd03744a3d73a8d:training-compute","entityId":"epoch:2da27fd03744a3d73a8d","metric":"training-compute","value":7.5609676e+22,"unit":"FLOP","observationPeriod":"2025-03-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The video tokenizer was trained for 300,000 steps with a batch size of 128 using 128 H100 GPUs. Input sequences consisted of Tv = 24 video frames sampled at their native capture frequencies (20, 25, or 30 Hz). Random spatial crops of size 448 \u00d7 960 were extracted from the frames. For each training sample, a camera view was randomly selected from the available N = 5\nperspectives.\"\n\n\"The latent world model was trained for 460,000 steps with a batch size of 256 on 256 H100 GPUs. Inputs consisted of 48 video frames at native capture frequencies (20, 25, or 30 Hz), spatial resolution 448 \u00d7 960 and across N = 5 cameras.  After encoding these videos to the latent space, this corresponds to T \u00d7 N \u00d7 H \u00d7 W = 6 \u00d7 5 \u00d7 14 \u00d7 30 = 12,600 input tokens.\"\n\nvideo tokenizer training compute estimate:\n6 FLOP / token / parameter * 285*10^6 parameters * 300000 steps * 128 samples per step * 12600 tokens per sample = 8.273664e+20 FLOP\n\n\nworld model training compute estimation:\n\n6 FLOP / token / parameter * 8.4 * 10^9 parameters * 460000 steps * 256 samples per step * 12600 tokens per sample = 7.478231e+22 FLOP\n\nTotal: 8.273664e+20 FLOP + 7.478231e+22 FLOP = 7.5609676e+22 FLOP"},{"id":"epoch:78141c3bdf39ab22eb5a:parameters","entityId":"epoch:78141c3bdf39ab22eb5a","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-03-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:222bb9ff86cf61301653:parameters","entityId":"epoch:222bb9ff86cf61301653","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2025-03-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3B"},{"id":"epoch:f8623733a00764fc9962:parameters","entityId":"epoch:f8623733a00764fc9962","metric":"parameters","value":671000000000.0,"unit":"parameters","observationPeriod":"2025-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token."},{"id":"epoch:f8623733a00764fc9962:training-compute","entityId":"epoch:f8623733a00764fc9962","metric":"training-compute","value":3.3e+24,"unit":"FLOP","observationPeriod":"2025-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 \u00d7 37B \u00d7 14.8T = 3.29e24 pretraining FLOPs, 1e22 posttrainining FLOPs - see doc below\n\nhttps://docs.google.com/document/d/1C4zBvHh4UNnzaorwQSdG61zQJsBV9X3i3A70NeelxHY/edit?usp=sharing"},{"id":"epoch:76ad829292ae8b545126:parameters","entityId":"epoch:76ad829292ae8b545126","metric":"parameters","value":1100000000.0,"unit":"parameters","observationPeriod":"2025-03-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"This model has 1.1B model parameters.\""},{"id":"epoch:2361069e08ab48ef3682:parameters","entityId":"epoch:2361069e08ab48ef3682","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-03-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Dense decoder-only Transformer model"},{"id":"epoch:19e6eaa5b3abd90f2a43:parameters","entityId":"epoch:19e6eaa5b3abd90f2a43","metric":"parameters","value":253000000000.0,"unit":"parameters","observationPeriod":"2025-03-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"253B\n\"Dense decoder-only Transformer model Network Architecture: Llama-3.1-405B-Instruct, customized through Neural Architecture Search (NAS)\n\n**This model was developed based on Llama-3.1-405B-Instruct\n** This model has 253B model parameters.\""},{"id":"epoch:19e6eaa5b3abd90f2a43:training-compute","entityId":"epoch:19e6eaa5b3abd90f2a43","metric":"training-compute","value":3.911001e+25,"unit":"FLOP","observationPeriod":"2025-03-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"*Training compute figure is total lifetime (Meta + Nvidia) compute, not incremental Nvidia training compute*\n\nTotal training compute: 3.8e+25 FLOP (base model) + 1.11e+24 FLOP (fine-tuning) = 3.9e25 FLOP\nSee calculation in the finetune compute notes."},{"id":"epoch:8f35c80d0e4c1ccafc27:parameters","entityId":"epoch:8f35c80d0e4c1ccafc27","metric":"parameters","value":49000000000.0,"unit":"parameters","observationPeriod":"2025-03-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The model is a derivative of Meta\u2019s Llama-3.3-70B-Instruct, using Neural Architecture Search (NAS). The NAS algorithm results in non-standard and non-repetitive blocks. This includes the following: Skip attention: In some blocks, the attention is skipped entirely, or replaced with a single linear layer. Variable FFN: The expansion/compression ratio in the FFN layer is different between blocks.\n\nWe utilize a block-wise distillation of the reference model, where for each block we create multiple variants providing different tradeoffs of quality vs. computational complexity, discussed in more depth below. We then search over the blocks to create a model which meets the required throughput and memory (optimized for a single H100-80GB GPU) while minimizing the quality degradation."},{"id":"epoch:ac4b7bd68531200d1051:parameters","entityId":"epoch:ac4b7bd68531200d1051","metric":"parameters","value":2190000000.0,"unit":"parameters","observationPeriod":"2025-03-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Specifically, our publicly released GR00T-N1-2B model has 2.2B parameters in total, with 1.34B in the VLM.\"\n\nhttps://huggingface.co/nvidia/GR00T-N1-2B\n2.19e9 parameters, bf16, 4.38 GB"},{"id":"epoch:ac4b7bd68531200d1051:training-compute","entityId":"epoch:ac4b7bd68531200d1051","metric":"training-compute","value":7.12368e+22,"unit":"FLOP","observationPeriod":"2025-03-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"GR00T-N1-2B used roughly 50,000 H100 GPU hours for pretraining\"\nAssume H100 SXM5 variant, BF16, 0.4 utilization (NVIDIA in-house)\nH100 SXM5 BF16 performance = 989400000000000 FLOP/s = 9.894e14 FLOP/s\n(0.4 * 9.894e14 FLOP/s/GPU) * (3600 s/1 hr) * 50e3 GPU*hr = 7.12368e+22 FLOP"},{"id":"epoch:962d16db993962670648:parameters","entityId":"epoch:962d16db993962670648","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-03-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:962d16db993962670648:training-compute","entityId":"epoch:962d16db993962670648","metric":"training-compute","value":3.6059017e+24,"unit":"FLOP","observationPeriod":"2025-03-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.4e+24 FLOP [base model compute] + 2.205901651968e+24 FLOP = 3.6059017e+24 FLOP"},{"id":"epoch:2bd1784d4add94a28a6e:parameters","entityId":"epoch:2bd1784d4add94a28a6e","metric":"parameters","value":24000000000.0,"unit":"parameters","observationPeriod":"2025-03-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"24B"},{"id":"epoch:bfb836ba2629c5b9e7b2:parameters","entityId":"epoch:bfb836ba2629c5b9e7b2","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-03-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:bfb836ba2629c5b9e7b2:training-compute","entityId":"epoch:bfb836ba2629c5b9e7b2","metric":"training-compute","value":1.26e+24,"unit":"FLOP","observationPeriod":"2025-03-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.25 \u00d7 10^24 (base model reported training compute) + 7.04 \u00d7 10^21 (finetune compute) = 1.26 \u00d7 10^24 FLOP\n\nTable 1"},{"id":"epoch:b5e13497ad41d7d0f728:parameters","entityId":"epoch:b5e13497ad41d7d0f728","metric":"parameters","value":7800000000.0,"unit":"parameters","observationPeriod":"2025-03-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.8B"},{"id":"epoch:b5e13497ad41d7d0f728:training-compute","entityId":"epoch:b5e13497ad41d7d0f728","metric":"training-compute","value":4.23e+23,"unit":"FLOP","observationPeriod":"2025-03-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4.21 \u00d7 10^23 (base model reported training compute) + 1.71 \u00d7 10^21 (finetune compute) = 4.23 \u00d7 10^23 FLOP\n\nTable 1"},{"id":"epoch:803e1b0f9aa959cad783:parameters","entityId":"epoch:803e1b0f9aa959cad783","metric":"parameters","value":2400000000.0,"unit":"parameters","observationPeriod":"2025-03-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.4B"},{"id":"epoch:803e1b0f9aa959cad783:training-compute","entityId":"epoch:803e1b0f9aa959cad783","metric":"training-compute","value":9.41e+22,"unit":"FLOP","observationPeriod":"2025-03-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9.36 \u00d7 10^22 (base model reported training compute) + 5.27 \u00d7 10^20 (finetune compute) = 9.41 \u00d7 10^22 FLOP\n\nTable 1"},{"id":"epoch:b6148581dfa781693925:parameters","entityId":"epoch:b6148581dfa781693925","metric":"parameters","value":424000000000.0,"unit":"parameters","observationPeriod":"2025-03-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"MoE:\ntotal parameters - 424B\nactive parameters - 47B\n\n\"ViT encoder comprising 630 million parameters\""},{"id":"epoch:130f2bda6b54a607e4d2:parameters","entityId":"epoch:130f2bda6b54a607e4d2","metric":"parameters","value":2400000000.0,"unit":"parameters","observationPeriod":"2025-03-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.4B"},{"id":"epoch:130f2bda6b54a607e4d2:training-compute","entityId":"epoch:130f2bda6b54a607e4d2","metric":"training-compute","value":9.504e+22,"unit":"FLOP","observationPeriod":"2025-03-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9.36 \u00d7 10^22 (base model reported training compute) + 1.44 \u00d7 10^21 (finetune compute) = 9.504e+22 FLOP"},{"id":"epoch:f645ebeedc5d674e6ffc:parameters","entityId":"epoch:f645ebeedc5d674e6ffc","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-03-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:f645ebeedc5d674e6ffc:training-compute","entityId":"epoch:f645ebeedc5d674e6ffc","metric":"training-compute","value":1.2692e+24,"unit":"FLOP","observationPeriod":"2025-03-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.25 \u00d7 10^24 (base model reported training compute) + 1.92 \u00d7 10^22 (finetune compute) = 1.2692e+24 FLOP"},{"id":"epoch:c8fcd4e7f2ed9f266582:parameters","entityId":"epoch:c8fcd4e7f2ed9f266582","metric":"parameters","value":7800000000.0,"unit":"parameters","observationPeriod":"2025-03-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.8B"},{"id":"epoch:c8fcd4e7f2ed9f266582:training-compute","entityId":"epoch:c8fcd4e7f2ed9f266582","metric":"training-compute","value":4.2568e+23,"unit":"FLOP","observationPeriod":"2025-03-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4.21 \u00d7 10^23 (base model reported training compute) + 4.68 \u00d7 10^21 (finetune compute) = 4.2568e+23 FLOP"},{"id":"epoch:ae095dee8cad2778800c:parameters","entityId":"epoch:ae095dee8cad2778800c","metric":"parameters","value":111000000000.0,"unit":"parameters","observationPeriod":"2025-03-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"a 111 billion parameter model optimized for demanding enterprises that require fast, secure, and high-quality AI"},{"id":"epoch:a9784ff9703668fbc9d9:parameters","entityId":"epoch:a9784ff9703668fbc9d9","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2025-03-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1B (Table 1)"},{"id":"epoch:a9784ff9703668fbc9d9:training-compute","entityId":"epoch:a9784ff9703668fbc9d9","metric":"training-compute","value":1.2292301e+21,"unit":"FLOP","observationPeriod":"2025-03-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1: equivalent to 19 days on 8 A100 GPUs using fp16 precision (same as 48 H100 GPU days)\n\n312*10^12 FLOP / sec / GPU * 8 GPUs * 19 days * 24 hours / day * 3600 sec / hour * 0.3 [assumed utilization] = 1.2292301e+21 FLOP"},{"id":"epoch:bb5a76fb8be0be20bcad:parameters","entityId":"epoch:bb5a76fb8be0be20bcad","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-03-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:bb5a76fb8be0be20bcad:training-compute","entityId":"epoch:bb5a76fb8be0be20bcad","metric":"training-compute","value":1.3e+24,"unit":"FLOP","observationPeriod":"2025-03-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"first table here: https://allenai.org/blog/olmo2-32B"},{"id":"epoch:21e4376fdd6db57cdc44:parameters","entityId":"epoch:21e4376fdd6db57cdc44","metric":"parameters","value":27000000000.0,"unit":"parameters","observationPeriod":"2025-03-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Vision Encoder: 417M\nEmbedding Parameters:  1,416M\nNon-embedding Parameters:  25,600M\n"},{"id":"epoch:21e4376fdd6db57cdc44:training-compute","entityId":"epoch:21e4376fdd6db57cdc44","metric":"training-compute","value":2.268e+24,"unit":"FLOP","observationPeriod":"2025-03-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND =  6 * 27B parameters * 14T training tokens = 2.268 \u00d7 10^24 FLOP"},{"id":"epoch:e4b9a931658f8c645392:parameters","entityId":"epoch:e4b9a931658f8c645392","metric":"parameters","value":12000000000.0,"unit":"parameters","observationPeriod":"2025-03-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Vision Encoder: 417M\nEmbedding Parameters: 1,012M\nNon-embedding Parameters: 10,759M"},{"id":"epoch:e4b9a931658f8c645392:training-compute","entityId":"epoch:e4b9a931658f8c645392","metric":"training-compute","value":8.64e+23,"unit":"FLOP","observationPeriod":"2025-03-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 * 12B parameters * 12T training tokens = 8.64 \u00d7 10^23 FLOP"},{"id":"epoch:605b4d866f5ce10cceed:parameters","entityId":"epoch:605b4d866f5ce10cceed","metric":"parameters","value":4000000000.0,"unit":"parameters","observationPeriod":"2025-03-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Vision Encoder: 417M\nEmbedding Parameters: 675M\nNon-embedding Parameters: 3,209M"},{"id":"epoch:605b4d866f5ce10cceed:training-compute","entityId":"epoch:605b4d866f5ce10cceed","metric":"training-compute","value":9.6e+22,"unit":"FLOP","observationPeriod":"2025-03-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 * 4B parameters * 4T training tokens = 9.6 \u00d7 10^22 FLOP"},{"id":"epoch:e13eaa75ecd355be938f:parameters","entityId":"epoch:e13eaa75ecd355be938f","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2025-03-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Embedding Parameters: 302M\nNon-embedding Parameters: 698M"},{"id":"epoch:e13eaa75ecd355be938f:training-compute","entityId":"epoch:e13eaa75ecd355be938f","metric":"training-compute","value":1.2e+22,"unit":"FLOP","observationPeriod":"2025-03-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 * 1B parameters * 2T training tokens = 1.2 \u00d7 10^22 FLOP"},{"id":"epoch:8e07ab422fdf09c8936a:parameters","entityId":"epoch:8e07ab422fdf09c8936a","metric":"parameters","value":560000000000.0,"unit":"parameters","observationPeriod":"2025-03-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" the model scales to 56B activated\nparameters and 560B total parameters"},{"id":"epoch:8e07ab422fdf09c8936a:training-compute","entityId":"epoch:8e07ab422fdf09c8936a","metric":"training-compute","value":5.4e+24,"unit":"FLOP","observationPeriod":"2025-03-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"using C = 6*ND, 6* 56 billion parameters * 16 trillion tokens = ~5.4e24 FLOP"},{"id":"epoch:b3b27ac1cef413b167d2:parameters","entityId":"epoch:b3b27ac1cef413b167d2","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-03-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:a64c6f6037746ad2a159:parameters","entityId":"epoch:a64c6f6037746ad2a159","metric":"parameters","value":21000000000.0,"unit":"parameters","observationPeriod":"2025-03-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1de29593fc079d6f66af:parameters","entityId":"epoch:1de29593fc079d6f66af","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2025-03-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:126b10007cbf9d8028bb:parameters","entityId":"epoch:126b10007cbf9d8028bb","metric":"parameters","value":16800000000.0,"unit":"parameters","observationPeriod":"2025-03-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"16.8 billion parameters with 2.75 billion activated parameters"},{"id":"epoch:126b10007cbf9d8028bb:training-compute","entityId":"epoch:126b10007cbf9d8028bb","metric":"training-compute","value":1.485e+23,"unit":"FLOP","observationPeriod":"2025-03-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 2.75 * 10^9 active parameters * 9 * 10^12 tokens = 1.485e+23 FLOP"},{"id":"epoch:c8a562d98eaef7b0ad86:parameters","entityId":"epoch:c8a562d98eaef7b0ad86","metric":"parameters","value":290000000000.0,"unit":"parameters","observationPeriod":"2025-03-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"290 billion parameters with 28.8 billion activated parameters"},{"id":"epoch:c8a562d98eaef7b0ad86:training-compute","entityId":"epoch:c8a562d98eaef7b0ad86","metric":"training-compute","value":1.5552e+24,"unit":"FLOP","observationPeriod":"2025-03-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 28.8 * 10^9 active parameters * 9 * 10^12 tokens = 1.5552e+24 FLOP"},{"id":"epoch:933f332efa7ad07d147b:parameters","entityId":"epoch:933f332efa7ad07d147b","metric":"parameters","value":32500000000.0,"unit":"parameters","observationPeriod":"2025-03-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Architecture: transformers with RoPE, SwiGLU, RMSNorm, and Attention QKV bias\nNumber of Parameters: 32.5B\nNumber of Paramaters (Non-Embedding): 31.0B\nNumber of Layers: 64\nNumber of Attention Heads (GQA): 40 for Q and 8 for KV"},{"id":"epoch:933f332efa7ad07d147b:training-compute","entityId":"epoch:933f332efa7ad07d147b","metric":"training-compute","value":3.51e+24,"unit":"FLOP","observationPeriod":"2025-03-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming the same dataset size as for Qwen2.5 training (18T tokens):\n\n6ND = 6 * 32500000000 parameters * 18 * 10^12 tokens =  3.51 \u00d7 10^24\n\n'Speculative' confidence"},{"id":"epoch:f1a3d0dd311044e7c190:parameters","entityId":"epoch:f1a3d0dd311044e7c190","metric":"parameters","value":52000000000.0,"unit":"parameters","observationPeriod":"2025-03-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12B active/52B total"},{"id":"epoch:ca54e85b3a9ce6544cc0:parameters","entityId":"epoch:ca54e85b3a9ce6544cc0","metric":"parameters","value":398000000000.0,"unit":"parameters","observationPeriod":"2025-03-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"94B active/398B total"},{"id":"epoch:f0b383e96de82446b514:parameters","entityId":"epoch:f0b383e96de82446b514","metric":"parameters","value":1900000000.0,"unit":"parameters","observationPeriod":"2025-03-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c43180ded75eec4dd24c:parameters","entityId":"epoch:c43180ded75eec4dd24c","metric":"parameters","value":3800000000.0,"unit":"parameters","observationPeriod":"2025-03-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.8-billion\ndense decoder-only Transformer model"},{"id":"epoch:c43180ded75eec4dd24c:training-compute","entityId":"epoch:c43180ded75eec4dd24c","metric":"training-compute","value":9.9561596e+22,"unit":"FLOP","observationPeriod":"2025-03-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 FLOP / token / parameter * 3800000000 parameters * 5000000000000 tokens = 1.14e+23 FLOP\n\n512 GPUs * 312000000000000 FLOP / sec * 504 hours * 3600 sec / hour * 0.3 [assumed utilization] = 8.6951854e+22 FLOP\n\ngeometric mean: \nsqrt(1.14e+23*8.6951854e+22) = 9.9561596e+22"},{"id":"epoch:3f042fd97eccbefa2f5b:parameters","entityId":"epoch:3f042fd97eccbefa2f5b","metric":"parameters","value":5600000000.0,"unit":"parameters","observationPeriod":"2025-03-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"5.6B\n\n1. base: Phi-4 Mini (3.8b parameters) \n2. \"The audio encoder and projector introduce 460M parameters while LoRAA consumes another 460M parameters.\"\n3. \"The image encoder and projector introduce 440M model parameters while the vision adapter LoRAV consumes another 370M model parameters.\""},{"id":"epoch:3f042fd97eccbefa2f5b:training-compute","entityId":"epoch:3f042fd97eccbefa2f5b","metric":"training-compute","value":1.1852519e+23,"unit":"FLOP","observationPeriod":"2025-03-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.14e+23 (base model training compute) + 7.1724e+21 (finetune compute) = 1.211724e+23\n\nHF instruct model:\n\"GPUs: 512 A100-80G\nTraining time: 28 days\"\n\n512 GPUs * 312000000000000 FLOP / sec * 28 days * 24 hours / day * 3600 sec / hour * 0.3 [assumed utilization] = 1.1593581e+23 FLOP\n\ngeometric mean: \nsqrt(1.211724e+23 * 1.1593581e+23) = 1.1852519e+23"},{"id":"epoch:eff4ef18c2f6bdd58093:parameters","entityId":"epoch:eff4ef18c2f6bdd58093","metric":"parameters","value":33100000000.0,"unit":"parameters","observationPeriod":"2025-03-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"According to huggingface.co/CohereLabs/aya-vision-32b."},{"id":"epoch:19d39c50be8694a76ede:parameters","entityId":"epoch:19d39c50be8694a76ede","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2025-03-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2d31eaa3a97805c153ba:training-compute","entityId":"epoch:2d31eaa3a97805c153ba","metric":"training-compute","value":3.8e+26,"unit":"FLOP","observationPeriod":"2025-02-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Analysis of GPT-4.5's training cluster yields a median estimate of 187M H100-hours of training. The utilization assumptions we used for the Grok 3 estimate (maybe worth revisiting) were 20 to 40% under the H100 FP8 spec of 2000 teraflop/s. This leads to an estimate of 2.7e26 to 5.4e26 FLOP, or a geomean of 3.8e26\n\nAlternatively, using a plausible range of 20 to 50% utilization, given the possibility of FP8 training, yields a median estimate of ~2e25 FLOP. See notebook below for details.\n\nhttps://colab.research.google.com/drive/1QBmVPm64Ti0xucN0EsZTgSz_I7Mj9hAZ#scrollTo=NYH1ABJuLJlw \n\nThis is consistent with OpenAI's statement that GPT-4.5 was a \u201cnew order of magnitude in compute\u201d compared to previous models (e.g. GPT-4, which was ~2e25), suggesting around 2e26 FLOP. But they could have meant this somewhat loosely. \n\nIn the \"Pretraining GPT-4.5\" interview, they state they used multi-cluster training: https://youtu.be/6nJZopACRuQ?si=FFJC-gEmGPZjvoPM&t=617 "},{"id":"epoch:b222914fde96d6b18d4b:parameters","entityId":"epoch:b222914fde96d6b18d4b","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-02-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:8afca030fc8466a7c8d1:parameters","entityId":"epoch:8afca030fc8466a7c8d1","metric":"parameters","value":2530000000.0,"unit":"parameters","observationPeriod":"2025-02-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a4a420046aa03cf384d0:parameters","entityId":"epoch:a4a420046aa03cf384d0","metric":"parameters","value":2500000000.0,"unit":"parameters","observationPeriod":"2025-02-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e8e869bed58f7626ef3f:parameters","entityId":"epoch:e8e869bed58f7626ef3f","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2025-02-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B"},{"id":"epoch:e8e869bed58f7626ef3f:training-compute","entityId":"epoch:e8e869bed58f7626ef3f","metric":"training-compute","value":2.5e+23,"unit":"FLOP","observationPeriod":"2025-02-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Through extensive experimentation, the model is validated at scale, reaching 14 billion parameters. Subsequently, Wan has seen large-scale data comprising billions of images and videos, amounting to O(1) trillions of tokens in total.\"\n\nSo likely between 1T and 10T tokens. Assume 3T.\n\nTransformer architecture, so 6ND should be a decent approximation.\n\n6ND = 6 * 14e9 * 3e12 ~= 2.5e+23 FLOP"},{"id":"epoch:70b848f1f66ddbf621c6:parameters","entityId":"epoch:70b848f1f66ddbf621c6","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-02-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2fc6d2f99e6cbd964fc4:parameters","entityId":"epoch:2fc6d2f99e6cbd964fc4","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-02-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B (same as the base model)"},{"id":"epoch:ab873f037b189d03b519:parameters","entityId":"epoch:ab873f037b189d03b519","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-02-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:ab873f037b189d03b519:training-compute","entityId":"epoch:ab873f037b189d03b519","metric":"training-compute","value":7.3536e+23,"unit":"FLOP","observationPeriod":"2025-02-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 8 * 10^9 parameters * 15.32 * 10^12 tokens [see dataset size notes] = 7.3536e+23 FLOP"},{"id":"epoch:93d739383209b1447f97:training-compute","entityId":"epoch:93d739383209b1447f97","metric":"training-compute","value":3.35e+25,"unit":"FLOP","observationPeriod":"2025-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://docs.google.com/spreadsheets/d/10bhwdVrfHI8tysVIz62ZxtvQ30L-HojYvmU18_b-WIM/edit?gid=0#gid=0"},{"id":"epoch:12aae2b350eaf0d652b3:parameters","entityId":"epoch:12aae2b350eaf0d652b3","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2025-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"30B"},{"id":"epoch:12aae2b350eaf0d652b3:training-compute","entityId":"epoch:12aae2b350eaf0d652b3","metric":"training-compute","value":4.1015808e+24,"unit":"FLOP","observationPeriod":"2025-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We have constructed a datacenter comprising thousands of NVIDIA H800 GPUs\"\n\"we have achieved 99% effective GPU\ntraining time over more than one month.\"\n\n989000000000000 FLOP / GPU / sec [bf16 assumed] * 720 hours * 3600 sec / hour * 5000 GPUs [assumption -> \"Likely\" confidence] * 0.32 [reported utilization] = 4.1015808e+24 FLOP"},{"id":"epoch:40f19aab3fb93be4835c:parameters","entityId":"epoch:40f19aab3fb93be4835c","metric":"parameters","value":7080000000.0,"unit":"parameters","observationPeriod":"2025-02-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our system comprises two main components: S2, a VLM backbone, and S1, a latent-conditional visuomotor transformer. S2 is built on a 7B-parameter open-source, open-weight VLM pretrained on internet-scale data. <..>\n\nS1, an 80M parameter cross-attention encoder-decoder transformer, handles low-level control.\"\n\n7B+80M = 7080000000"},{"id":"epoch:573818d456e7a42a9f97:parameters","entityId":"epoch:573818d456e7a42a9f97","metric":"parameters","value":1140000000.0,"unit":"parameters","observationPeriod":"2025-02-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1B"},{"id":"epoch:573818d456e7a42a9f97:training-compute","entityId":"epoch:573818d456e7a42a9f97","metric":"training-compute","value":8.208e+22,"unit":"FLOP","observationPeriod":"2025-02-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 1140000000 parameters * 12000000000000 tokens [see dataset size notes] = 8.208e+22 FLOP"},{"id":"epoch:2451cc602d6224bae47f:parameters","entityId":"epoch:2451cc602d6224bae47f","metric":"parameters","value":40300000000.0,"unit":"parameters","observationPeriod":"2025-02-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1 lists 40.3B parameters as model size."},{"id":"epoch:2451cc602d6224bae47f:training-compute","entityId":"epoch:2451cc602d6224bae47f","metric":"training-compute","value":2.25e+24,"unit":"FLOP","observationPeriod":"2025-02-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"40.3e9 parameters * 9.3e12 training datapoints * 6 = 2.25e24.\nSame FLOPS estimate given by authors in Table 1."},{"id":"epoch:767372a5526df34aba22:parameters","entityId":"epoch:767372a5526df34aba22","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-02-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:767372a5526df34aba22:training-compute","entityId":"epoch:767372a5526df34aba22","metric":"training-compute","value":1.008e+23,"unit":"FLOP","observationPeriod":"2025-02-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7e9 parameters *2.4e12 training datapoints*6=1.008e23 FLOP "},{"id":"epoch:09b1ac81f48df4cd7f28:parameters","entityId":"epoch:09b1ac81f48df4cd7f28","metric":"parameters","value":2920000000.0,"unit":"parameters","observationPeriod":"2025-02-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/google/paligemma-3b-mix-224"},{"id":"epoch:09b1ac81f48df4cd7f28:training-compute","entityId":"epoch:09b1ac81f48df4cd7f28","metric":"training-compute","value":4.0397694e+22,"unit":"FLOP","observationPeriod":"2025-02-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.12e+22 FLOP [base LLM compute] + 4.9467301e+21 [base VAE compute] + 4.2509643e+21 FLOP [finetune compute] = 4.0397694e+22FLOP"},{"id":"epoch:0a604da3d83629a088b9:parameters","entityId":"epoch:0a604da3d83629a088b9","metric":"parameters","value":2920000000.0,"unit":"parameters","observationPeriod":"2025-02-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/google/paligemma-3b-mix-448"},{"id":"epoch:4bccda90fc0562a35773:parameters","entityId":"epoch:4bccda90fc0562a35773","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2025-02-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The model is initialized with pre-trained weights from the Qwen2.5 LLM"},{"id":"epoch:4bccda90fc0562a35773:training-compute","entityId":"epoch:4bccda90fc0562a35773","metric":"training-compute","value":9.5712e+24,"unit":"FLOP","observationPeriod":"2025-02-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.8e+24 FLOP [base LLM compute] + 1.7712e+24 FLOP [VL training compute] = 9.5712e+24 FLOP"},{"id":"epoch:09b4b4b91bb6349b48e8:parameters","entityId":"epoch:09b4b4b91bb6349b48e8","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-02-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The model is initialized with pre-trained weights from the Qwen2.5 LLM"},{"id":"epoch:09b4b4b91bb6349b48e8:training-compute","entityId":"epoch:09b4b4b91bb6349b48e8","metric":"training-compute","value":9.9408e+23,"unit":"FLOP","observationPeriod":"2025-02-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8.2188e+23 FLOP [base LLM compute] + 1.722e+23 FLOP [VL training compute] = 9.9408e+23 FLOP"},{"id":"epoch:a6e3626a334a0025b4e7:parameters","entityId":"epoch:a6e3626a334a0025b4e7","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2025-02-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The model is initialized with pre-trained weights from the Qwen2.5 LLM"},{"id":"epoch:a6e3626a334a0025b4e7:training-compute","entityId":"epoch:a6e3626a334a0025b4e7","metric":"training-compute","value":4.0752e+23,"unit":"FLOP","observationPeriod":"2025-02-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.3372e+23 FLOP [base LLM compute] + 7.38e+22 FLOP [VL training compute] = 4.0752e+23 FLOP"},{"id":"epoch:d0e6f264d799068c49b8:parameters","entityId":"epoch:d0e6f264d799068c49b8","metric":"parameters","value":671000000000.0,"unit":"parameters","observationPeriod":"2025-02-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"671B params"},{"id":"epoch:0bb118ba49757d5d7929:parameters","entityId":"epoch:0bb118ba49757d5d7929","metric":"parameters","value":400000000.0,"unit":"parameters","observationPeriod":"2025-02-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"a total of \u223c 400M parameters (258M for Conv, 138M for Trans)"},{"id":"epoch:0bb118ba49757d5d7929:training-compute","entityId":"epoch:0bb118ba49757d5d7929","metric":"training-compute","value":1.62e+18,"unit":"FLOP","observationPeriod":"2025-02-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"125000000000000 FLOP / GPU / sec [V100 reported, bf16 assumed] * 1 GPU  * 12 hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.62e+18 FLOP"},{"id":"epoch:ea1cf395c2f390a0ad81:parameters","entityId":"epoch:ea1cf395c2f390a0ad81","metric":"parameters","value":130000000000.0,"unit":"parameters","observationPeriod":"2025-02-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"130B"},{"id":"epoch:ea1cf395c2f390a0ad81:training-compute","entityId":"epoch:ea1cf395c2f390a0ad81","metric":"training-compute","value":1.92504e+24,"unit":"FLOP","observationPeriod":"2025-02-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"base LLM 6.24e+23 FLOP + 6 FLOP / parameter / token * 130*10^9 parameters * 1668000000000 tokens [see training dataset size notes] = 6.24e+23 FLOP + 1.30104e+24 FLOP = 1.92504e+24 FLOP"},{"id":"epoch:b772cb7586411f1de506:parameters","entityId":"epoch:b772cb7586411f1de506","metric":"parameters","value":130000000000.0,"unit":"parameters","observationPeriod":"2025-02-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"130B"},{"id":"epoch:b772cb7586411f1de506:training-compute","entityId":"epoch:b772cb7586411f1de506","metric":"training-compute","value":6.24e+23,"unit":"FLOP","observationPeriod":"2025-02-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 130 * 10^9 parameters * 800 * 10^9 tokens = 6.24e+23 FLOP\n\n[1 epoch assumed]"},{"id":"epoch:4227d2ffa9e4651e03f4:parameters","entityId":"epoch:4227d2ffa9e4651e03f4","metric":"parameters","value":130000000000.0,"unit":"parameters","observationPeriod":"2025-02-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"130B"},{"id":"epoch:4227d2ffa9e4651e03f4:training-compute","entityId":"epoch:4227d2ffa9e4651e03f4","metric":"training-compute","value":2.54904e+24,"unit":"FLOP","observationPeriod":"2025-02-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"base LLM 6.24e+23 FLOP + 6 FLOP / parameter / token * 130*10^9 parameters * 2468000000000 tokens [see training dataset size notes] = 6.24e+23 FLOP + 1.92504e+24 FLOP = 2.54904e+24 FLOP"},{"id":"epoch:1a86e4bd8eb269cfea5e:parameters","entityId":"epoch:1a86e4bd8eb269cfea5e","metric":"parameters","value":3000000000000.0,"unit":"parameters","observationPeriod":"2025-02-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1a86e4bd8eb269cfea5e:training-compute","entityId":"epoch:1a86e4bd8eb269cfea5e","metric":"training-compute","value":3.5e+26,"unit":"FLOP","observationPeriod":"2025-02-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Estimate based on a cluster of 80,000 H100s per the xai website and an estimated training time of approximately three months.\n\nFull estimate here: https://docs.google.com/document/d/1MIUFviULJ3YI_XjyzL8cwG0cBRANKNxVEB4DrUcFiNs/edit?usp=sharing"},{"id":"epoch:dd1bbe6844d6d0ba348e:parameters","entityId":"epoch:dd1bbe6844d6d0ba348e","metric":"parameters","value":24000000000.0,"unit":"parameters","observationPeriod":"2025-02-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"24B"},{"id":"epoch:d9714a20e22519273f16:parameters","entityId":"epoch:d9714a20e22519273f16","metric":"parameters","value":4000000000.0,"unit":"parameters","observationPeriod":"2025-02-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4B"},{"id":"epoch:d738d4fd6481d2b3886b:parameters","entityId":"epoch:d738d4fd6481d2b3886b","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-02-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:64a95769bd37ebf999d3:parameters","entityId":"epoch:64a95769bd37ebf999d3","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-02-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:74e7398adff56b86ad55:parameters","entityId":"epoch:74e7398adff56b86ad55","metric":"parameters","value":2980000000.0,"unit":"parameters","observationPeriod":"2025-02-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.98B"},{"id":"epoch:d24fb548778992549d02:parameters","entityId":"epoch:d24fb548778992549d02","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2025-02-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3B"},{"id":"epoch:b9f28d76fb6f47c6107d:parameters","entityId":"epoch:b9f28d76fb6f47c6107d","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-02-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:4a87511ac17470fa0023:parameters","entityId":"epoch:4a87511ac17470fa0023","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-02-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a9232b8e226b588fb4c0:parameters","entityId":"epoch:a9232b8e226b588fb4c0","metric":"parameters","value":1600000000.0,"unit":"parameters","observationPeriod":"2025-02-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.6B\n"},{"id":"epoch:f812c402c5b80b3898ad:parameters","entityId":"epoch:f812c402c5b80b3898ad","metric":"parameters","value":13493916736.0,"unit":"parameters","observationPeriod":"2025-02-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/yili18/Hamster_dev\nVLM:  VILA 1.5-13b, 13493916736 parameters\npolicy: either RVT-2 or 3D-DA\n\nthis record is for the HAMSTER VLM, so the policy is not included\n\n\n"},{"id":"epoch:f812c402c5b80b3898ad:training-compute","entityId":"epoch:f812c402c5b80b3898ad","metric":"training-compute","value":2.4081408e+21,"unit":"FLOP","observationPeriod":"2025-02-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.3003136e+21 FLOP [base model compute] + 1.078272e20 FLOP [fimetune compute] = 2.4081408e+21 FLOP"},{"id":"epoch:29a97b10c054797cd752:parameters","entityId":"epoch:29a97b10c054797cd752","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-02-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"implemented through the Goku model family, which comprises Transformer architectures with 2B and 8B parameters.\""},{"id":"epoch:c7ebb3e788d0e95d7975:parameters","entityId":"epoch:c7ebb3e788d0e95d7975","metric":"parameters","value":600000000.0,"unit":"parameters","observationPeriod":"2025-02-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"600M"},{"id":"epoch:c7ebb3e788d0e95d7975:training-compute","entityId":"epoch:c7ebb3e788d0e95d7975","metric":"training-compute","value":1.954368e+22,"unit":"FLOP","observationPeriod":"2025-02-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000 FLOP / GPU / sec [bf16 assumed] * 58000 GPU-hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.954368e+22 FLOP"},{"id":"epoch:2f8d541a9c139715e4ff:parameters","entityId":"epoch:2f8d541a9c139715e4ff","metric":"parameters","value":300000000.0,"unit":"parameters","observationPeriod":"2025-02-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"300M"},{"id":"epoch:2f8d541a9c139715e4ff:training-compute","entityId":"epoch:2f8d541a9c139715e4ff","metric":"training-compute","value":7.07616e+21,"unit":"FLOP","observationPeriod":"2025-02-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000 FLOP / GPU / sec [bf16 assumed] * 21000 GPU-hours * 3600 sec / hour * 0.3 [assumed utilization] = 7.07616e+21 FLOP"},{"id":"epoch:8af77cf34686a353ae2b:parameters","entityId":"epoch:8af77cf34686a353ae2b","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-01-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:85970874180caad54ab0:parameters","entityId":"epoch:85970874180caad54ab0","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-01-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:40b13cfadebc57cfbeb8:parameters","entityId":"epoch:40b13cfadebc57cfbeb8","metric":"parameters","value":24000000000.0,"unit":"parameters","observationPeriod":"2025-01-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"24B"},{"id":"epoch:40b13cfadebc57cfbeb8:training-compute","entityId":"epoch:40b13cfadebc57cfbeb8","metric":"training-compute","value":1.152e+24,"unit":"FLOP","observationPeriod":"2025-01-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*8T tokens * 24B parameters = 1.152e+24 FLOP"},{"id":"epoch:69ae84a9bab231182d17:parameters","entityId":"epoch:69ae84a9bab231182d17","metric":"parameters","value":405000000000.0,"unit":"parameters","observationPeriod":"2025-01-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"405B"},{"id":"epoch:12c6527f3b11fc6f88f2:parameters","entityId":"epoch:12c6527f3b11fc6f88f2","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-01-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B\n\nIn our experiments, we utilize DeepSeek-LLM (1.5B and 7B) [3] with a maximum supported sequence length of 4096 as the base language model. For the vision encoder used in understanding tasks, we select SigLIP-Large-Patch16-384 [53]."},{"id":"epoch:8e4d20ec331aa724f79b:parameters","entityId":"epoch:8e4d20ec331aa724f79b","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2025-01-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1B\n\nIn our experiments, we utilize DeepSeek-LLM (1.5B and 7B) [3] with a maximum supported sequence length of 4096 as the base language model. For the vision encoder used in understanding\ntasks, we select SigLIP-Large-Patch16-384 [53].\n\n** I am not sure which DeepSeek-LLM 1.5B they are referring to, there is no such model in the linked paper"},{"id":"epoch:07d2702791c144a41c09:parameters","entityId":"epoch:07d2702791c144a41c09","metric":"parameters","value":82000000.0,"unit":"parameters","observationPeriod":"2025-01-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"82M"},{"id":"epoch:07d2702791c144a41c09:training-compute","entityId":"epoch:07d2702791c144a41c09","metric":"training-compute","value":1.6848e+20,"unit":"FLOP","observationPeriod":"2025-01-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000 FLOP / GPU / sec * 500 GPU - hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.6848e+20 FLOP"},{"id":"epoch:0851dc5448ece9fc2711:parameters","entityId":"epoch:0851dc5448ece9fc2711","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2025-01-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Safetensors: 11B params"},{"id":"epoch:0851dc5448ece9fc2711:training-compute","entityId":"epoch:0851dc5448ece9fc2711","metric":"training-compute","value":3.3e+22,"unit":"FLOP","observationPeriod":"2025-01-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"500B tokens of \"high-quality data (text, audio, and vision)\"\n11B parameters (safetensors)\nthe model was trained by modules, epochs unknown -> with 'likely' confidence approximate training compute is\n\n6 FLOP / parameter / token * 11 * 10^9 parameters * 500 * 10^9 tokens = 3.3e+22 FLOP"},{"id":"epoch:29920ddce720f70b65c1:parameters","entityId":"epoch:29920ddce720f70b65c1","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2025-01-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:c47af3168e3c7732deed:parameters","entityId":"epoch:c47af3168e3c7732deed","metric":"parameters","value":14800000000.0,"unit":"parameters","observationPeriod":"2025-01-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14.8B (safetensors)"},{"id":"epoch:3b72a0b74ea51b4d66aa:parameters","entityId":"epoch:3b72a0b74ea51b4d66aa","metric":"parameters","value":1780000000.0,"unit":"parameters","observationPeriod":"2025-01-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.78B (safetensors)"},{"id":"epoch:da73a813dcc2dbf7ff5e:parameters","entityId":"epoch:da73a813dcc2dbf7ff5e","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-01-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:56766c2a2a33e1c698b0:parameters","entityId":"epoch:56766c2a2a33e1c698b0","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-01-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:c2b4bdc4c2252d78624c:parameters","entityId":"epoch:c2b4bdc4c2252d78624c","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-01-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:a053d5a764a5679a021e:parameters","entityId":"epoch:a053d5a764a5679a021e","metric":"parameters","value":149000000.0,"unit":"parameters","observationPeriod":"2025-01-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"149M"},{"id":"epoch:a053d5a764a5679a021e:training-compute","entityId":"epoch:a053d5a764a5679a021e","metric":"training-compute","value":3.676128e+21,"unit":"FLOP","observationPeriod":"2025-01-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 149 * 10^6 parameters * 1028000000000 tokens * 4 epochs [see dataset size notes] = 3.676128e+21 FLOP"},{"id":"epoch:cd5d7546f9d55b391cfe:parameters","entityId":"epoch:cd5d7546f9d55b391cfe","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2025-01-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B (assuming same as base model Llama 3.3 70B)"},{"id":"epoch:9200b4f8baeda487e8c3:parameters","entityId":"epoch:9200b4f8baeda487e8c3","metric":"parameters","value":671000000000.0,"unit":"parameters","observationPeriod":"2025-01-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"671B total\n37B activated\nhttps://github.com/deepseek-ai/DeepSeek-R1/tree/main"},{"id":"epoch:9200b4f8baeda487e8c3:training-compute","entityId":"epoch:9200b4f8baeda487e8c3","metric":"training-compute","value":3.5e+24,"unit":"FLOP","observationPeriod":"2025-01-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.29e24 + 1.8e23 = 3.47e24, round to 3.5e24 due to lack of sigfigs"},{"id":"epoch:b9cd8d357d6b7abc5f53:parameters","entityId":"epoch:b9cd8d357d6b7abc5f53","metric":"parameters","value":671000000000.0,"unit":"parameters","observationPeriod":"2025-01-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"671B total\n37B activated\nhttps://github.com/deepseek-ai/DeepSeek-R1/tree/main"},{"id":"epoch:cd78e1cac8fe8b493ff3:parameters","entityId":"epoch:cd78e1cac8fe8b493ff3","metric":"parameters","value":8930000000.0,"unit":"parameters","observationPeriod":"2025-01-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 4\n\nhttps://huggingface.co/nvidia/Eagle2-9B\n\"8.93B params\"\n"},{"id":"epoch:cd78e1cac8fe8b493ff3:training-compute","entityId":"epoch:cd78e1cac8fe8b493ff3","metric":"training-compute","value":4.7156e+22,"unit":"FLOP","observationPeriod":"2025-01-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Appendix A. Compute: \"We show our training resource for Eagle2-9B in Tab. A. In actual development, we rarely iterate the Stage-1 model. Usually, we iterate Stage-1.5 once after iterating Stage-2 >10 times.\"\nAssume \">10 times\" -> 12\nAssume \"H100\" -> NVIDIA H100 SXM5 80GB\nAssume bf16\nAssume 0.4 utilization (NVIDIA in-house)\nH100 SXM5 performance = 989400000000000 FLOP/s = 9.894e14 FLOP/s\n\nStage 1:     (H100 * 128) * (2.5 hr * 1)\nStage 1.5:  (H100 * 256) * (28 hr * 2)\nStage 2:     (H100 * 256) * (6 hr * 12)\n\nStage 1:      0.4 * (9.894e14 FLOP/s * 128) * (3600 s / 1 hr) * (2.5 hr *1) ~= 4.56e20 FLOP\nStage 1.5:   0.4 * (9.894e14 FLOP/s * 256) * (3600 s / 1 hr) * (28 hr * 2) ~= 2.04e22 FLOP\nStage 2:      0.4 * (9.894e14 FLOP/s * 256) * (3600 s / 1 hr) * (6 hr * 12) ~= 2.63e22 FLOP\n\n(4.56e20 + 2.04e22 + 2.63e22) FLOP ~= 4.72e22 FLOP\n"},{"id":"epoch:3738bdceed5cdd432951:training-compute","entityId":"epoch:3738bdceed5cdd432951","metric":"training-compute","value":3.234816e+19,"unit":"FLOP","observationPeriod":"2025-01-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000 FLOP / GPU / sec * 8 GPUs * 12 hours * 3600 sec / hour * 0.3 [assumed utilization] = 3.234816e+19 FLOP"},{"id":"epoch:7743775c0cf6e5ff5387:parameters","entityId":"epoch:7743775c0cf6e5ff5387","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-01-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f295ba43c85b3d71f45e:parameters","entityId":"epoch:f295ba43c85b3d71f45e","metric":"parameters","value":46800000.0,"unit":"parameters","observationPeriod":"2025-01-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"MatterGen contains 46.8M parameters\"\n\nhttps://huggingface.co/microsoft/mattergen\n\n\"The model architecture is based on GemNet (Gasteiger et al. 2021).\""},{"id":"epoch:f295ba43c85b3d71f45e:training-compute","entityId":"epoch:f295ba43c85b3d71f45e","metric":"training-compute","value":2.69568e+19,"unit":"FLOP","observationPeriod":"2025-01-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000*3600*10*8*0.3 = 2.69568e+19"},{"id":"epoch:04ff01570610955898b0:parameters","entityId":"epoch:04ff01570610955898b0","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a247524b734310832af2:parameters","entityId":"epoch:a247524b734310832af2","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:12fb606c2c2f3bced1e5:parameters","entityId":"epoch:12fb606c2c2f3bced1e5","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2025-01-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:fd642dedcce7a47f50a5:parameters","entityId":"epoch:fd642dedcce7a47f50a5","metric":"parameters","value":456000000000.0,"unit":"parameters","observationPeriod":"2025-01-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Total Parameters: 456B\nActivated Parameters per Token: 45.9B\nNumber Layers: 80\nHybrid Attention: a softmax attention is positioned after every 7 lightning attention.\nNumber of attention heads: 64\nAttention head dimension: 128\nMixture of Experts:\nNumber of experts: 32\nExpert hidden dimension: 9216\nTop-2 routing strategy\nPositional Encoding: Rotary Position Embedding (RoPE) applied to half of the attention head dimension with a base frequency of 10,000,000\nHidden Size: 6144\nVocab Size: 200,064"},{"id":"epoch:fd642dedcce7a47f50a5:training-compute","entityId":"epoch:fd642dedcce7a47f50a5","metric":"training-compute","value":3.1417632e+24,"unit":"FLOP","observationPeriod":"2025-01-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 45.9 * 10^9 activated parameters * 1.1408e+13 tokens = 3.1417632e+24 FLOP\n\n\"Likely\" confidence because the model is MoE (formula might not be that accurate) \n\ndata trained on is  1.1408e+13 = ((16 million * 500)+7.2 trillion + 3.2 trillion + 1 trillion)"},{"id":"epoch:4ed76505d2e6eb6cfe15:training-compute","entityId":"epoch:4ed76505d2e6eb6cfe15","metric":"training-compute","value":2.1238848e+24,"unit":"FLOP","observationPeriod":"2025-01-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.98288e+24 FLOP [base model compute] + 1.410048e+23 FLOP [addtional vision-language finetune compute] = 2.1238848e+24 FLOP"},{"id":"epoch:81192e958ec5ce28fafb:parameters","entityId":"epoch:81192e958ec5ce28fafb","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2025-01-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:eee88059bb62545c3113:parameters","entityId":"epoch:eee88059bb62545c3113","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2025-01-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:d6e2e46e294669fb3369:parameters","entityId":"epoch:d6e2e46e294669fb3369","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2025-01-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B"},{"id":"epoch:d6e2e46e294669fb3369:training-compute","entityId":"epoch:d6e2e46e294669fb3369","metric":"training-compute","value":2.8e+24,"unit":"FLOP","observationPeriod":"2025-01-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train all of the WFM models reported in the paper using a cluster of 10,000 NVIDIA H100 GPUs in a time span of three months.\"\n\nI assign the FLOPs from this cluster proportional to the parameter size of the model trained. There are a total of 76B parameters between the 8 models. Therefore, assuming 20% utilization (starting with 33% but then accounting for time between experiments), we get\n(10k H100s)*(90 days)*(24*60*60)*(979e12)*(0.2 utilization)*(14/76) = 2.8e24 FLOPs\n"},{"id":"epoch:6d69d1bb3e8d28f7675f:parameters","entityId":"epoch:6d69d1bb3e8d28f7675f","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-01-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:6d69d1bb3e8d28f7675f:training-compute","entityId":"epoch:6d69d1bb3e8d28f7675f","metric":"training-compute","value":1.4e+24,"unit":"FLOP","observationPeriod":"2025-01-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train all of the WFM models reported in the paper using a cluster of 10,000 NVIDIA H100 GPUs in a time span of three months.\"\n\nI assign the FLOPs from this cluster proportional to the parameter size of the model trained. There are a total of 76B parameters between the 8 models. Therefore, assuming 20% utilization (starting with 33% but then accounting for time between experiments), we get\n(10k H100s)*(90 days)*(24*60*60)*(979e12)*(0.2 utilization)*(7/76) = 1.4e+24 FLOPs\n"},{"id":"epoch:07b99d4cbc42bb6899ce:parameters","entityId":"epoch:07b99d4cbc42bb6899ce","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2025-01-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B"},{"id":"epoch:07b99d4cbc42bb6899ce:training-compute","entityId":"epoch:07b99d4cbc42bb6899ce","metric":"training-compute","value":2.8e+24,"unit":"FLOP","observationPeriod":"2025-01-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train all of the WFM models reported in the paper using a cluster of 10,000 NVIDIA H100 GPUs in a time span of three months.\"\n\nI assign the FLOPs from this cluster proportional to the parameter size of the model trained. There are a total of 76B parameters between the 8 models. Therefore, assuming 20% utilization (starting with 33% but then accounting for time between experiments), we get\n(10k H100s)*(90 days)*(24*60*60)*(979e12)*(0.2 utilization)*(14/76) = 2.8e24 FLOPs\n"},{"id":"epoch:dc17df2ce9503c6beec8:parameters","entityId":"epoch:dc17df2ce9503c6beec8","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2025-01-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:dc17df2ce9503c6beec8:training-compute","entityId":"epoch:dc17df2ce9503c6beec8","metric":"training-compute","value":4.01915904e+20,"unit":"FLOP","observationPeriod":"2025-01-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Hardware-based estimate from the RL training: 592 steps (from Table 1) * 680s/step (Table 2, PRIME row) * 8 (they used A800) * 312,000,000,000,000 FLOPS * 0.4 assumed utilization = 4.01915904E20 FLOPs"},{"id":"epoch:dfac7220bc94dd6d4f6e:parameters","entityId":"epoch:dfac7220bc94dd6d4f6e","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:dfac7220bc94dd6d4f6e:training-compute","entityId":"epoch:dfac7220bc94dd6d4f6e","metric":"training-compute","value":1.8e+23,"unit":"FLOP","observationPeriod":"2024-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.8*10^23 FLOPs (Table 6 - developers calculated using 6ND formula)"},{"id":"epoch:14eb15efc820c9c710ff:parameters","entityId":"epoch:14eb15efc820c9c710ff","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:14eb15efc820c9c710ff:training-compute","entityId":"epoch:14eb15efc820c9c710ff","metric":"training-compute","value":4.6e+23,"unit":"FLOP","observationPeriod":"2024-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4.6*10^23 FLOPs (Table 6 - developers calculated using 6ND formula)\n"},{"id":"epoch:cf0dddf2a4de37767528:parameters","entityId":"epoch:cf0dddf2a4de37767528","metric":"parameters","value":100598707.0,"unit":"parameters","observationPeriod":"2024-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"A.3 Baseline Implementations: \"The number of trainable parameters of these models are controlled to be similar, i.e., GS-LRM has 86.68M parameters, LGM has 103.29M parameters, while our default STORM has 100.60M parameters.\n\n3.3 Implementation, Model architecture: \"By default, we use a 12-layer Vision Transformer (ViT-B) (Dosovitskiy, 2020) with full attention and a patch size of 8, along with M = 16 motion tokens.\"\n\nhttps://github.com/NVlabs/GaussianSTORM?tab=readme-ov-file#training\nTraining: \"Multi-GPU example that reproduces the paper's STORM-B/8 model:\n\nhttps://github.com/NVlabs/GaussianSTORM/blob/main/storm/models/storm.py\ndef STORM_B_8(**kwargs):\n    return STORM(patch_size=8, embed_dim=768, depth=12, num_heads=12, **kwargs)\ndef STORM_L_8(**kwargs):\n    return STORM(patch_size=8, embed_dim=1024, depth=24, num_heads=16, **kwargs)\ndef STORM_B_16(**kwargs):\n    return STORM(patch_size=16, embed_dim=768, depth=12, num_heads=12, **kwargs)\ndef STORM_L_16(**kwargs):\n    return STORM(patch_size=16, embed_dim=1024, depth=24, num_heads=16, **kwargs)\ndef STORM_XL_8(**kwargs):\n    return STORM(patch_size=8, embed_dim=1152, depth=28, num_heads=16, **kwargs)\ndef STORM_H_8(**kwargs):\n    return STORM(patch_size=8, embed_dim=1280, depth=32, num_heads=16, **kwargs)\ndef STORM_H_16(**kwargs):\n    return STORM(patch_size=16, embed_dim=1280, depth=32, num_heads=16, **kwargs)\n\nPyTorch model.parameters(): \n\nwith \"dummy decoder\":\nSTORM-B/8:  100.60M (100,598,707)\nSTORM-L/8:  327.08M (327,081,139)\nSTORM-XL/8: 476.55M (476,547,379)\nSTORM-H/8:  665.91M (665,905,587)\nSTORM-B/16: 103.42M (103,417,907)\nSTORM-L/16: 330.82M (330,817,331)\nSTORM-H/16: 670.56M (670,558,771)\n\nwith \"conv decoder\", i.e. \"Latent-STORM\" (adds ~30M):\nSTORM-B/8:  133.23M (133,230,019)\nSTORM-L/8:  359.13M (359,125,699)\nSTORM-XL/8: 508.30M (508,298,563)\nSTORM-H/8:  697.36M (697,363,395)\nSTORM-B/16: 134.21M (134,211,523)\nSTORM-L/16: 360.43M (360,434,371)\nSTORM-H/16: 699.00M (698,999,235)"},{"id":"epoch:95144373ba5740851973:parameters","entityId":"epoch:95144373ba5740851973","metric":"parameters","value":1900000000.0,"unit":"parameters","observationPeriod":"2024-12-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.9B (Table 1)"},{"id":"epoch:4493cf1854ba081ed2c0:parameters","entityId":"epoch:4493cf1854ba081ed2c0","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2024-12-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"10B - 13B \nDiffusion Transformer (DiT)"},{"id":"epoch:f6cd6345b4711fdf31c6:parameters","entityId":"epoch:f6cd6345b4711fdf31c6","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-12-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:271f0a6b06c8bf1a7921:parameters","entityId":"epoch:271f0a6b06c8bf1a7921","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2024-12-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1b625c076569ff710e71:parameters","entityId":"epoch:1b625c076569ff710e71","metric":"parameters","value":82000000.0,"unit":"parameters","observationPeriod":"2024-12-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"82M"},{"id":"epoch:1b625c076569ff710e71:training-compute","entityId":"epoch:1b625c076569ff710e71","metric":"training-compute","value":1.6848e+20,"unit":"FLOP","observationPeriod":"2024-12-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000 FLOP / GPU / sec * 500 GPU - hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.6848e+20 FLOP"},{"id":"epoch:b1274c4a792a06e35d9e:parameters","entityId":"epoch:b1274c4a792a06e35d9e","metric":"parameters","value":671000000000.0,"unit":"parameters","observationPeriod":"2024-12-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token."},{"id":"epoch:b1274c4a792a06e35d9e:training-compute","entityId":"epoch:b1274c4a792a06e35d9e","metric":"training-compute","value":3.3e+24,"unit":"FLOP","observationPeriod":"2024-12-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 \u00d7 37B \u00d7 14.8T = 3.29e24 pretraining FLOPs, 1e22 posttrainining FLOPs - see doc below\n\nhttps://docs.google.com/document/d/1C4zBvHh4UNnzaorwQSdG61zQJsBV9X3i3A70NeelxHY/edit?usp=sharing"},{"id":"epoch:77c890306155810a96f2:parameters","entityId":"epoch:77c890306155810a96f2","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-12-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:18b6be7a23c5df3f7489:parameters","entityId":"epoch:18b6be7a23c5df3f7489","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2024-12-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3B"},{"id":"epoch:47fff2b76dcf35b9b978:parameters","entityId":"epoch:47fff2b76dcf35b9b978","metric":"parameters","value":221000000.0,"unit":"parameters","observationPeriod":"2024-12-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"221M"},{"id":"epoch:85099029b16de656dc62:parameters","entityId":"epoch:85099029b16de656dc62","metric":"parameters","value":9000000000.0,"unit":"parameters","observationPeriod":"2024-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:85099029b16de656dc62:training-compute","entityId":"epoch:85099029b16de656dc62","metric":"training-compute","value":4.484146e+23,"unit":"FLOP","observationPeriod":"2024-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Gemma 2 9B base model: 4.32e+23\nAdditional pretraining compute: \n10*24*60*60*989500000000000*64*0.3=1.64146e+22\nTotal: 1.64146\u00d710^22 + 4.32\u00d710^23 = 4.484146\u00d710^23"},{"id":"epoch:0a5c3738bc8b5ecbb05b:parameters","entityId":"epoch:0a5c3738bc8b5ecbb05b","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0a5c3738bc8b5ecbb05b:training-compute","entityId":"epoch:0a5c3738bc8b5ecbb05b","metric":"training-compute","value":1.23330162e+24,"unit":"FLOP","observationPeriod":"2024-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Llama3.1 8B base model: 1.224e+24\nAdditional pretraining compute: \n136*60*60*989500000000000*64*0.3=9.30162\u00d710^21\nTotal: 9.30162 10^21 + 1.224*^24 =1.23330162 \u00d7 10^24"},{"id":"epoch:e21fcef0b522b27af382:parameters","entityId":"epoch:e21fcef0b522b27af382","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e21fcef0b522b27af382:training-compute","entityId":"epoch:e21fcef0b522b27af382","metric":"training-compute","value":8.0103891e+24,"unit":"FLOP","observationPeriod":"2024-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Llama3.1 70B base model: 7.929e+24\nAdditional pretraining compute: \nStage 1:\n200*60*60*989500000000000*64*0.3=1.36788\u00d710^22\nStage 2: \n495*60*60*989500000000000*128*0.3=6.77103\u00d710^22\nTotal: 7.929\u00d710^24 + 1.36788\u00d710^22 + 6.77103\u00d710^22 = 8.0103891 \u00d7 10^24"},{"id":"epoch:c1abd0e370922a7568ff:parameters","entityId":"epoch:c1abd0e370922a7568ff","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:89d5d7f1f8e2d79516c7:parameters","entityId":"epoch:89d5d7f1f8e2d79516c7","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:f694d702a47960a80e47:parameters","entityId":"epoch:f694d702a47960a80e47","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:ad98d2ce1e46a8b75eb7:parameters","entityId":"epoch:ad98d2ce1e46a8b75eb7","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2024-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3B"},{"id":"epoch:ca63201b2fc68511955a:parameters","entityId":"epoch:ca63201b2fc68511955a","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2024-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1B"},{"id":"epoch:7d11e3e1c0dc981f1347:parameters","entityId":"epoch:7d11e3e1c0dc981f1347","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B (assuming same as base model)"},{"id":"epoch:cbf44b1ec61a50727b64:parameters","entityId":"epoch:cbf44b1ec61a50727b64","metric":"parameters","value":9688000000.0,"unit":"parameters","observationPeriod":"2024-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 26: \n9.688B"},{"id":"epoch:e4e7c32b37ed01b421c1:parameters","entityId":"epoch:e4e7c32b37ed01b421c1","metric":"parameters","value":2500000000.0,"unit":"parameters","observationPeriod":"2024-12-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.5B\nModel Architecture: Granite-3.1-2B-Base is based on a decoder-only dense transformer architecture. Core components of this architecture are: GQA and RoPE, MLP with SwiGLU, RMSNorm, and shared input/output embeddings."},{"id":"epoch:e4e7c32b37ed01b421c1:training-compute","entityId":"epoch:e4e7c32b37ed01b421c1","metric":"training-compute","value":1.8e+23,"unit":"FLOP","observationPeriod":"2024-12-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 2.5 * 10^9 parameters * 12*10^12 tokens = 1.8e+23 FLOP"},{"id":"epoch:35568ed51bdcaf030769:parameters","entityId":"epoch:35568ed51bdcaf030769","metric":"parameters","value":8100000000.0,"unit":"parameters","observationPeriod":"2024-12-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8.1B\nModel Architecture: Granite-3.1-8B-Base is based on a decoder-only dense transformer architecture. Core components of this architecture are: GQA and RoPE, MLP with SwiGLU, RMSNorm, and shared input/output embeddings."},{"id":"epoch:35568ed51bdcaf030769:training-compute","entityId":"epoch:35568ed51bdcaf030769","metric":"training-compute","value":5.832e+23,"unit":"FLOP","observationPeriod":"2024-12-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 FLOP / parameter / token * 8.1*10^9 parameters * 12*10^12 tokens = 5.832e+23 FLOP"},{"id":"epoch:1468900f935e4203a7c9:parameters","entityId":"epoch:1468900f935e4203a7c9","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-12-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1468900f935e4203a7c9:training-compute","entityId":"epoch:1468900f935e4203a7c9","metric":"training-compute","value":5.88e+23,"unit":"FLOP","observationPeriod":"2024-12-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 7 * 10^9 parameters * 14 * 10^12 tokens = 5.88e+23 FLOP"},{"id":"epoch:6f5343d83d745bf89a14:parameters","entityId":"epoch:6f5343d83d745bf89a14","metric":"parameters","value":7500000000.0,"unit":"parameters","observationPeriod":"2024-12-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/aisingapore/SEA-LION-v1-7B-IT"},{"id":"epoch:2e64767f08872268b6ee:parameters","entityId":"epoch:2e64767f08872268b6ee","metric":"parameters","value":335800000.0,"unit":"parameters","observationPeriod":"2024-12-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The F5-TTS base model has 22 layers, 16 attention heads, 1024/2048 embedding/feed-forward network (FFN) dimension for DiT; and 4 layers, 512/1024 embedding/FFN dimension for ConvNeXt V2; in total 335.8M parameters."},{"id":"epoch:2e64767f08872268b6ee:training-compute","entityId":"epoch:2e64767f08872268b6ee","metric":"training-compute","value":4.5287424e+20,"unit":"FLOP","observationPeriod":"2024-12-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000*8*168*3600*0.3 = 4.5287424e+20"},{"id":"epoch:9bd422515e17a375a0fb:parameters","entityId":"epoch:9bd422515e17a375a0fb","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2024-12-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We employed the Qwen2.5 (Yang et al., 2024) series of Large Language Models (LLMs) at varying scales to serve as the backbone for Apollo. Specifically, we utilized models with 1.5B, 3B, and 7B parameters. Following our analysis in Sec. 4, we used a SigLIP-SO400M (Zhai et al., 2023) encoder combined with an InternVideo2 (Wang et al., 2024d) video encoder\"\n\nthere is a confusion with Qwen2.5 link - it points to Qwen 2 paper instead of Qwen 2.5 release notes"},{"id":"epoch:196ed7bbe6d0dc88a532:parameters","entityId":"epoch:196ed7bbe6d0dc88a532","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2024-12-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We employed the Qwen2.5 (Yang et al., 2024) series of Large Language Models (LLMs) at varying scales to serve as the backbone for Apollo. Specifically, we utilized models with 1.5B, 3B, and 7B parameters. Following our analysis in Sec. 4, we used a SigLIP-SO400M (Zhai et al., 2023) encoder combined with an InternVideo2 (Wang et al., 2024d) video encoder\"\n\nthere is a confusion with Qwen2.5 link - it points to Qwen 2 paper instead of Qwen 2.5 release notes"},{"id":"epoch:67d1d2434cc4c31d8148:parameters","entityId":"epoch:67d1d2434cc4c31d8148","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-12-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We employed the Qwen2.5 (Yang et al., 2024) series of Large Language Models (LLMs) at varying scales to serve as the backbone for Apollo. Specifically, we utilized models with 1.5B, 3B, and 7B parameters. Following our analysis in Sec. 4, we used a SigLIP-SO400M (Zhai et al., 2023) encoder combined with an InternVideo2 (Wang et al., 2024d) video encoder\"\n\nthere is a confusion with Qwen2.5 link - it points to Qwen 2 paper instead of Qwen 2.5 release notes"},{"id":"epoch:2118814a7ff4ba6fbaaf:parameters","entityId":"epoch:2118814a7ff4ba6fbaaf","metric":"parameters","value":20000000000.0,"unit":"parameters","observationPeriod":"2024-12-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"20B total parameters\n3.3B active parameters"},{"id":"epoch:2118814a7ff4ba6fbaaf:training-compute","entityId":"epoch:2118814a7ff4ba6fbaaf","metric":"training-compute","value":9.9e+22,"unit":"FLOP","observationPeriod":"2024-12-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 3.3 * 10^9 active parameters * 5 * 10^12 tokens [speculatively] = 9.9e+22 FLOP"},{"id":"epoch:88f3ec4e0e3206b81136:parameters","entityId":"epoch:88f3ec4e0e3206b81136","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2024-12-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B parameters, dense decoder-only Transformer model"},{"id":"epoch:88f3ec4e0e3206b81136:training-compute","entityId":"epoch:88f3ec4e0e3206b81136","metric":"training-compute","value":9.3202015e+23,"unit":"FLOP","observationPeriod":"2024-12-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6* 14*10^9 parameters * 10*10^12 tokens = 8.4e+23 FLOP\n\n989500000000000 FLOP / sec [assumed bf16 precision] * 1920 GPUs * 504 hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.0341209e+24 FLOP\n\ngeometric mean\nsqrt(8.4e+23 * 1.0341209e+24) = 9.3202015e+23"},{"id":"epoch:892a6f5c845ebdbf9a59:parameters","entityId":"epoch:892a6f5c845ebdbf9a59","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2024-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:491c4929a73df4f896b4:parameters","entityId":"epoch:491c4929a73df4f896b4","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:58b67a2cc7ecae6df820:parameters","entityId":"epoch:58b67a2cc7ecae6df820","metric":"parameters","value":2400000000.0,"unit":"parameters","observationPeriod":"2024-12-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" 2.4B"},{"id":"epoch:58b67a2cc7ecae6df820:training-compute","entityId":"epoch:58b67a2cc7ecae6df820","metric":"training-compute","value":9.36e+22,"unit":"FLOP","observationPeriod":"2024-12-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9.36 \u00d7 10^22 (Table 2)"},{"id":"epoch:2c5f5719bd499e322c8f:parameters","entityId":"epoch:2c5f5719bd499e322c8f","metric":"parameters","value":7800000000.0,"unit":"parameters","observationPeriod":"2024-12-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.8B"},{"id":"epoch:2c5f5719bd499e322c8f:training-compute","entityId":"epoch:2c5f5719bd499e322c8f","metric":"training-compute","value":4.21e+23,"unit":"FLOP","observationPeriod":"2024-12-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4.21 \u00d7 10^23 (Table 2)"},{"id":"epoch:3f7adc3999378258ce2e:parameters","entityId":"epoch:3f7adc3999378258ce2e","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2024-12-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:3f7adc3999378258ce2e:training-compute","entityId":"epoch:3f7adc3999378258ce2e","metric":"training-compute","value":1.25e+24,"unit":"FLOP","observationPeriod":"2024-12-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.25 \u00d7 10^24 (Table 2) "},{"id":"epoch:73d6ed9180e113c7a205:parameters","entityId":"epoch:73d6ed9180e113c7a205","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2024-12-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a576808d7139ea58c68c:parameters","entityId":"epoch:a576808d7139ea58c68c","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-12-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:a576808d7139ea58c68c:training-compute","entityId":"epoch:a576808d7139ea58c68c","metric":"training-compute","value":6.8649768e+24,"unit":"FLOP","observationPeriod":"2024-12-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 FLOP / parameter / token * 70*10^9 parameters * 15*10^12 tokens = 6.3e+24 FLOP\n\n7000000 GPU-hours * 3600 sec / hour * 989500000000000 FLOP / second * 0.3 [assumed utilization]= 7.48062e+24 FLOP\n\nsqrt(7.48062e+24*6.3e+24) = 6.8649768e+24"},{"id":"epoch:188844eaebbed20bf37e:parameters","entityId":"epoch:188844eaebbed20bf37e","metric":"parameters","value":25500000000.0,"unit":"parameters","observationPeriod":"2024-12-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"25.5B"},{"id":"epoch:715a42e76a3f6311a176:parameters","entityId":"epoch:715a42e76a3f6311a176","metric":"parameters","value":38400000000.0,"unit":"parameters","observationPeriod":"2024-12-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"38.4B"},{"id":"epoch:85c140d610fb8cfa4e8c:parameters","entityId":"epoch:85c140d610fb8cfa4e8c","metric":"parameters","value":78400000000.0,"unit":"parameters","observationPeriod":"2024-12-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"78.4B"},{"id":"epoch:a6a435c7c50b177dce3e:parameters","entityId":"epoch:a6a435c7c50b177dce3e","metric":"parameters","value":15000000000.0,"unit":"parameters","observationPeriod":"2024-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"15B"},{"id":"epoch:4a06ce10cae57edd73ab:parameters","entityId":"epoch:4a06ce10cae57edd73ab","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:4a06ce10cae57edd73ab:training-compute","entityId":"epoch:4a06ce10cae57edd73ab","metric":"training-compute","value":2.2794518e+21,"unit":"FLOP","observationPeriod":"2024-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train all models using 128 NVIDIA H100 GPUs with a global batch size of 2048 across all stages.\"\n\nfrom Figure 1\nNVILA 8B takes 4.5x times less gpu hours than LLAVA OneVision which is reported to take 400 GPU days (\"For example, training a state-of-the-art 7B VLM [5] can take up to 400 GPU days\")\n\n400/4.5 ~ 89 GPU days ~ 2133 GPU hours\n\n989500000000000 FLOP / sec / GPU * 2133 GPU*hours * 3600 sec / hour * 0.3 [assumed utilization] = 2.2794518e+21 FLOP\n\nConfidence: Likely (precision is FP8 not FP16 and utilization could be different from 0.3)"},{"id":"epoch:67dfb7c0f7ac64d53723:parameters","entityId":"epoch:67dfb7c0f7ac64d53723","metric":"parameters","value":2000000000.0,"unit":"parameters","observationPeriod":"2024-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:22094d38a8580e63b407:parameters","entityId":"epoch:22094d38a8580e63b407","metric":"parameters","value":350000000.0,"unit":"parameters","observationPeriod":"2024-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"350M\n"},{"id":"epoch:22094d38a8580e63b407:training-compute","entityId":"epoch:22094d38a8580e63b407","metric":"training-compute","value":2.6788982e+21,"unit":"FLOP","observationPeriod":"2024-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 350 * 10^6 parameters * 1,086,324,736,000 tokens = 2.2812819e+21 FLOP\n\n989400000000000 FLOP / GPU / sec [bf16 assumed] * 46 HOURS * 3600 sec / hour * 64 GPUs * 0.3 [assumed utilization] = 3.1458171e+21 FLOP\n\nsqrt(2.2812819e+21*3.1458171e+21) = 2.6788982e+21 "},{"id":"epoch:29cf8710f9c1a9a1ab86:parameters","entityId":"epoch:29cf8710f9c1a9a1ab86","metric":"parameters","value":1200000000.0,"unit":"parameters","observationPeriod":"2024-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.2B"},{"id":"epoch:29cf8710f9c1a9a1ab86:training-compute","entityId":"epoch:29cf8710f9c1a9a1ab86","metric":"training-compute","value":2.9770787e+22,"unit":"FLOP","observationPeriod":"2024-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 1.2 * 10^9 parameters * 5 * 10^12 tokens = 3.6e+22 FLOP\n\n989400000000000 FLOP / GPU / sec [bf16 assumed] * 192 GPUs * 5 days * 24 hour / day * 3600 sec / hour * 0.3 [assumed utilization] = 2.4619438e+22 FLOP\n\nsqrt(3.6e+22*2.4619438e+22) = 2.9770787e+22 "},{"id":"epoch:6023fa03f0df6c839d2f:parameters","entityId":"epoch:6023fa03f0df6c839d2f","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2024-12-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B"},{"id":"epoch:5f78e1c1fe393cc43dfc:training-compute","entityId":"epoch:5f78e1c1fe393cc43dfc","metric":"training-compute","value":6.000010000000001e+24,"unit":"FLOP","observationPeriod":"2024-12-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"probably just below 1e25 stemming from the Llama 70B serving speed.  If Llama 70B is trained proportionally to 405B, then it's at ~ 6.6e24. Nova Pro is served at 100tk/s, while Llama 70B is served at 70tk/s on average, and 100tk/s by together.ai at FP8. So Nova Pro would be >1e25 if they roughly 2x the amount of training compared to Llama 70B which [seems unlikely]\""},{"id":"epoch:c08bcf050804e0641bfa:parameters","entityId":"epoch:c08bcf050804e0641bfa","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-12-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13b"},{"id":"epoch:c08bcf050804e0641bfa:training-compute","entityId":"epoch:c08bcf050804e0641bfa","metric":"training-compute","value":1.4814815e+23,"unit":"FLOP","observationPeriod":"2024-12-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from Figure 10:\n\nthe optimal model has 13b parameters, 5.8e+07PF (image training) + 7.0e+07PF (video training) of compute and 740B (image tokens) + 928B (video tokens) \n\n5.8e+07PF + 7.0e+07PF = 12.8e+07PF = 12.8*10^7*10^20/(24*3600) = 1.4814815e+23 FLOPs\n\n6ND = 6*13*10^9*(740+928)*10^9 = 1.30104e+23\n"},{"id":"epoch:10cadffe32e764bced53:parameters","entityId":"epoch:10cadffe32e764bced53","metric":"parameters","value":32000000.0,"unit":"parameters","observationPeriod":"2024-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1ed4cb58fe3da0865b28:parameters","entityId":"epoch:1ed4cb58fe3da0865b28","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2024-11-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1ed4cb58fe3da0865b28:training-compute","entityId":"epoch:1ed4cb58fe3da0865b28","metric":"training-compute","value":6.000001e+22,"unit":"FLOP","observationPeriod":"2024-11-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"10B parameters trained on 1T tokens: 6 * 10B * 1T = 6e22"},{"id":"epoch:480164c44cf2a38c9e37:parameters","entityId":"epoch:480164c44cf2a38c9e37","metric":"parameters","value":102000000000.0,"unit":"parameters","observationPeriod":"2024-11-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B-102B"},{"id":"epoch:c26784ed4e13649c7b2d:parameters","entityId":"epoch:c26784ed4e13649c7b2d","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:74d3cf44b367eb16dddb:parameters","entityId":"epoch:74d3cf44b367eb16dddb","metric":"parameters","value":28900000000.0,"unit":"parameters","observationPeriod":"2024-11-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:219d2ca06d0c472234ba:parameters","entityId":"epoch:219d2ca06d0c472234ba","metric":"parameters","value":2500000000.0,"unit":"parameters","observationPeriod":"2024-11-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f18ffa487388fbf3c500:parameters","entityId":"epoch:f18ffa487388fbf3c500","metric":"parameters","value":150000.0,"unit":"parameters","observationPeriod":"2024-11-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ea4709661d9456834619:training-compute","entityId":"epoch:ea4709661d9456834619","metric":"training-compute","value":2.6599104e+21,"unit":"FLOP","observationPeriod":"2024-11-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Previous model Fish-Speech 1.4 used 1.9151355e+21 FLOP for training (see model card) and was trained on 720000 hours of audio, while this model is similar but trained on 1M hours of audio -> its training compute couldbe estimated as \n\n1.9151355e+21 * 10^6 / 720000 = 2.6599104e+21 FLOP"},{"id":"epoch:db4a618bc1509ae4074e:parameters","entityId":"epoch:db4a618bc1509ae4074e","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2024-11-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.5B"},{"id":"epoch:db4a618bc1509ae4074e:training-compute","entityId":"epoch:db4a618bc1509ae4074e","metric":"training-compute","value":1.35e+22,"unit":"FLOP","observationPeriod":"2024-11-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 1.5 * 10^9 parameters * 1.5 * 10^12 tokens = 1.35e+22 FLOP"},{"id":"epoch:676775761b501b6ac411:parameters","entityId":"epoch:676775761b501b6ac411","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-11-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:37122fb7df6f1b0d4340:parameters","entityId":"epoch:37122fb7df6f1b0d4340","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-11-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:61e691d146c9d9e6672d:parameters","entityId":"epoch:61e691d146c9d9e6672d","metric":"parameters","value":124000000000.0,"unit":"parameters","observationPeriod":"2024-11-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"123B multimodal decoder, 1B parameter vision encoder"},{"id":"epoch:faa5c9034c96b716e6c1:parameters","entityId":"epoch:faa5c9034c96b716e6c1","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-11-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:faa5c9034c96b716e6c1:training-compute","entityId":"epoch:faa5c9034c96b716e6c1","metric":"training-compute","value":4.242e+23,"unit":"FLOP","observationPeriod":"2024-11-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 7000000000 parameters * 10100000000000 tokens = 4.242e+23 FLOP"},{"id":"epoch:dff310a7d443a0f55107:parameters","entityId":"epoch:dff310a7d443a0f55107","metric":"parameters","value":117000000.0,"unit":"parameters","observationPeriod":"2024-11-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"a 117M parameter Transformer encoder\""},{"id":"epoch:dff310a7d443a0f55107:training-compute","entityId":"epoch:dff310a7d443a0f55107","metric":"training-compute","value":1.8354513e+19,"unit":"FLOP","observationPeriod":"2024-11-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 117 * 10^6 parameters * 32254000000 tokens = 2.2642308e+19 FLOP\n\n35580000000000 FLOP / GPU / sec * 8 GPUs * 48.4 hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.4878702e+19 FLOP\n\nsqrt(2.2642308e+19 * 1.4878702e+19) = 1.8354513e+19 FLOP"},{"id":"epoch:d8534b0170dda0123308:parameters","entityId":"epoch:d8534b0170dda0123308","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2024-11-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9eecb64d925785ce8495:parameters","entityId":"epoch:9eecb64d925785ce8495","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2024-11-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"11b"},{"id":"epoch:4d7838b4ce2377a3f4c9:parameters","entityId":"epoch:4d7838b4ce2377a3f4c9","metric":"parameters","value":123000000000.0,"unit":"parameters","observationPeriod":"2024-11-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5bc1c73e0419286a6f33:parameters","entityId":"epoch:5bc1c73e0419286a6f33","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2024-11-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"72B"},{"id":"epoch:9522d2c83190da764d4c:parameters","entityId":"epoch:9522d2c83190da764d4c","metric":"parameters","value":9000000000.0,"unit":"parameters","observationPeriod":"2024-11-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9B"},{"id":"epoch:7069d81f42dc67c6be9a:parameters","entityId":"epoch:7069d81f42dc67c6be9a","metric":"parameters","value":32500000000.0,"unit":"parameters","observationPeriod":"2024-11-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32.5B (31B - non emb)"},{"id":"epoch:7069d81f42dc67c6be9a:training-compute","entityId":"epoch:7069d81f42dc67c6be9a","metric":"training-compute","value":1.0725e+24,"unit":"FLOP","observationPeriod":"2024-11-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming 1 epoch\n\n6ND = 6*32.5 parameters *10^9*5.5*10^12 tokens = 1.0725e+24"},{"id":"epoch:97baaf3e3ff4b875a98d:parameters","entityId":"epoch:97baaf3e3ff4b875a98d","metric":"parameters","value":665000000.0,"unit":"parameters","observationPeriod":"2024-11-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/EarthSpeciesProject/NatureLM-audio\n\n\"we use an LLM to produce text, in this case Llama 3.1-\n8b (Dubey et al., 2024a), which is fine-tuned with LoRA (Hu et al., 2022). During training, only the\nadapter layers of the LLM are updated, while the base LLM parameters remain frozen. I\""},{"id":"epoch:97baaf3e3ff4b875a98d:training-compute","entityId":"epoch:97baaf3e3ff4b875a98d","metric":"training-compute","value":1.4108774e+21,"unit":"FLOP","observationPeriod":"2024-11-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"756000000000000 FLOP / GPU / sec [H100 reported, bf16 assumed] * 8 GPUs * 216 hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.4108774e+21 FLOP"},{"id":"epoch:61a4e3fd46b01633f409:training-compute","entityId":"epoch:61a4e3fd46b01633f409","metric":"training-compute","value":1.9151355e+21,"unit":"FLOP","observationPeriod":"2024-11-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The AR training utilized an 8*H100 80G GPUs for one week, while the vocoder training employed an 8*4090 GPUs for an additional week. Note that these timelines exclude the DPO stage.\"\n\n(989400000000000 FLOP / GPU / sec [H100, bf16 assumed] + 330000000000000 FLOP / GPU / sec [4090, bf16 assumed]) * 8 GPUs * 168 hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.9151355e+21 FLOP"},{"id":"epoch:a17c330b4b94a91d9fad:parameters","entityId":"epoch:a17c330b4b94a91d9fad","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-11-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:70b1640349e6f01ab7f2:parameters","entityId":"epoch:70b1640349e6f01ab7f2","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2024-11-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:021bb7b674daff433e4b:parameters","entityId":"epoch:021bb7b674daff433e4b","metric":"parameters","value":389000000000.0,"unit":"parameters","observationPeriod":"2024-11-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"a total of 389 billion parameters and 52 billion activation parameters\""},{"id":"epoch:021bb7b674daff433e4b:training-compute","entityId":"epoch:021bb7b674daff433e4b","metric":"training-compute","value":3.49237e+24,"unit":"FLOP","observationPeriod":"2024-11-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"52B activated parameters\n\n6ND = 6*52*10^9*7*10^12 = 2.184 \u00d7 10^24\n\nThey also suggest more precise formula to calculate MoE compute budget:\n\n9.59ND + 2.3 \u00d7 10^8D = 9.59*52*10^9*7*10^12 + 2.3 \u00d7 10^8 \u00d7  7*10^12 = 3.49237\u00d710^24\n\nwhich seems closer to projected compute on Figure 3"},{"id":"epoch:b595ce3f0bfc850a08da:parameters","entityId":"epoch:b595ce3f0bfc850a08da","metric":"parameters","value":178045568.0,"unit":"parameters","observationPeriod":"2024-11-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Checked using HuggingFace Library API."},{"id":"epoch:b595ce3f0bfc850a08da:training-compute","entityId":"epoch:b595ce3f0bfc850a08da","metric":"training-compute","value":3.18e+19,"unit":"FLOP","observationPeriod":"2024-11-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training used Nvidia H100 Hopper nodes for 400 hours (https://huggingface.co/recursionpharma/OpenPhenom#training-evaluation-and-testing-details). Nvidia H100 Hopper has peak FLOPS of 67 teraFLOPS at single-precision (https://en.wikipedia.org/wiki/Hopper_(microarchitecture)#H100_accelerator_and_DGX_H100).\n\nAssuming 33% utilization rate and the use 1 Nvidia H100 Hopper,\nTraining compute\n= 0.33 * 6.7e13 FLOPS * 400 h * 3600 s / h\n~= 3.18e19 FLOPS"},{"id":"epoch:63bdd772210fb9599373:parameters","entityId":"epoch:63bdd772210fb9599373","metric":"parameters","value":4000000000.0,"unit":"parameters","observationPeriod":"2024-11-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4b\n\nLlama-3.1-Minitron-4B-Width-Base uses a model embedding size of 3072, 32 attention heads, MLP intermediate dimension of 9216, with 32 layers in total. Additionally, it uses Grouped-Query Attention (GQA) and Rotary Position Embeddings (RoPE).\n\nArchitecture Type: Transformer Decoder (Auto-Regressive Language Model)\n\nNetwork Architecture: Llama-3.1"},{"id":"epoch:5e3d6fb7346d46c9d77e:parameters","entityId":"epoch:5e3d6fb7346d46c9d77e","metric":"parameters","value":8300000000.0,"unit":"parameters","observationPeriod":"2024-11-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8.3B (table 2)\n\n(including  6.2B non-embedding parameters)\n\nMinitron-8B-Base is a large language model (LLM) obtained by pruning Nemotron-4 15B; specifically, we prune model embedding size, number of attention heads, and MLP intermediate dimension.\n\nMinitron-8B-Base uses a model embedding size of 4096, 48 attention heads, and an MLP intermediate dimension of 16384. It also uses Grouped-Query Attention (GQA) and Rotary Position Embeddings (RoPE)."},{"id":"epoch:1f416b5c73ad0a85a93d:parameters","entityId":"epoch:1f416b5c73ad0a85a93d","metric":"parameters","value":4200000000.0,"unit":"parameters","observationPeriod":"2024-11-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" 4.2B (table 4)\n\nincluding 2.6B non-embedding parameters\n\n\"Minitron-4B-Base uses a model embedding size of 3072, 32 attention heads, and an MLP intermediate dimension of 9216. It also uses Grouped-Query Attention (GQA) and Rotary Position Embeddings (RoPE).\""},{"id":"epoch:45581e3e6357f98fe35c:parameters","entityId":"epoch:45581e3e6357f98fe35c","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ec2f41541e71fafd3fab:parameters","entityId":"epoch:ec2f41541e71fafd3fab","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2024-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:53fa4fd8018ecca8ade7:parameters","entityId":"epoch:53fa4fd8018ecca8ade7","metric":"parameters","value":9209788.0,"unit":"parameters","observationPeriod":"2024-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In total, the model contains about 9,209,788 trainable parameters.\""},{"id":"epoch:4eeacfad0b4463f93466:parameters","entityId":"epoch:4eeacfad0b4463f93466","metric":"parameters","value":8800000000.0,"unit":"parameters","observationPeriod":"2024-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B (LLama) + 1843000000 (ViT) = 8.8B"},{"id":"epoch:4eeacfad0b4463f93466:training-compute","entityId":"epoch:4eeacfad0b4463f93466","metric":"training-compute","value":1.425e+23,"unit":"FLOP","observationPeriod":"2024-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Finetune: 10*60*60*312000000000000*0.5=5.616e+18\n\nBase models: 8.4e+22, 5.85e+22\nTotal: 1.4250562e+23"},{"id":"epoch:1981cf103fed71b30e92:parameters","entityId":"epoch:1981cf103fed71b30e92","metric":"parameters","value":9000000000.0,"unit":"parameters","observationPeriod":"2024-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:802148bbb91d6f4b8d43:parameters","entityId":"epoch:802148bbb91d6f4b8d43","metric":"parameters","value":3300000000.0,"unit":"parameters","observationPeriod":"2024-10-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"While in principle our model can be initialized from scratch or fine-tuned from any VLM backbone, in practice we use PaliGemma [5] as our base model. PaliGemma is an opensource 3 billion parameter VLM that offers a convenient tradeoff between size and performance. We add 300M parameters for the action expert (which is initialized from scratch) for a total of 3.3 billion parameters.\""},{"id":"epoch:d84a0a5bf6f7ba81f2e4:parameters","entityId":"epoch:d84a0a5bf6f7ba81f2e4","metric":"parameters","value":229000000.0,"unit":"parameters","observationPeriod":"2024-10-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The parameter\ncounts of our 3D-aided face latent model and diffusion transformer model are about 200M and 29M respectively\""},{"id":"epoch:d84a0a5bf6f7ba81f2e4:training-compute","entityId":"epoch:d84a0a5bf6f7ba81f2e4","metric":"training-compute","value":4.012416e+19,"unit":"FLOP","observationPeriod":"2024-10-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"38700000000000 FLOP/GPU/sec * 240 hours * 3600 sec / hour * 4 GPUs * 0.3 [assumed utilization] = 40124160000000000000 FLOP"},{"id":"epoch:46a4adbe88314e3e8965:parameters","entityId":"epoch:46a4adbe88314e3e8965","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-10-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Currently, AstroOne, boasting 70 billion parameters and 32 billion tokens of astronomical text corpus, has been evaluated by 10,000 astronomy assessment datasets.\""},{"id":"epoch:64d00eb0dbae400ca4bf:parameters","entityId":"epoch:64d00eb0dbae400ca4bf","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-10-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d96cafffb11948d4a67c:parameters","entityId":"epoch:d96cafffb11948d4a67c","metric":"parameters","value":600000000.0,"unit":"parameters","observationPeriod":"2024-10-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. ASR decoder: \"Our models are based on an ASR decoder architecture called the Recurrent Neural Network Transducer (RNN-T), which offers advantages in scalability, robustness against hallucinations, and timestamp accuracy\u2014key factors for real-world usage at scale. We use a 600M parameter Conformer RNN-T model.\"\n\n2. Text Formatting module."},{"id":"epoch:9e089659630a6f1cb45f:parameters","entityId":"epoch:9e089659630a6f1cb45f","metric":"parameters","value":2500000000.0,"unit":"parameters","observationPeriod":"2024-10-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.5B"},{"id":"epoch:22dbc3b6f33c7d180986:parameters","entityId":"epoch:22dbc3b6f33c7d180986","metric":"parameters","value":650000000.0,"unit":"parameters","observationPeriod":"2024-10-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:22dbc3b6f33c7d180986:training-compute","entityId":"epoch:22dbc3b6f33c7d180986","metric":"training-compute","value":8.18e+20,"unit":"FLOP","observationPeriod":"2024-10-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Tokens per step; 4096*8*32=1048576\nTotal training tokens 1048576*200000=209715200000\nEpochs: 209715200000 / 28800000001 = 7.28\nFLOP: 6*650000000*209715200000=8.1788928e+20"},{"id":"epoch:767ff8e835759bfb567d:parameters","entityId":"epoch:767ff8e835759bfb567d","metric":"parameters","value":500000000000.0,"unit":"parameters","observationPeriod":"2024-10-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"[Speculative] Doubao's large language model has scaled up from 35 billion parameters to 800 billion, with 500 billion and 800 billion parameter models currently under training.\nhttps://xueqiu.com/9637001584/309910396?md5__1038=7qmx2DyDuie4cDBqDTQEWqDtMvO4iTphD\n"},{"id":"epoch:767ff8e835759bfb567d:training-compute","entityId":"epoch:767ff8e835759bfb567d","metric":"training-compute","value":2.505e+25,"unit":"FLOP","observationPeriod":"2024-10-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 * 500*10^9 * 8350*10^9 = 2.505e+25"},{"id":"epoch:6aab3905b5ab2f4e44b2:parameters","entityId":"epoch:6aab3905b5ab2f4e44b2","metric":"parameters","value":32300000000.0,"unit":"parameters","observationPeriod":"2024-10-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6aab3905b5ab2f4e44b2:training-compute","entityId":"epoch:6aab3905b5ab2f4e44b2","metric":"training-compute","value":6.688684e+21,"unit":"FLOP","observationPeriod":"2024-10-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"'Likely' confidence for dataset size and epochs\n\n34513333333*32300000000*6 = 6.688684e+21"},{"id":"epoch:adac81a5bfe6e362d7e4:parameters","entityId":"epoch:adac81a5bfe6e362d7e4","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-10-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:adac81a5bfe6e362d7e4:training-compute","entityId":"epoch:adac81a5bfe6e362d7e4","metric":"training-compute","value":1.65664e+21,"unit":"FLOP","observationPeriod":"2024-10-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"'Likely' confidence for dataset size and epochs\n\n34513333333*8000000000*6 = 1.65664e+21"},{"id":"epoch:b9388e9864228922adb7:parameters","entityId":"epoch:b9388e9864228922adb7","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-10-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:84f9563c70594f079031:parameters","entityId":"epoch:84f9563c70594f079031","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2024-10-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"featuring a 10 billion parameter diffusion model built on our novel Asymmetric Diffusion Transformer (AsymmDiT) architecture"},{"id":"epoch:30b3f6adfca42eeb11b3:parameters","entityId":"epoch:30b3f6adfca42eeb11b3","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2024-10-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"72B"},{"id":"epoch:30b3f6adfca42eeb11b3:training-compute","entityId":"epoch:30b3f6adfca42eeb11b3","metric":"training-compute","value":3.02e+24,"unit":"FLOP","observationPeriod":"2024-10-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Uses Qwen2-72B as a backbone, which trained with 3.02e24 FLOP, as well as InternViT-6B. It's unclear how many FLOP were spent training but probably negligible; e.g. PaLI trained ViT-e with ~4B parameters using 1.07e23 FLOP.\n\nFine-tuning FLOPs:\n57,016,320,000 image/text tokens over all stages\n6 * 72B * 57,016,320,000 = 2.463e22\n"},{"id":"epoch:e8f8c4c0afc39685ea73:parameters","entityId":"epoch:e8f8c4c0afc39685ea73","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2024-10-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"72B"},{"id":"epoch:e8f8c4c0afc39685ea73:training-compute","entityId":"epoch:e8f8c4c0afc39685ea73","metric":"training-compute","value":3.02e+24,"unit":"FLOP","observationPeriod":"2024-10-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Additional compute in this paper is negligible relative to the compute used to train the language model backbone (Qwen2-72B at 3.02e24 FLOP)"},{"id":"epoch:121504cf1ac6efb5dcdc:parameters","entityId":"epoch:121504cf1ac6efb5dcdc","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2024-10-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"72B"},{"id":"epoch:121504cf1ac6efb5dcdc:training-compute","entityId":"epoch:121504cf1ac6efb5dcdc","metric":"training-compute","value":3.0398181e+24,"unit":"FLOP","observationPeriod":"2024-10-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.02e24 FLOP (Qwen2-72B compute) + 19818086000000000000000 = 3.0398181e+24"},{"id":"epoch:3f6b67189f72212eb23b:parameters","entityId":"epoch:3f6b67189f72212eb23b","metric":"parameters","value":8100000000.0,"unit":"parameters","observationPeriod":"2024-10-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8.1B"},{"id":"epoch:d4969beb5baa86735ffa:parameters","entityId":"epoch:d4969beb5baa86735ffa","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-10-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e87d57971f0c0bf31846:parameters","entityId":"epoch:e87d57971f0c0bf31846","metric":"parameters","value":8100000000.0,"unit":"parameters","observationPeriod":"2024-10-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8.1B"},{"id":"epoch:e87d57971f0c0bf31846:training-compute","entityId":"epoch:e87d57971f0c0bf31846","metric":"training-compute","value":5.832e+23,"unit":"FLOP","observationPeriod":"2024-10-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 FLOP / parameter / token * 8.1*10^9 parameters * 12*10^12 tokens = 5.832e+23 FLOP\n\n\"All our Granite 3.0 models are trained using a\ncompute budget of 8.35 \u00d7 10^23 FLOPS.\"\n\n8.35 \u00d7 10^23 * 757.0 (model's power consumption) / (174.6+757.0+64.5+121.2) = 5.6573436e+23\n\nhardware estimation:\n832102 GPU-hours * 3600 sec / hour * 989500000000000 FLOP / GPU / sec * 0.3 [assumed utilization] = 8.8923412e+23"},{"id":"epoch:ee0792b0bfc49501447e:parameters","entityId":"epoch:ee0792b0bfc49501447e","metric":"parameters","value":2500000000.0,"unit":"parameters","observationPeriod":"2024-10-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.5B"},{"id":"epoch:ee0792b0bfc49501447e:training-compute","entityId":"epoch:ee0792b0bfc49501447e","metric":"training-compute","value":1.8e+23,"unit":"FLOP","observationPeriod":"2024-10-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 FLOP / token / parameter * 2.5*10^9 parameters * 12*10^12 tokens = 1.8e+23 FLOP\n\n\"\"All our Granite 3.0 models are trained using a\ncompute budget of 8.35 \u00d7 10^23 FLOPS.\"\n\n8.35 \u00d7 10^23 * 174.6 (model's power consumption) / (174.6+757.0+64.5+121.2) =1.304851e+23\n\nhardware estimation:\n192030 GPU-hours * 3600 sec / hour *989500000000000 FLOP / GPU / sec * 0.3 [assumed utilization] = 2.0521478e+23 FLOP"},{"id":"epoch:65a18ab77d2a505e76e7:parameters","entityId":"epoch:65a18ab77d2a505e76e7","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-10-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d40c8de53dd4267c9d0a:parameters","entityId":"epoch:d40c8de53dd4267c9d0a","metric":"parameters","value":2975000000.0,"unit":"parameters","observationPeriod":"2024-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"VAE: 175M\nDiT: 2.8B"},{"id":"epoch:d40c8de53dd4267c9d0a:training-compute","entityId":"epoch:d40c8de53dd4267c9d0a","metric":"training-compute","value":6.565847e+22,"unit":"FLOP","observationPeriod":"2024-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming (!) 10 days of training (we might make a better assumption if we know what is the average throughput steps/sec for such setups because steps are given):\n989500000000000*10*24*3600*256*0.3 = 6.565847e+22"},{"id":"epoch:0b5887c2434b1c9eb52e:parameters","entityId":"epoch:0b5887c2434b1c9eb52e","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2024-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"giant model -1.3B parameters"},{"id":"epoch:c5d53cc5a4e1185155f7:parameters","entityId":"epoch:c5d53cc5a4e1185155f7","metric":"parameters","value":335300000.0,"unit":"parameters","observationPeriod":"2024-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"335.3M"},{"id":"epoch:8695c5b7f3a4c019f776:parameters","entityId":"epoch:8695c5b7f3a4c019f776","metric":"parameters","value":35000000000.0,"unit":"parameters","observationPeriod":"2024-10-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:16fe2ddf46f2d624d8b4:parameters","entityId":"epoch:16fe2ddf46f2d624d8b4","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2024-10-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.3B\n! In the paper they say that they \"utilize DeepSeek-LLM (1.3B) [5] with a maximum supported sequence length of 4096 as the base language model.\"\nbut I haven't found any mentioning of 1.3B model in that Deepseek paper (only 7B and 67B)"},{"id":"epoch:2b508c6a6a233a5fad06:training-compute","entityId":"epoch:2b508c6a6a233a5fad06","metric":"training-compute","value":1.5e+24,"unit":"FLOP","observationPeriod":"2024-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The CEO of 01.AI tweeted that Yi-Lightning was trained for 1 month on 2000 H100s: https://x.com/kaifulee/status/1846310645849047524\nAssuming this is accurate:\n(9.9e14 * 2000) FLOP/s * 1 month * 30.5 days/month * 24hr/day * 3600 s/hr * 0.3 utilization assumption = 1.565e24"},{"id":"epoch:feeaed89739957983225:parameters","entityId":"epoch:feeaed89739957983225","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2024-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a4ca199ff9751a268c7e:parameters","entityId":"epoch:a4ca199ff9751a268c7e","metric":"parameters","value":8019808256.0,"unit":"parameters","observationPeriod":"2024-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Architecture\tDense Transformer\nParameters\t8,019,808,256\nLayers\t36\nHeads\t32"},{"id":"epoch:da040096546e697c90b4:parameters","entityId":"epoch:da040096546e697c90b4","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:7e641ea755d5f3ac0f55:training-compute","entityId":"epoch:7e641ea755d5f3ac0f55","metric":"training-compute","value":7.7605724e+21,"unit":"FLOP","observationPeriod":"2024-10-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From paper: 128 A100s for 30 days; assumptions: 30% utilization rate, FP16 precision"},{"id":"epoch:bc48e724937cc9071e19:parameters","entityId":"epoch:bc48e724937cc9071e19","metric":"parameters","value":1648000000.0,"unit":"parameters","observationPeriod":"2024-10-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1648M parameters"},{"id":"epoch:9ab2ab8008be31cd8437:parameters","entityId":"epoch:9ab2ab8008be31cd8437","metric":"parameters","value":304000000.0,"unit":"parameters","observationPeriod":"2024-10-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"304M"},{"id":"epoch:cd23abd0386ae3517f33:training-compute","entityId":"epoch:cd23abd0386ae3517f33","metric":"training-compute","value":7.929e+24,"unit":"FLOP","observationPeriod":"2024-10-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Taken from Llama 3.1 70B as the finetuning compute is multiple orders of magnitude lower"},{"id":"epoch:84594d7190ef1f9de633:parameters","entityId":"epoch:84594d7190ef1f9de633","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-10-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:eef844cd3d048348e8ec:parameters","entityId":"epoch:eef844cd3d048348e8ec","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2024-10-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We utilized xTrimoPGLM (27), a state-of-the-art\nprotein language model, as the protein encoder, and Vicuna-\n13B (25), fine-tuned from Llama-2 (21), as the LLM of\nProteinChat.\""},{"id":"epoch:03aaf29eb1203c80df06:parameters","entityId":"epoch:03aaf29eb1203c80df06","metric":"parameters","value":1200000000.0,"unit":"parameters","observationPeriod":"2024-10-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Model Training and Inference: \"We scale the size of RDT up to 1.2B parameters, establishing it as the currently largest diffusion-based robotic foundation model.\""},{"id":"epoch:03aaf29eb1203c80df06:training-compute","entityId":"epoch:03aaf29eb1203c80df06","metric":"training-compute","value":4.06e+22,"unit":"FLOP","observationPeriod":"2024-10-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Model Training and Inference: \"The model is pre-trained on 48 H100 80GB GPUs for a month, giving a total of 1M training iteration steps. It takes three days to fine-tune this model using the same GPUs for 130K steps.\"\nTable 10: \"Mixed Precision, bf16\"\nAssume 48xGPU, \"scheduling reasons\"  -> NVIDIA H100 SXM5 -> 9.894e14 FLOP/s/GPU\nAssume 0.3 utilization\nAssume 1 month + 3 days = 33 days\n0.3 * 48 GPU * (33 days) * 9.894e14 FLOP/s/GPU ~=  4.02e22 FLOP\n\n"},{"id":"epoch:7ee0ab4c7a16786eb9ab:parameters","entityId":"epoch:7ee0ab4c7a16786eb9ab","metric":"parameters","value":764000000.0,"unit":"parameters","observationPeriod":"2024-10-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e02f5ebced6e92584668:parameters","entityId":"epoch:e02f5ebced6e92584668","metric":"parameters","value":150000000000.0,"unit":"parameters","observationPeriod":"2024-10-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Source: https://venturebeat.com/ai/writers-palmyra-x-004-takes-the-lead-in-ai-function-calling-surpassing-tech-giants/"},{"id":"epoch:bc17b48a65f2e3c4948f:parameters","entityId":"epoch:bc17b48a65f2e3c4948f","metric":"parameters","value":5000000000.0,"unit":"parameters","observationPeriod":"2024-10-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"5 billion"},{"id":"epoch:c773b14a44dd7f94dac5:parameters","entityId":"epoch:c773b14a44dd7f94dac5","metric":"parameters","value":230000000.0,"unit":"parameters","observationPeriod":"2024-10-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"the default GR-2 model contains 230M parameters, of which 95M are trainable"},{"id":"epoch:f083ba4c75f81bbe3a25:parameters","entityId":"epoch:f083ba4c75f81bbe3a25","metric":"parameters","value":2000000000.0,"unit":"parameters","observationPeriod":"2024-10-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We utilize the prevailing MM-DiT architecture from SD3 Medium (Esser\net al., 2024) as the base model, with 2B parameters in total\" (https://arxiv.org/pdf/2410.05954, page 7)."},{"id":"epoch:f083ba4c75f81bbe3a25:training-compute","entityId":"epoch:f083ba4c75f81bbe3a25","metric":"training-compute","value":7.67e+21,"unit":"FLOP","observationPeriod":"2024-10-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\u201cIt is trained only on open-source datasets within 20.7k A100 GPU hours\u201d (https://huggingface.co/spaces/Pyramid-Flow/pyramid-flow), (https://arxiv.org/pdf/2410.05954, page 1). The training was done in BF16 (https://arxiv.org/pdf/2410.05954, page 18). \n\nBoth A100 models deliver peak 312 teraFLOPS without sparsity (https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/a100/pdf/a100-80gb-datasheet-update-nvidia-us-1521051-r2-web.pdf).\n\nAssuming 33% utilization rate,\nTraining compute\n= 0.33 * 3.12e14 FLOPS * 2.07e4 hours * 3.6e3 s / hour\n~= 7.67e21 FLOPS"},{"id":"epoch:ab5457d4c51e348c30b0:parameters","entityId":"epoch:ab5457d4c51e348c30b0","metric":"parameters","value":24900000000.0,"unit":"parameters","observationPeriod":"2024-10-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"ARIA MoE activates 3.5B parameters per text token and has a total of 24.9B parameter"},{"id":"epoch:ab5457d4c51e348c30b0:training-compute","entityId":"epoch:ab5457d4c51e348c30b0","metric":"training-compute","value":1.428e+23,"unit":"FLOP","observationPeriod":"2024-10-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 3,5 * 10^9 active parameters * 6,8 * 10^12 tokens = 1.428e+23 FLOP"},{"id":"epoch:1ddba47d7e2ddcdb16cc:parameters","entityId":"epoch:1ddba47d7e2ddcdb16cc","metric":"parameters","value":44000000.0,"unit":"parameters","observationPeriod":"2024-10-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a20a5709c5cc6d332996:training-compute","entityId":"epoch:a20a5709c5cc6d332996","metric":"training-compute","value":8.6e+18,"unit":"FLOP","observationPeriod":"2024-10-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.5*24*60*60*2*35580000000000*0.4=8.6075136e+18"},{"id":"epoch:e0f337a0879b33e25861:parameters","entityId":"epoch:e0f337a0879b33e25861","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2024-10-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"30B"},{"id":"epoch:e0f337a0879b33e25861:training-compute","entityId":"epoch:e0f337a0879b33e25861","metric":"training-compute","value":1.65e+24,"unit":"FLOP","observationPeriod":"2024-10-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Model size = 30B\nBroken down by training stage (table 3):\n256px T2I: samples seen = 1.94E9; sample token length = 256; flops = 6ND = 8.94E22\n256px T2I/V: samples seen = 3.95E8; sample token length = 8192; flops = 6ND = 5.82E23\n768px T2I/V: samples seen = 7.38E7; sample token length = 73,728; flops = 6ND = 9.79E23\nTotal flops = 1.65E24"},{"id":"epoch:5c00ca23c432627ccdc3:parameters","entityId":"epoch:5c00ca23c432627ccdc3","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-10-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:5c00ca23c432627ccdc3:training-compute","entityId":"epoch:5c00ca23c432627ccdc3","metric":"training-compute","value":1.4e+23,"unit":"FLOP","observationPeriod":"2024-10-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Pre trained for 14 days on 384 H100s. I assumed a 0.3 utilization rate"},{"id":"epoch:b85c0f9fec5531865859:parameters","entityId":"epoch:b85c0f9fec5531865859","metric":"parameters","value":110000000.0,"unit":"parameters","observationPeriod":"2024-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The model consists of 12 230 layers with a dimension of 512, with 8 attention heads and a total of 110M parameters\""},{"id":"epoch:e1d1c7a0f5baac1e29ef:parameters","entityId":"epoch:e1d1c7a0f5baac1e29ef","metric":"parameters","value":40300000000.0,"unit":"parameters","observationPeriod":"2024-09-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"A 40.3B Mixture of Experts (MoE) model, designed for tackling more complex tasks.\"\n\n\"LFM-40B offers a new balance between model size and output quality. It leverages 12B activated parameters at use.\""},{"id":"epoch:2059dff13ae79eb8dd8e:parameters","entityId":"epoch:2059dff13ae79eb8dd8e","metric":"parameters","value":779000000.0,"unit":"parameters","observationPeriod":"2024-09-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0054b0d37aa217b76ccc:training-compute","entityId":"epoch:0054b0d37aa217b76ccc","metric":"training-compute","value":1.6000000000000008e+19,"unit":"FLOP","observationPeriod":"2024-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware: 1x NVIDIA A100 PCIe GPU (3.12e14 FP16 Tensor FLOP/s)\n2. Training duration: 36 hours (directly provided) = 129,600 seconds\n3. Utilization: 40% (assumed)\n4. Calculation: 3.12e14 FLOP/s \u00d7 1 GPU \u00d7 129,600s \u00d7 0.4 = 1.6e19 FLOPs"},{"id":"epoch:4d1f0afac4982d7a84e4:parameters","entityId":"epoch:4d1f0afac4982d7a84e4","metric":"parameters","value":1200000000.0,"unit":"parameters","observationPeriod":"2024-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"ConoGen is a conotoxin generation model that adopts an architecture consistent with the pre-training model ProGen (Figure 2A) 15 . ConoGen is also constructed using a Transformer-based neural network architecture. One major advantage of the Transformer lies in its self-attention mechanism, which enables the model to encode distant signals of information when making sequence predictions, thereby affording it the capability to comprehend complex semantics. Its Transformer architecture consists of 36 layers, each comprising 8 self-attention heads, totaling 1.2 billion trainable parameters to ensure the training and optimization of the model.\""},{"id":"epoch:039d1f56df96c3f217bd:parameters","entityId":"epoch:039d1f56df96c3f217bd","metric":"parameters","value":7520000000.0,"unit":"parameters","observationPeriod":"2024-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.52\u00d710^9"},{"id":"epoch:039d1f56df96c3f217bd:training-compute","entityId":"epoch:039d1f56df96c3f217bd","metric":"training-compute","value":4.983e+22,"unit":"FLOP","observationPeriod":"2024-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4.53\u00d710^10 FLOPs per token (Table 11)\n\n1100000000000 tokens *  4.53\u00d710^10 FLOPs per token = 4.983e+22\n\n1100000000000 tokens * 7*10^9 parameters * 6 = 4.62e+22"},{"id":"epoch:59a9169177aa1033f02b:parameters","entityId":"epoch:59a9169177aa1033f02b","metric":"parameters","value":3100000000.0,"unit":"parameters","observationPeriod":"2024-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" 3.10\u00d710^9"},{"id":"epoch:59a9169177aa1033f02b:training-compute","entityId":"epoch:59a9169177aa1033f02b","metric":"training-compute","value":2.0944e+22,"unit":"FLOP","observationPeriod":"2024-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.87\u00d710^10 FLOPs per token (Table 11)\n\n1120000000000 tokens*1.87\u00d710^10 = 2.0944e+22\n\n1120000000000 tokens *3*10^9 parameters *6 = 2.016e+22"},{"id":"epoch:10500385697bef8d54a7:parameters","entityId":"epoch:10500385697bef8d54a7","metric":"parameters","value":117000000.0,"unit":"parameters","observationPeriod":"2024-09-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:cef359fc13271662a3f8:training-compute","entityId":"epoch:cef359fc13271662a3f8","metric":"training-compute","value":3.7e+19,"unit":"FLOP","observationPeriod":"2024-09-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 4x NVIDIA Tesla P100 GPUs (1.90\u00d710\u00b9\u00b3 FLOP/s per GPU)\n\n2. Training duration: 14 days (directly provided) = 1,209,600 seconds\n\n3. Utilization rate: 40%\n\n4. Final calculation:\n1.90\u00d710\u00b9\u00b3 FLOP/s/GPU \u00d7 4 GPUs \u00d7 1,209,600 seconds \u00d7 0.4 = 3.7\u00d710\u00b9\u2079 FLOPs"},{"id":"epoch:25b138744963ae7903ea:parameters","entityId":"epoch:25b138744963ae7903ea","metric":"parameters","value":1000000.0,"unit":"parameters","observationPeriod":"2024-09-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:87f60eaeddb542ef2bd3:parameters","entityId":"epoch:87f60eaeddb542ef2bd3","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2024-09-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:87f60eaeddb542ef2bd3:training-compute","entityId":"epoch:87f60eaeddb542ef2bd3","metric":"training-compute","value":1.33583e+22,"unit":"FLOP","observationPeriod":"2024-09-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from Table 8:\nh100 gpus [assuming NVIDIA H100 SXM5 80GB]\npre-train 128 gpus 33.3 hours 4.2k gpu-hours\nfine-tune 256 gpus 32.4 hours 8.3k gpu-hours\n\n989500000000000 FLOP/sec [bf16 precision] * 3600 s/hour * 0.3 [assumed utilization] * (4200 + 8300) = 1.33583\u00d710^22 FLOP\n\n\"We set a maximum sequence length of 2304 for both pre-training and fine-tuning\"\nfrom Table 6:\n6 FLOP/parameter/token * 2304 max tokens / sequence * 72B paramters \n[Pre-train]: ((128 sequences [batch size] * 22300 steps) +\n[Fine-tune]: (256 sequences [batch size] * 20000 steps)) = 7.9371436032 \u00d7 10^21 FLOP\n\nhardware estimation seems more correct since operation counting doesn't account for vision module training"},{"id":"epoch:915400054017fa833b34:parameters","entityId":"epoch:915400054017fa833b34","metric":"parameters","value":63600000.0,"unit":"parameters","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:915400054017fa833b34:training-compute","entityId":"epoch:915400054017fa833b34","metric":"training-compute","value":2.7713e+19,"unit":"FLOP","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup:\n- Initial: 8x NVIDIA H100 SXM5 (7.56e14 FLOP/s per GPU)\n- Further: 1x NVIDIA A800 (7.80e13 FLOP/s)\n\n2. Training duration (calculated from steps):\n- Initial: 119,000 steps \u00d7 (256\u00d72314) tokens/step = 7.04e10 tokens\n- Further: 31,000 steps \u00d7 (32\u00d72241) tokens/step = 2.22e9 tokens\n\n4. Final calculation:\nInitial: 6 \u00d7 6.36e7 params \u00d7 7.04e10 tokens = 2.68e19 FLOPs\nFurther: 6 \u00d7 6.36e7 params \u00d7 2.22e9 tokens = 8.46e17 FLOPs\nTotal: 2.7712964e+19"},{"id":"epoch:d271e2ee6ce074545c38:parameters","entityId":"epoch:d271e2ee6ce074545c38","metric":"parameters","value":650000000.0,"unit":"parameters","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d271e2ee6ce074545c38:training-compute","entityId":"epoch:d271e2ee6ce074545c38","metric":"training-compute","value":3.900000000000003e+22,"unit":"FLOP","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup:\n- ProtBFN: 256 TPU v4 chips (2.75e14 FLOP/s per chip)\n- AbBFN: 128 TPU v4 chips (2.75e14 FLOP/s per chip)\n\n2. Training duration:\n- ProtBFN: 14 days = 1,209,600 seconds\n- AbBFN: 4 days = 345,600 seconds\n\n3. Utilization rate: 40%\n\n4. Final calculation:\n- ProtBFN: 2.75e14 \u00d7 256 \u00d7 1,209,600 \u00d7 0.4 = 3.406e22 FLOPs\n- AbBFN: 2.75e14 \u00d7 128 \u00d7 345,600 \u00d7 0.4 = 4.866e21 FLOPs\n- Total: 3.406e22 + 4.866e21 = 3.9e22 FLOPs"},{"id":"epoch:375ecbd3c670b7c2676e:parameters","entityId":"epoch:375ecbd3c670b7c2676e","metric":"parameters","value":10600000000.0,"unit":"parameters","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/meta-llama/Llama-3.2-11B-Vision"},{"id":"epoch:375ecbd3c670b7c2676e:training-compute","entityId":"epoch:375ecbd3c670b7c2676e","metric":"training-compute","value":5.79e+23,"unit":"FLOP","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Tensor type is BF16 (https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct).\n\n\u201cTraining utilized a cumulative of 2.02M GPU hours of computation on H100-80GB (TDP of 700W) type hardware, per the table below. Training time is the total GPU time required for training each model and power consumption is the peak power capacity per GPU device used, adjusted for power usage efficiency\u2026 Training time: Stage 1 pretraining: 147K H100 hours Stage 2 annealing: 98K H100 hours SFT: 896 H100 hours RLHF: 224 H100 hours.\u201d (https://github.com/meta-llama/llama-models/blob/main/models/llama3_2/MODEL_CARD_VISION.md#hardware-and-software).\n\nThe Nvidia H100 80GB is the H100 SXM. BFLOAT16 Tensor Core peak FLOPS with sparsity is 1,979 teraFLOPS (https://resources.nvidia.com/en-us-tensor-core/nvidia-tensor-core-gpu-datasheet?ncid=no-ncid).\n\nAssuming 33% utilization rate,\nTraining compute\n~= 0.33 * ( 147000 + 98000 + 896 + 224 ) hours * 3600 s / hour * 1979e12 FLOPS / GPU\n~= 5.79e23 FLOPS"},{"id":"epoch:ff1055388484de06f19c:parameters","entityId":"epoch:ff1055388484de06f19c","metric":"parameters","value":88600000000.0,"unit":"parameters","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/meta-llama/Llama-3.2-90B-Vision"},{"id":"epoch:bbb9004e9d3e3db8379e:parameters","entityId":"epoch:bbb9004e9d3e3db8379e","metric":"parameters","value":1230000000.0,"unit":"parameters","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/meta-llama/Llama-3.2-1B"},{"id":"epoch:bbb9004e9d3e3db8379e:training-compute","entityId":"epoch:bbb9004e9d3e3db8379e","metric":"training-compute","value":6.642e+22,"unit":"FLOP","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*1230000000.00*9000000000000 = 6.642e+22\n\n370000 hours * 3600 s * 133800000000000 FLOPS/s* 0.3 = 5.346648e+22"},{"id":"epoch:43a7df5b5f90a5a1bcab:parameters","entityId":"epoch:43a7df5b5f90a5a1bcab","metric":"parameters","value":3210000000.0,"unit":"parameters","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/meta-llama/Llama-3.2-1B"},{"id":"epoch:43a7df5b5f90a5a1bcab:training-compute","entityId":"epoch:43a7df5b5f90a5a1bcab","metric":"training-compute","value":1.7334e+23,"unit":"FLOP","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*3210000000.00*9000000000000 = 1.7334e+23\n\n460000 hours * 3600 s * 133800000000000 FLOPS/s* 0.3 = 6.647184e+22"},{"id":"epoch:c0c7cbcf57ef50d57766:parameters","entityId":"epoch:c0c7cbcf57ef50d57766","metric":"parameters","value":650000000.0,"unit":"parameters","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:975585972f822e1c2924:parameters","entityId":"epoch:975585972f822e1c2924","metric":"parameters","value":738000000.0,"unit":"parameters","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"This model comprises 36  layers and has a dimensionality of 1280, amounting to a total of 738 million parameters [28].\""},{"id":"epoch:975585972f822e1c2924:training-compute","entityId":"epoch:975585972f822e1c2924","metric":"training-compute","value":1.8397012e+16,"unit":"FLOP","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*4154700*738000000=1.8397012e+16"},{"id":"epoch:f477116777265d90a6ce:parameters","entityId":"epoch:f477116777265d90a6ce","metric":"parameters","value":943964.0,"unit":"parameters","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9387c350eec21d5739a7:parameters","entityId":"epoch:9387c350eec21d5739a7","metric":"parameters","value":20000000.0,"unit":"parameters","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In particular, the student model in MTDP has a significantly smaller parameter scale (~20 million)\" from 2.1 Model Structure"},{"id":"epoch:9387c350eec21d5739a7:training-compute","entityId":"epoch:9387c350eec21d5739a7","metric":"training-compute","value":600000000000000.0,"unit":"FLOP","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 * 20,000,000 connections * 500,000 training examples * 10 epochs = 6.0e14 FLOP"},{"id":"epoch:91ea78a9069dc9bfe545:training-compute","entityId":"epoch:91ea78a9069dc9bfe545","metric":"training-compute","value":5.75e+22,"unit":"FLOP","observationPeriod":"2024-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.394e9 * 6e13 * 0.4 = 5.75e22"},{"id":"epoch:0985bac24a051c60c9a2:training-compute","entityId":"epoch:0985bac24a051c60c9a2","metric":"training-compute","value":1.6500000000000016e+19,"unit":"FLOP","observationPeriod":"2024-09-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 1x NVIDIA A100 80GB (3.12 x 10^14 FLOP/s)\n\n2. Training duration: \n   - pst-large: 33.7 hours (directly provided)\n   - pst-small: 3.15 hours (estimated based on parameter ratio 178M/5M)\n\n3. Utilization rate: 40%\n\n4. Calculation:\n   pst-large: 3.12 x 10^14 \u00d7 121,320 \u00d7 0.40 = 1.51 x 10^19 FLOPs\n   pst-small: 3.12 x 10^14 \u00d7 11,340 \u00d7 0.40 = 1.42 x 10^18 FLOPs\n   Total: 1.65 x 10^19 FLOPs"},{"id":"epoch:750b7cc03dc1af3442f7:parameters","entityId":"epoch:750b7cc03dc1af3442f7","metric":"parameters","value":7900000.0,"unit":"parameters","observationPeriod":"2024-09-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"As a result, PocketGen requires significantly fewer trainable parameters than RFDiffusionAA [16] (7.9M versus 82.9M trainable parameters).\""},{"id":"epoch:750b7cc03dc1af3442f7:training-compute","entityId":"epoch:750b7cc03dc1af3442f7","metric":"training-compute","value":2.1e+19,"unit":"FLOP","observationPeriod":"2024-09-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"It takes around 48 hours to finish training on 1 Tesla A100 GPU from scratch.\" Assume 40% utilization, FP16 tensor precision. "},{"id":"epoch:b16c8b51cde975a84ac0:training-compute","entityId":"epoch:b16c8b51cde975a84ac0","metric":"training-compute","value":3.5000000000000287e+18,"unit":"FLOP","observationPeriod":"2024-09-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 1x NVIDIA GeForce RTX 3090 (1.60 x 10^14 FLOPs/s per GPU)\n\n2. Training duration: Estimated\n   * 71,000 samples, batch size 64 \u2192 1,100 batches/epoch\n   * 0.5s/batch \u2192 9 min/epoch\n   * 100 epochs \u2192 15 hours (54,000 seconds)\n\n3. Utilization rate: 40%\n\n4. Final calculation:\n   1.60 x 10^14 FLOPs/s \u00d7 1 GPU \u00d7 54,000s \u00d7 0.4 = 3.456 x 10^18 FLOPs"},{"id":"epoch:84d1d48aeb7d1a2c48c5:parameters","entityId":"epoch:84d1d48aeb7d1a2c48c5","metric":"parameters","value":350000000.0,"unit":"parameters","observationPeriod":"2024-09-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:84d1d48aeb7d1a2c48c5:training-compute","entityId":"epoch:84d1d48aeb7d1a2c48c5","metric":"training-compute","value":1.1000000000000008e+22,"unit":"FLOP","observationPeriod":"2024-09-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware: A100 GPUs with 3.12\u00d710\u00b9\u2074 FLOP/s per GPU (bf16/fp16)\n\n2. Duration: Directly provided - 1,014 GPU days = 8.75\u00d710\u2077 seconds\n\n3. Utilization: 40%\n\n4. Calculation:\n3.12\u00d710\u00b9\u2074 FLOP/s \u00d7 8.75\u00d710\u2077 s \u00d7 0.40 = 1.09\u00d710\u00b2\u00b2 FLOPs"},{"id":"epoch:d353ea1142815cdf5b65:parameters","entityId":"epoch:d353ea1142815cdf5b65","metric":"parameters","value":960000.0,"unit":"parameters","observationPeriod":"2024-09-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:8719e0044290593e9dc9:training-compute","entityId":"epoch:8719e0044290593e9dc9","metric":"training-compute","value":8.599999999999978e+20,"unit":"FLOP","observationPeriod":"2024-09-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 8x NVIDIA A100 SXM4 80GB GPUs, 3.12e14 FLOP/s per GPU\n\n2. Training duration: Directly provided as 10 days = 864,000 seconds\n\n3. Utilization rate: 40%\n\n4. Final calculation:\n3.12e14 FLOP/s \u00d7 864,000s \u00d7 0.4 \u00d7 8 GPUs = 8.6e20 FLOPs"},{"id":"epoch:83e58ea2e87b06cc2d4f:parameters","entityId":"epoch:83e58ea2e87b06cc2d4f","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2024-09-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b3b7ee8c3f03a88ccab4:parameters","entityId":"epoch:b3b7ee8c3f03a88ccab4","metric":"parameters","value":115000000000.0,"unit":"parameters","observationPeriod":"2024-09-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b3b7ee8c3f03a88ccab4:training-compute","entityId":"epoch:b3b7ee8c3f03a88ccab4","metric":"training-compute","value":6.9e+24,"unit":"FLOP","observationPeriod":"2024-09-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND: 6 * 115B * 10T = 6.9e24"},{"id":"epoch:ac129cf4f39b493a4cec:parameters","entityId":"epoch:ac129cf4f39b493a4cec","metric":"parameters","value":2300000000.0,"unit":"parameters","observationPeriod":"2024-09-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"a 2.3 billion parameter foundation model developed using 160 variables from the Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2)."},{"id":"epoch:ac129cf4f39b493a4cec:training-compute","entityId":"epoch:ac129cf4f39b493a4cec","metric":"training-compute","value":7.15e+19,"unit":"FLOP","observationPeriod":"2024-09-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we train the model on 64 A100 GPUs and batch size 1 for 100,000 gradient descent steps\"\n\"With these choices we are dealing with 51,840\ntokens per sample yet are keeping the length of the global and local sequence roughly balanced\"\n\nTotal number of tokens: 100000*51840=5184000000\nTraining compute: 6*5184000000*2300000000=7.15392e+19"},{"id":"epoch:ca56c6bb974f96f713ac:parameters","entityId":"epoch:ca56c6bb974f96f713ac","metric":"parameters","value":72700000000.0,"unit":"parameters","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"72.7B"},{"id":"epoch:ca56c6bb974f96f713ac:training-compute","entityId":"epoch:ca56c6bb974f96f713ac","metric":"training-compute","value":7.8e+24,"unit":"FLOP","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training dataset size was 18 trillion\n\n6ND = 6 * 72.7 billion parameters * 18 trillion tokens = 7.8e24"},{"id":"epoch:4829719699ed230be230:parameters","entityId":"epoch:4829719699ed230be230","metric":"parameters","value":5083905.0,"unit":"parameters","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Loaded checkpoint and counted parameters."},{"id":"epoch:cc575170cba1a9155957:training-compute","entityId":"epoch:cc575170cba1a9155957","metric":"training-compute","value":6.480000000000066e+18,"unit":"FLOP","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware: 1x NVIDIA A30 GPU (1.50\u00d710\u00b9\u2074 FLOP/s using fp16 tensor cores)\n\n2. Training duration: Directly provided - 15 epochs \u00d7 2 hours/epoch = 30 hours = 108,000 seconds\n\n3. Utilization rate: 40%\n\n4. Calculation:\n   1.50\u00d710\u00b9\u2074 FLOP/s \u00d7 1 GPU \u00d7 108,000s \u00d7 0.40 = 6.48\u00d710\u00b9\u2078 FLOP"},{"id":"epoch:d6513ab5b0c7fec0ca77:parameters","entityId":"epoch:d6513ab5b0c7fec0ca77","metric":"parameters","value":7610000000.0,"unit":"parameters","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:3cd2e63ebab00594e59e:parameters","entityId":"epoch:3cd2e63ebab00594e59e","metric":"parameters","value":72700000000.0,"unit":"parameters","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Number of Parameters: 72.7B\nNumber of Paramaters (Non-Embedding): 70.0B"},{"id":"epoch:3cd2e63ebab00594e59e:training-compute","entityId":"epoch:3cd2e63ebab00594e59e","metric":"training-compute","value":7.8516e+24,"unit":"FLOP","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0999f97e50fd5a34bf14:parameters","entityId":"epoch:0999f97e50fd5a34bf14","metric":"parameters","value":3090000000.0,"unit":"parameters","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.09B"},{"id":"epoch:0999f97e50fd5a34bf14:training-compute","entityId":"epoch:0999f97e50fd5a34bf14","metric":"training-compute","value":3.3372e+23,"unit":"FLOP","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training dataset size was 18 trillion\n\n6 FLOP/parameter/token * 3090000000 parameters * 18000000000000 tokens = 3.3372e+23 FLOP"},{"id":"epoch:880b93fc19c0e6e6dbd5:parameters","entityId":"epoch:880b93fc19c0e6e6dbd5","metric":"parameters","value":7610000000.0,"unit":"parameters","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.61B"},{"id":"epoch:880b93fc19c0e6e6dbd5:training-compute","entityId":"epoch:880b93fc19c0e6e6dbd5","metric":"training-compute","value":8.2188e+23,"unit":"FLOP","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training dataset size was 18 trillion\n\n6ND = 6 * 7.61 billion parameters * 18 trillion tokens = 8.2188e+23"},{"id":"epoch:9719d12dd9cac22087f4:parameters","entityId":"epoch:9719d12dd9cac22087f4","metric":"parameters","value":1540000000.0,"unit":"parameters","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.54B"},{"id":"epoch:9719d12dd9cac22087f4:training-compute","entityId":"epoch:9719d12dd9cac22087f4","metric":"training-compute","value":1.6632e+23,"unit":"FLOP","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training dataset size was 18 trillion\n\n6ND = 6 * 1.54B billion parameters * 18 trillion tokens = 1.6632e+23"},{"id":"epoch:f9685715553e4f0db1f0:parameters","entityId":"epoch:f9685715553e4f0db1f0","metric":"parameters","value":14700000000.0,"unit":"parameters","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14.7B according to the model card\nhttps://qwenlm.github.io/blog/qwen2.5-llm/"},{"id":"epoch:f9685715553e4f0db1f0:training-compute","entityId":"epoch:f9685715553e4f0db1f0","metric":"training-compute","value":1.58760000000001e+24,"unit":"FLOP","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training dataset size was 18 trillion\n\n6ND = 6 * 14.7 billion parameters * 18 trillion tokens = 1.59e24"},{"id":"epoch:10dbd1c5d0326edfa7b6:parameters","entityId":"epoch:10dbd1c5d0326edfa7b6","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:10dbd1c5d0326edfa7b6:training-compute","entityId":"epoch:10dbd1c5d0326edfa7b6","metric":"training-compute","value":8.6388e+23,"unit":"FLOP","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"base model: 8.2188e+23 FLOP\nfine-tuning: 4.2e+22 FLOP \n\nTotal: 8.6388e+23 FLOP"},{"id":"epoch:a4173af20c262e8548ab:parameters","entityId":"epoch:a4173af20c262e8548ab","metric":"parameters","value":34000000000.0,"unit":"parameters","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e805435a7cbb2857bbb0:parameters","entityId":"epoch:e805435a7cbb2857bbb0","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:59c7fef4d3c183ce417b:parameters","entityId":"epoch:59c7fef4d3c183ce417b","metric":"parameters","value":32500000000.0,"unit":"parameters","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32.5B"},{"id":"epoch:eea630894545f1989023:parameters","entityId":"epoch:eea630894545f1989023","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.5B"},{"id":"epoch:eea630894545f1989023:training-compute","entityId":"epoch:eea630894545f1989023","metric":"training-compute","value":1.7532e+23,"unit":"FLOP","observationPeriod":"2024-09-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"base model: 1.6632e+23 FLOP\nfine-tuning: 9e+21 FLOP\n\nTotal: 1.7532e+23 FLOP"},{"id":"epoch:d2d2a1b5b8c6d2a41dd9:parameters","entityId":"epoch:d2d2a1b5b8c6d2a41dd9","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2024-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"72 billion (language model) and 675M (vision encoder)"},{"id":"epoch:d2d2a1b5b8c6d2a41dd9:training-compute","entityId":"epoch:d2d2a1b5b8c6d2a41dd9","metric":"training-compute","value":6.048e+23,"unit":"FLOP","observationPeriod":"2024-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 FLOP / token / parameter \u00d7 1.4\u00d710^12 tokens \u00d7 7.2\u00d710^10 parameters = 6.048e+23 FLOP"},{"id":"epoch:d52f99f0e902e4f9c2cd:parameters","entityId":"epoch:d52f99f0e902e4f9c2cd","metric":"parameters","value":2000000000.0,"unit":"parameters","observationPeriod":"2024-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.5 billion (language model) and 675M (vision encoder)"},{"id":"epoch:d52f99f0e902e4f9c2cd:training-compute","entityId":"epoch:d52f99f0e902e4f9c2cd","metric":"training-compute","value":1.68e+22,"unit":"FLOP","observationPeriod":"2024-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 FLOP / parameter / token \u00d7 1.4\u00d710^12 tokens \u00d7 2\u00d710^9 parameters = 1.68e+22 FLOP"},{"id":"epoch:a113cb751693e63c34ce:parameters","entityId":"epoch:a113cb751693e63c34ce","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.6 billion (language model) and 675M (vision encoder)"},{"id":"epoch:a113cb751693e63c34ce:training-compute","entityId":"epoch:a113cb751693e63c34ce","metric":"training-compute","value":6.72e+22,"unit":"FLOP","observationPeriod":"2024-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 FLOP / parameter / token \u00d7 1.4\u00d710^12 tokens \u00d7 8\u00d710^9 parameters = 6.72e+22 FLOP"},{"id":"epoch:d136d9a558e2b9cd828b:parameters","entityId":"epoch:d136d9a558e2b9cd828b","metric":"parameters","value":7610000000.0,"unit":"parameters","observationPeriod":"2024-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Number of Parameters: 7.61B"},{"id":"epoch:d136d9a558e2b9cd828b:training-compute","entityId":"epoch:d136d9a558e2b9cd828b","metric":"training-compute","value":2.5113e+23,"unit":"FLOP","observationPeriod":"2024-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 FLOP / token / parameter *7610000000 parameters *5.5T tokens =2.5113e+23 FLOP"},{"id":"epoch:38175a201dca6c5e701d:parameters","entityId":"epoch:38175a201dca6c5e701d","metric":"parameters","value":1540000000.0,"unit":"parameters","observationPeriod":"2024-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Number of Parameters: 1.54B"},{"id":"epoch:38175a201dca6c5e701d:training-compute","entityId":"epoch:38175a201dca6c5e701d","metric":"training-compute","value":5.082e+22,"unit":"FLOP","observationPeriod":"2024-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND =  6*1540000000 parameters *5.5T tokens =5.082e+22"},{"id":"epoch:bdab79d877a37c838075:parameters","entityId":"epoch:bdab79d877a37c838075","metric":"parameters","value":22000000000.0,"unit":"parameters","observationPeriod":"2024-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"22B"},{"id":"epoch:c86e70d9bec1fdd67315:parameters","entityId":"epoch:c86e70d9bec1fdd67315","metric":"parameters","value":3800000000.0,"unit":"parameters","observationPeriod":"2024-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:69c77afdbab8b595ef74:parameters","entityId":"epoch:69c77afdbab8b595ef74","metric":"parameters","value":12400000000.0,"unit":"parameters","observationPeriod":"2024-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"New 400M parameter vision encoder trained from scratch\"\n+\n\"12B parameter multimodal decoder based on Mistral Nemo\""},{"id":"epoch:746bdb8393e1dc932882:parameters","entityId":"epoch:746bdb8393e1dc932882","metric":"parameters","value":32500000000.0,"unit":"parameters","observationPeriod":"2024-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32.5B"},{"id":"epoch:746bdb8393e1dc932882:training-compute","entityId":"epoch:746bdb8393e1dc932882","metric":"training-compute","value":3.51e+24,"unit":"FLOP","observationPeriod":"2024-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 32.5B parameters * 18 trillion tokens = 3.51 \u00d7 10^24 FLOP"},{"id":"epoch:3b36e35f21cce147e65c:parameters","entityId":"epoch:3b36e35f21cce147e65c","metric":"parameters","value":110000000.0,"unit":"parameters","observationPeriod":"2024-09-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our final network has \u2248 110M parameters in total and all of the SF-level networks were trained for 30 epochs, using a batch size of 255\""},{"id":"epoch:689ef13b087c06fa47dd:parameters","entityId":"epoch:689ef13b087c06fa47dd","metric":"parameters","value":10200000000.0,"unit":"parameters","observationPeriod":"2024-09-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:504c4482c0e725e3dbb6:training-compute","entityId":"epoch:504c4482c0e725e3dbb6","metric":"training-compute","value":2.481192e+19,"unit":"FLOP","observationPeriod":"2024-09-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Autoencoder: 8.5h per epoch on single H100, 2 epochs total\nDiffusion model: 3.5h per epoch on single H100, 3 epochs total\nReward model: 7h on a single A100\n\nTotal training time: 2*8.5+3*3.5+7=34.5\n\nCompute: \n((2*8.5+3*3.5)*756000000000000+7*312000000000000)*60*60*0.3=2.481192e+19"},{"id":"epoch:961faf23bb503db4455c:parameters","entityId":"epoch:961faf23bb503db4455c","metric":"parameters","value":650000000.0,"unit":"parameters","observationPeriod":"2024-09-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f6baaf8e20667382cd88:parameters","entityId":"epoch:f6baaf8e20667382cd88","metric":"parameters","value":650000000.0,"unit":"parameters","observationPeriod":"2024-09-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:286510009fd272d6e96e:parameters","entityId":"epoch:286510009fd272d6e96e","metric":"parameters","value":17800000.0,"unit":"parameters","observationPeriod":"2024-09-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Taken from Table 2"},{"id":"epoch:286510009fd272d6e96e:training-compute","entityId":"epoch:286510009fd272d6e96e","metric":"training-compute","value":2.59999999999998e+20,"unit":"FLOP","observationPeriod":"2024-09-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 1x NVIDIA A100 GPU with 3.12\u00d710\u00b9\u2074 FLOP/s (fp16_tensor)\n\n2. Training duration: Directly provided as 24 GPU days (9 days initial + 15 days second phase) = 2.0736\u00d710\u2076 seconds\n\n3. Utilization rate: 40%\n\n4. Final calculation:\n3.12\u00d710\u00b9\u2074 FLOP/s \u00d7 1 GPU \u00d7 2.0736\u00d710\u2076 s \u00d7 0.4 = 2.6\u00d710\u00b2\u2070 FLOP"},{"id":"epoch:c457dead213a434fcce2:parameters","entityId":"epoch:c457dead213a434fcce2","metric":"parameters","value":32000000.0,"unit":"parameters","observationPeriod":"2024-09-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32M (safetensors)"},{"id":"epoch:c457dead213a434fcce2:training-compute","entityId":"epoch:c457dead213a434fcce2","metric":"training-compute","value":1.1e+19,"unit":"FLOP","observationPeriod":"2024-09-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Novae was trained on a Nvidia HGX A100 GPU for 24 hours.\" Assume FP16 tensor precision and 40% utilization. "},{"id":"epoch:39fb4f6e1a5398003b03:parameters","entityId":"epoch:39fb4f6e1a5398003b03","metric":"parameters","value":89600000.0,"unit":"parameters","observationPeriod":"2024-09-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"total number of parameters to 89.6 million.\""},{"id":"epoch:ff96ae963d580e522207:training-compute","entityId":"epoch:ff96ae963d580e522207","metric":"training-compute","value":5.500000000000003e+19,"unit":"FLOP","observationPeriod":"2024-09-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 1x NVIDIA RTX 3090 (1.60e14 FLOP/s FP16 tensor)\n\n2. Training duration: 10 days provided directly (864,000 seconds)\n\n3. Utilization rate: 40%\n\n4. Calculation: 1.60e14 FLOP/s \u00d7 1 GPU \u00d7 864,000s \u00d7 0.4 = 5.5e19 FLOPs"},{"id":"epoch:9b2e73296d2062d02448:parameters","entityId":"epoch:9b2e73296d2062d02448","metric":"parameters","value":27200000000.0,"unit":"parameters","observationPeriod":"2024-09-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"27.2B"},{"id":"epoch:2b841703061cebebd0d3:parameters","entityId":"epoch:2b841703061cebebd0d3","metric":"parameters","value":335000000.0,"unit":"parameters","observationPeriod":"2024-09-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The architecture incorporated U-Net [23] style skip connections, 24 layers, 16 attention heads, an embedding dimension of 1024, a linear layer dimension of 4096. The character embedding vocabulary size was 399.5 The total number of parameters amounted to 335 million\""},{"id":"epoch:2b841703061cebebd0d3:training-compute","entityId":"epoch:2b841703061cebebd0d3","metric":"training-compute","value":4.939776e+20,"unit":"FLOP","observationPeriod":"2024-09-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 245760000000 tokens * 335000000 parameters = 4.939776e+20 FLOP "},{"id":"epoch:9192d769f247ac9da918:parameters","entityId":"epoch:9192d769f247ac9da918","metric":"parameters","value":5820515.0,"unit":"parameters","observationPeriod":"2024-09-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The model was based on a recurrent neural network with long short-term memory (LSTM) cells and consisted of four layers with a total of 5,820,515\nparameters: layer 1, BatchNormalization; layer 2, LSTM with 1024 units;\nlayer 3, LSTM with 256 units; layer 4, BatchNormalization.\""},{"id":"epoch:56e50afad8d9231a33f0:parameters","entityId":"epoch:56e50afad8d9231a33f0","metric":"parameters","value":22000000000.0,"unit":"parameters","observationPeriod":"2024-09-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"22B"},{"id":"epoch:eaa56ca09545a70680a2:parameters","entityId":"epoch:eaa56ca09545a70680a2","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2024-09-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"In June 2024, the company released 0.7B, 7B, and 32B large models based on small computing power."},{"id":"epoch:00d8f7376694e812ad0f:parameters","entityId":"epoch:00d8f7376694e812ad0f","metric":"parameters","value":4000000.0,"unit":"parameters","observationPeriod":"2024-09-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"CPDiffusion trains a denoising diffusion model with 4 million\nlearnable parameters from natural protein structures. "},{"id":"epoch:7b67d5b6fa67e0cc4a8d:parameters","entityId":"epoch:7b67d5b6fa67e0cc4a8d","metric":"parameters","value":38700000.0,"unit":"parameters","observationPeriod":"2024-09-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Taken from table 7, \"model size\" row; \"medium (38.7M)\""},{"id":"epoch:7b67d5b6fa67e0cc4a8d:training-compute","entityId":"epoch:7b67d5b6fa67e0cc4a8d","metric":"training-compute","value":4.26816e+18,"unit":"FLOP","observationPeriod":"2024-09-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We utilized an NVIDIA A100 GPU to train Molphenix using Phenom1 and MolGPS embeddings, which takes approximately \u223c4.75 hours each.\" Assume FP16 precision, 40% utilization. \n\n9.5*60*60*312000000000000*0.4=4.26816e+18"},{"id":"epoch:c1535564e2a29c616577:parameters","entityId":"epoch:c1535564e2a29c616577","metric":"parameters","value":288915.0,"unit":"parameters","observationPeriod":"2024-09-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Loaded model from https://github.com/ gzBiomedical/EpiScan and counted parameters."},{"id":"epoch:88cf21ac913d814ad92d:parameters","entityId":"epoch:88cf21ac913d814ad92d","metric":"parameters","value":660000.0,"unit":"parameters","observationPeriod":"2024-09-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:56dc4a7cfb0573ac7dd7:parameters","entityId":"epoch:56dc4a7cfb0573ac7dd7","metric":"parameters","value":734000000.0,"unit":"parameters","observationPeriod":"2024-09-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:56dc4a7cfb0573ac7dd7:training-compute","entityId":"epoch:56dc4a7cfb0573ac7dd7","metric":"training-compute","value":4.2506168e+21,"unit":"FLOP","observationPeriod":"2024-09-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Finetuned ProtGPT2 on 57M sequences with average length of ~120 (Figure 6)\n\nFT compute: 734000000*6840000000*5*6=1.506168e+20\n\nBase model compute: 4.1e+21\n\nTotal: 4.2506168e+21"},{"id":"epoch:90be8ebc4f9be282b5c7:parameters","entityId":"epoch:90be8ebc4f9be282b5c7","metric":"parameters","value":217000000.0,"unit":"parameters","observationPeriod":"2024-09-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"the Base model consists of 217M learnable parameters\""},{"id":"epoch:90be8ebc4f9be282b5c7:training-compute","entityId":"epoch:90be8ebc4f9be282b5c7","metric":"training-compute","value":3.6084096e+21,"unit":"FLOP","observationPeriod":"2024-09-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"265 hours *3600 sec / hour *64 GPUs *197000000000000 FLOP / s *0.3 [assumed utilization] = 3.6084096e+21 FLOP"},{"id":"epoch:d270a679c50082acee83:parameters","entityId":"epoch:d270a679c50082acee83","metric":"parameters","value":650000000.0,"unit":"parameters","observationPeriod":"2024-09-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"so the ESM2 model with 650 million parameters is used to continue training here.\""},{"id":"epoch:9a9b2faf81714359dd53:training-compute","entityId":"epoch:9a9b2faf81714359dd53","metric":"training-compute","value":3.074112e+17,"unit":"FLOP","observationPeriod":"2024-09-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware: 1x NVIDIA GeForce RTX 3090\n\n2. Training duration: 6 hours (max of 1-6 hour range) = 21,600 seconds\n\n3. Utilization: 40%\n\n4. Calculation: \n\n\n35580000000000*6*60*60*0.4=3.074112e+17"},{"id":"epoch:22507184f5d43e0a6d2f:parameters","entityId":"epoch:22507184f5d43e0a6d2f","metric":"parameters","value":236000000000.0,"unit":"parameters","observationPeriod":"2024-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"21B active params, 236B total"},{"id":"epoch:22507184f5d43e0a6d2f:training-compute","entityId":"epoch:22507184f5d43e0a6d2f","metric":"training-compute","value":1.7892e+24,"unit":"FLOP","observationPeriod":"2024-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"V2.5 is a merge of V2-coder and V2-chat\nV2-coder is trained for 6T additional tokens from an intermediate checkpoint of V2, which had been trained for 4.2T tokens. Total: 10.2T\nV2-chat is fine-tuned from V2, saw 8.2T tokens in pre-training\nUnique steps: 8.2T + 6T = 14.2T\nFLOPs: 6 * 21B * 14.2T = 1.7892e24"},{"id":"epoch:fe6ace81e637cf273231:parameters","entityId":"epoch:fe6ace81e637cf273231","metric":"parameters","value":1467000000.0,"unit":"parameters","observationPeriod":"2024-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:3f4cb7853be499a80a70:parameters","entityId":"epoch:3f4cb7853be499a80a70","metric":"parameters","value":141000000000.0,"unit":"parameters","observationPeriod":"2024-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f6bb9df9e5c13536ff57:parameters","entityId":"epoch:f6bb9df9e5c13536ff57","metric":"parameters","value":4000000000.0,"unit":"parameters","observationPeriod":"2024-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The MiniCPM3.0 model boasts 4 billion parameters\" (https://www.aibase.com/news/11581)."},{"id":"epoch:c6ed8784428c518e30db:parameters","entityId":"epoch:c6ed8784428c518e30db","metric":"parameters","value":670000000.0,"unit":"parameters","observationPeriod":"2024-09-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ed34f50ba12dcbea8db5:parameters","entityId":"epoch:ed34f50ba12dcbea8db5","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-09-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"6.9B total parameters, of which only 1.3B are activated for each input token\""},{"id":"epoch:ed34f50ba12dcbea8db5:training-compute","entityId":"epoch:ed34f50ba12dcbea8db5","metric":"training-compute","value":5.1741608015e+22,"unit":"FLOP","observationPeriod":"2024-09-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we pretrain OLMOE-1B-7B for 5.1 trillion tokens with 6.9B total parameters, of which only 1.3B are activated for each input token\"\nSo 6 * 5.1T * 1.3B = 3.978e22\n\nIn appendix B: \"We pretrain OLMOE-1B-7B on 256 H100 GPUs for approximately 10 days ... For adaptation, we use 32 H100 GPUs for 33 hours to instruction tune and for another 14 hours to preference tune via DPO.\"\nWhich would suggest 9.90e14 * (256*240 + 32*(33+14)) * 3600 * 0.3 = 6.730e22 including instruction/preference tuning\n\ngeometric mean: sqrt(3.978e22 * 6.730e22) = 5.174e22"},{"id":"epoch:27bf07bd883f7d988fc5:training-compute","entityId":"epoch:27bf07bd883f7d988fc5","metric":"training-compute","value":1.64975616e+21,"unit":"FLOP","observationPeriod":"2024-09-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training stages, Table 2:\n267+105+11+28+9+39=459 hours\n\n459*60*60*312000000000000*8*0.4=1649756160000000000000"},{"id":"epoch:799401f70fee987b88af:training-compute","entityId":"epoch:799401f70fee987b88af","metric":"training-compute","value":7.6197888e+20,"unit":"FLOP","observationPeriod":"2024-09-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 8x NVIDIA A100 GPUs, 3.12e+14 FLOP/s per GPU\n2. Training duration: 371 hours (267h AlphaFLOW + 104h ESMFLOW) = 1,335,600 seconds\n3. Utilization rate: 40%\n4. Final calculation: 8 GPUs \u00d7 3.12e+14 FLOP/s \u00d7 1,335,600s \u00d7 0.4 = 1.34e+21 FLOPs\n\nESMFlow training time, table 2\n104+37+5+34+9+23=212\n\n212hours*8*3.12e14*0.4=761978880000000000000"},{"id":"epoch:6aa7cc5470bebae98d5a:parameters","entityId":"epoch:6aa7cc5470bebae98d5a","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-08-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:daef624f1e7b9232ef2e:parameters","entityId":"epoch:daef624f1e7b9232ef2e","metric":"parameters","value":9560000000.0,"unit":"parameters","observationPeriod":"2024-08-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our state-of-the-art, in-house, audio encoder: Saaras v1\nMeta\u2019s Llama3-8B-Instruct as the decoder\nThe encoder and decoder are connected by a small projector with ~60M parameters.\""},{"id":"epoch:312ae6aba84baaa5a8b0:parameters","entityId":"epoch:312ae6aba84baaa5a8b0","metric":"parameters","value":21660193.0,"unit":"parameters","observationPeriod":"2024-08-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"transformer encoder stack + gated aggregator linear map + classifier + gat = (3*65536+65536+256*1024+1024*256)*6 + 512*256+(2048*256)*32+33+33280 "},{"id":"epoch:ee19142a5be7e359b9c8:parameters","entityId":"epoch:ee19142a5be7e359b9c8","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2024-08-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"11B"},{"id":"epoch:7aa44557e68bbb3bab33:parameters","entityId":"epoch:7aa44557e68bbb3bab33","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2024-08-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.3b"},{"id":"epoch:7aa44557e68bbb3bab33:training-compute","entityId":"epoch:7aa44557e68bbb3bab33","metric":"training-compute","value":7.8e+20,"unit":"FLOP","observationPeriod":"2024-08-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*1300000000.00*100*10^9 = 7.8e+20"},{"id":"epoch:0a29f0e072590034f566:parameters","entityId":"epoch:0a29f0e072590034f566","metric":"parameters","value":340000000.0,"unit":"parameters","observationPeriod":"2024-08-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"340m\n"},{"id":"epoch:0a29f0e072590034f566:training-compute","entityId":"epoch:0a29f0e072590034f566","metric":"training-compute","value":3.06e+19,"unit":"FLOP","observationPeriod":"2024-08-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*340*10^6*15*10^9 = 3.06e+19"},{"id":"epoch:29dda5a7c49c17526cd6:parameters","entityId":"epoch:29dda5a7c49c17526cd6","metric":"parameters","value":7041544704.0,"unit":"parameters","observationPeriod":"2024-08-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:29dda5a7c49c17526cd6:training-compute","entityId":"epoch:29dda5a7c49c17526cd6","metric":"training-compute","value":4.43e+23,"unit":"FLOP","observationPeriod":"2024-08-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"reported by the authors: 2.75*10^23 + 1.68*10^23 = 4.43*10^23 FLOP\n\nhttps://huggingface.co/Aleph-Alpha/Pharia-1-LLM-7B-control#compute--training-efficiency"},{"id":"epoch:db9cafd757a9f19ae99b:parameters","entityId":"epoch:db9cafd757a9f19ae99b","metric":"parameters","value":1200000000.0,"unit":"parameters","observationPeriod":"2024-08-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.2B"},{"id":"epoch:db9cafd757a9f19ae99b:training-compute","entityId":"epoch:db9cafd757a9f19ae99b","metric":"training-compute","value":7.1497946e+20,"unit":"FLOP","observationPeriod":"2024-08-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 1.2*10^9 parameters * 104857600000 tokens = 7.5497472e+20 FLOP\n\n989500000000000 FLOP / GPU / sec [bf16 assumed] *19.8 hours * 3600 sec / hour * 32 GPUs * 0.3 [assumed utilization] = 6.7710298e+20 FLOP\n\nsqrt(6.7710298e+20*7.5497472e+20) = 7.1497946e+20"},{"id":"epoch:21706cd9b9cc1f6fea6e:parameters","entityId":"epoch:21706cd9b9cc1f6fea6e","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-08-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:69981e1745f45cde7501:parameters","entityId":"epoch:69981e1745f45cde7501","metric":"parameters","value":398000000000.0,"unit":"parameters","observationPeriod":"2024-08-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"94B active/398B total"},{"id":"epoch:81d0c4cd0a2e221fea1a:parameters","entityId":"epoch:81d0c4cd0a2e221fea1a","metric":"parameters","value":52000000000.0,"unit":"parameters","observationPeriod":"2024-08-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12B active/52B total"},{"id":"epoch:0f6b001d52ba116be404:parameters","entityId":"epoch:0f6b001d52ba116be404","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2024-08-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"KOSMOS-2.5 contains a total of 1.3 billion parameters\"\n\n\"The vision encoder is initialized from the Pix2Struct-Large model\u2019s encoder (Lee et al. 2023), which is based on the Vision Transformer (ViT) (Dosovitskiy et al. 2021)\""},{"id":"epoch:0f6b001d52ba116be404:training-compute","entityId":"epoch:0f6b001d52ba116be404","metric":"training-compute","value":2.2018015e+21,"unit":"FLOP","observationPeriod":"2024-08-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 1.3 * 10^9 parameters * 260 * 10^9 tokens [1 epoch assumed] = 2.028e+21 FLOP\n\n+ 1.7380147e+20 FLOP [\"likely\" base model Pix2struct-Large training compute] = 2.2018015e+21 FLOP"},{"id":"epoch:12481267f85e9b4e2260:parameters","entityId":"epoch:12481267f85e9b4e2260","metric":"parameters","value":24670596.0,"unit":"parameters","observationPeriod":"2024-08-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Loaded provided model file from Zenodo page into PyTorch. "},{"id":"epoch:12481267f85e9b4e2260:training-compute","entityId":"epoch:12481267f85e9b4e2260","metric":"training-compute","value":1.7100000000000148e+18,"unit":"FLOP","observationPeriod":"2024-08-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 1x NVIDIA A100 GPU (3.12x10^14 FLOP/s using FP16 Tensor Core)\n\n2. Training duration: Directly provided - 0.76 hours per epoch x 5 folds = 3.8 hours (13,680 seconds)\n\n3. Utilization rate: 40%\n\n4. Calculation: 3.12x10^14 FLOP/s \u00d7 13,680s \u00d7 1 GPU \u00d7 0.4 = 1.71\u00d710^18 FLOPs"},{"id":"epoch:434a1adb16bd16d687de:parameters","entityId":"epoch:434a1adb16bd16d687de","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-08-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://arxiv.org/abs/2408.10188"},{"id":"epoch:4b8baf87a727fb650dbf:parameters","entityId":"epoch:4b8baf87a727fb650dbf","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2024-08-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We distill the Phi-1.5-1.3B model into Phi-Mamba1.5B as well as Hybrid-Phi-Mamba-1.5B\" (https://arxiv.org/pdf/2408.10189, page 8)."},{"id":"epoch:4b8baf87a727fb650dbf:training-compute","entityId":"epoch:4b8baf87a727fb650dbf","metric":"training-compute","value":1.215e+21,"unit":"FLOP","observationPeriod":"2024-08-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.17e+21 FLOP [base model compute] + 4.5e+19 FLOP [finetune compute] = 1.215e+21 FLOP"},{"id":"epoch:e5a2dfcf7034063b3898:training-compute","entityId":"epoch:e5a2dfcf7034063b3898","metric":"training-compute","value":2.11e+17,"unit":"FLOP","observationPeriod":"2024-08-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3*60*60*65130000000000*0.3=2.110212e+17"},{"id":"epoch:d5de6e40061f38215f92:parameters","entityId":"epoch:d5de6e40061f38215f92","metric":"parameters","value":4000000000.0,"unit":"parameters","observationPeriod":"2024-08-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1: shows the model is called xGen-MM-base (4B)"},{"id":"epoch:d5de6e40061f38215f92:training-compute","entityId":"epoch:d5de6e40061f38215f92","metric":"training-compute","value":2.4e+21,"unit":"FLOP","observationPeriod":"2024-08-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming 1 epoch and dense architecture (phi-3 mini is a dense model), compute = 6 * 4B * 100B = 2.4e+21 FLOP."},{"id":"epoch:a4b3674d160011001136:parameters","entityId":"epoch:a4b3674d160011001136","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e87a5ae24aa01f614965:parameters","entityId":"epoch:e87a5ae24aa01f614965","metric":"parameters","value":405000000000.0,"unit":"parameters","observationPeriod":"2024-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"405B"},{"id":"epoch:249c0532cb4e60e6e3a4:parameters","entityId":"epoch:249c0532cb4e60e6e3a4","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:eecd35c973c011bf339b:parameters","entityId":"epoch:eecd35c973c011bf339b","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:b8991944737cb69cdb11:training-compute","entityId":"epoch:b8991944737cb69cdb11","metric":"training-compute","value":2.96e+25,"unit":"FLOP","observationPeriod":"2024-08-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Estimate based on xAI statements comparing Grok-2 compute to GPT-4 and Grok-3. Full estimate here: https://docs.google.com/document/d/1C_dABuZrAqYE_ui4_GZ4bRLtq3TBjIGoBSktaPElhEU/edit?usp=sharing"},{"id":"epoch:bad9b5a160286e89c77d:parameters","entityId":"epoch:bad9b5a160286e89c77d","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-08-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"For OpenTable experiments, we use the LLaMA-3-70B-Instruct model as the initial policy"},{"id":"epoch:42beba81f08d09ea6a71:parameters","entityId":"epoch:42beba81f08d09ea6a71","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2024-08-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1837802f03ef2a081518:parameters","entityId":"epoch:1837802f03ef2a081518","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-08-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1837802f03ef2a081518:training-compute","entityId":"epoch:1837802f03ef2a081518","metric":"training-compute","value":3.9391101e+23,"unit":"FLOP","observationPeriod":"2024-08-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"989400000000000 FLOP / GPU / sec [bf16 assumed] * 1440 hours [see training time notes] * 3600 sec / hour * 256 GPUs * 0.3 [assumed utilization] = 3.9391101e+23 FLOP"},{"id":"epoch:9e99b4ba0f2365e7f91b:parameters","entityId":"epoch:9e99b4ba0f2365e7f91b","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-08-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:9e99b4ba0f2365e7f91b:training-compute","entityId":"epoch:9e99b4ba0f2365e7f91b","metric":"training-compute","value":7.20000786432e+23,"unit":"FLOP","observationPeriod":"2024-08-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Llama 3 compute, other models were also trained\nFinetune compute 6*8000000000*16384000=786432000000000000\nLlama 3 base: 7.2e+23\nTotal: 720000786432000000000000"},{"id":"epoch:b7f052dc292ba2ad85db:parameters","entityId":"epoch:b7f052dc292ba2ad85db","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-08-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:4ec7497d76de2b00129d:parameters","entityId":"epoch:4ec7497d76de2b00129d","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-08-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:56163aea4ded77d5f877:parameters","entityId":"epoch:56163aea4ded77d5f877","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-08-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:5e233641c2d1e4e03000:parameters","entityId":"epoch:5e233641c2d1e4e03000","metric":"parameters","value":350000000.0,"unit":"parameters","observationPeriod":"2024-08-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The Safetensors card in [1] lists the model size as 362M parameters, but the Limitations and Model details sections in [1], as well as the Abstract and Introduction sections in [2], specifically say the model size is 350M parameters. Since there are more sources saying the model has 350M parameters, I went with this figure.\n1. https://huggingface.co/OuteAI/OuteTTS-0.1-350M\n2. https://outeai.com/blog/outetts-0.1-350m"},{"id":"epoch:5e233641c2d1e4e03000:training-compute","entityId":"epoch:5e233641c2d1e4e03000","metric":"training-compute","value":6.3e+19,"unit":"FLOP","observationPeriod":"2024-08-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training compute\n= # of active parameters / forward pass * # of tokens * 6 FLOPS / token\n= 350e6 parameters * 30e9 tokens * 6 FLOPS / token\n= 63000e15 FLOPS\n= 6.3e19 FLOPS,\nusing the 6ND approximation and assuming the model architecture is dense."},{"id":"epoch:605eaaa53c930feceec1:parameters","entityId":"epoch:605eaaa53c930feceec1","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2024-08-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"To assess the performance of larger models, we note that we additionally fine-tuned ESM-2-3B for peptide generation (PepMLM-3B)"},{"id":"epoch:5bb421a2fbc335ff685f:parameters","entityId":"epoch:5bb421a2fbc335ff685f","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2024-08-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5f967c6218d2293bca6d:parameters","entityId":"epoch:5f967c6218d2293bca6d","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2024-08-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ba257fc6506f983f481d:parameters","entityId":"epoch:ba257fc6506f983f481d","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-08-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:13331eb43ee1577de344:parameters","entityId":"epoch:13331eb43ee1577de344","metric":"parameters","value":7200000000.0,"unit":"parameters","observationPeriod":"2024-08-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7 B active parameters in the decoder (inherits Qwen-2-7B-instruct) + 158 M in the Paraformer audio encoder \u2192 ~ 7.2 B total."},{"id":"epoch:4e1d53daca6b0dd60dc3:parameters","entityId":"epoch:4e1d53daca6b0dd60dc3","metric":"parameters","value":60200000.0,"unit":"parameters","observationPeriod":"2024-08-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"as in resnet-152"},{"id":"epoch:caeef4720a2af0d84ca9:parameters","entityId":"epoch:caeef4720a2af0d84ca9","metric":"parameters","value":7820000000.0,"unit":"parameters","observationPeriod":"2024-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.8B"},{"id":"epoch:caeef4720a2af0d84ca9:training-compute","entityId":"epoch:caeef4720a2af0d84ca9","metric":"training-compute","value":4.0000000000000003e+23,"unit":"FLOP","observationPeriod":"2024-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*7.8B parameters *8T tokens  = 3.744e+23 FLOP\n\n\"EXAONE language models were trained using Google Cloud Platform and a cluster powered by NVIDIA H100 GPUs and NVIDIA NeMo Framework. Then, they were optimized by NVIDIA TensorRT-LLM. The total amount of computation used for model training was about 4 \u00d7 1023 FLOPS\""},{"id":"epoch:1fb2360285ae8332cba1:parameters","entityId":"epoch:1fb2360285ae8332cba1","metric":"parameters","value":185000.0,"unit":"parameters","observationPeriod":"2024-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"17 low level controllers with 10k parameters each: \n\"Each policy is a dilated-gated CNN\n[22] following the architecture in [23] with 10k parameters... The final\nsystem contained 17 LLCs\"\n\nOne high-level controller with 4.5k parameters: \"The style policy architecture, similar to the LLC but with\nonly 4.5k parameters, has a (8, 128) observation space\"\n\nspin classifier that is a 2-layer MLP of hidden sizes (128, 64) and input size 18, which is 10k parameters per o1 and Claude.\n\nSo ~185k parameters total"},{"id":"epoch:55609d1e87139638a55e:parameters","entityId":"epoch:55609d1e87139638a55e","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-08-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8b"},{"id":"epoch:068659a4b8438d95f010:parameters","entityId":"epoch:068659a4b8438d95f010","metric":"parameters","value":7600000000.0,"unit":"parameters","observationPeriod":"2024-08-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:fd4575d9141de1af2025:parameters","entityId":"epoch:fd4575d9141de1af2025","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2024-08-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:fd4575d9141de1af2025:training-compute","entityId":"epoch:fd4575d9141de1af2025","metric":"training-compute","value":3.036551985824e+24,"unit":"FLOP","observationPeriod":"2024-08-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"FineTune: 6*72000000000*38314782000=1.655199e+22\n\nBase model: 3.02e+24\nTotal: 3.036552e+24"},{"id":"epoch:32e6385feaf55def7555:parameters","entityId":"epoch:32e6385feaf55def7555","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-08-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:32e6385feaf55def7555:training-compute","entityId":"epoch:32e6385feaf55def7555","metric":"training-compute","value":9.654e+23,"unit":"FLOP","observationPeriod":"2024-08-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8.1e+23 FLOP (base model training compute) + 1.554e+23 FLOP (finetune compute) = 9.654e+23 FLOP"},{"id":"epoch:c57d45e0b7601035f440:parameters","entityId":"epoch:c57d45e0b7601035f440","metric":"parameters","value":1100000000.0,"unit":"parameters","observationPeriod":"2024-08-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:32cf68d27a0e2822932e:parameters","entityId":"epoch:32cf68d27a0e2822932e","metric":"parameters","value":12000000000.0,"unit":"parameters","observationPeriod":"2024-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"All public FLUX.1 models are based on a hybrid architecture of multimodal and parallel diffusion transformer blocks and scaled to 12B parameters.\""},{"id":"epoch:c32d410dd7490d2effcb:parameters","entityId":"epoch:c32d410dd7490d2effcb","metric":"parameters","value":12000000000.0,"unit":"parameters","observationPeriod":"2024-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12 billion parameters, guidance-distilled from Flux.1 [pro]"},{"id":"epoch:559b95d21ee7f1888c9d:parameters","entityId":"epoch:559b95d21ee7f1888c9d","metric":"parameters","value":447149.0,"unit":"parameters","observationPeriod":"2024-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:559b95d21ee7f1888c9d:training-compute","entityId":"epoch:559b95d21ee7f1888c9d","metric":"training-compute","value":3.3696e+17,"unit":"FLOP","observationPeriod":"2024-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1*60*60*312000000000000*0.3=3.3696e+17"},{"id":"epoch:633be05500695671882c:parameters","entityId":"epoch:633be05500695671882c","metric":"parameters","value":20000000000.0,"unit":"parameters","observationPeriod":"2024-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"20b"},{"id":"epoch:969ac374f0b67c1508b2:parameters","entityId":"epoch:969ac374f0b67c1508b2","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-07-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"pretrained on top of the Llama3 base model that is 8 billion parameters\""},{"id":"epoch:969ac374f0b67c1508b2:training-compute","entityId":"epoch:969ac374f0b67c1508b2","metric":"training-compute","value":7.23e+23,"unit":"FLOP","observationPeriod":"2024-07-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Llama 3 base model: 7.2e+23\nSEA LION extended pre train: 2*24*60*60*64*989500000000000*0.3=3.28292352e+21\nTotal: 7.2328292e+23"},{"id":"epoch:849695ddf59a7449f8bf:parameters","entityId":"epoch:849695ddf59a7449f8bf","metric":"parameters","value":2730000000.0,"unit":"parameters","observationPeriod":"2024-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1, sum of non-embedding and embedding parameters"},{"id":"epoch:849695ddf59a7449f8bf:training-compute","entityId":"epoch:849695ddf59a7449f8bf","metric":"training-compute","value":4.5126e+23,"unit":"FLOP","observationPeriod":"2024-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Model was initialized from a pruned version of a 6.4B parameter model trained using the same recipe as AFM-server. Assuming \"same recipe\" involves training for the full 6.3T tokens, this implies 6 * 6.3T * 6.4B = 2.42e23 FLOP. \n\nThe pruning masks are learned by training over 188B tokens, which suggests 6 * 188B * 6.4B = 7.22e21 FLOPs.\n\nPretraining is then run over 6.3T tokens; however, labels are a convex combination of true labels and the predicted labels from the unpruned 6.4B model. Since this involves running the 6.3T tokens forward through both the 6.4B and the 2.73B model, but only calculating gradients for the smaller model, FLOPs here are equal to (6 * 6.3T * 2.73B) + (2 * 6.3T * 6.4B) = 1.84e23. \n\nFinally, there is a 1T \"continuation\" pretraining stage without distillation loss, for 6 * 1T * 2.73B = 1.64e22 FLOP, and a 100B context-lengthening stage for another 6 * 100B * 2.73B = 1.64e21 FLOP\n\nIn total: 2.42e23 + 7.22e21 + 1.84e23 + 1.64e22 + 1.64e21 = 4.51e23 FLOP"},{"id":"epoch:36f6d09db23de6432aa0:training-compute","entityId":"epoch:36f6d09db23de6432aa0","metric":"training-compute","value":4.3e+24,"unit":"FLOP","observationPeriod":"2024-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The AFM base models are dense decoder-only models that build on the\nTransformer architecture\"\n\n\"We train AFM-server from scratch for 6.3T tokens on 8192\nTPUv4 chips, using a sequence length of 4096 and a batch-size of 4096 sequences.\"\n\n\"For both models we perform continued pre-training at a sequence length of\n8192, with another 1T tokens from a mixture that upweights math and code,\nand down-weights the bulk web-crawl.\"\n\n\"The sustained model-flop-utilization (MFU) for this training run was approximately 52%.\"\n\nParameter count is not specified other than it being \"larger\" than 3 billion.\n\nCounting FLOP: Chinchilla scaling laws would suggest 7.3T / 20 = 365B parameters. \n\n365B parameters * 7.3T tokens * 6 ~= 1.6e25 FLOP.\n\nHowever, the attention to inference optimization in the technical report suggests a smaller size, even for this \"server\" model. One point of reference is Llama 3 70B being overtrained by a factor of 10. If this is true of AFM-server, the parameter count would be ~37B and training compute would be 1.6e24 FLOP.\n\nGPU-time: assume a wall-clock training time of 30 days based on the current trend value for notable models.\n\n8192 chips * 275e12 FLOP/s per chip * 0.52 utilization * 30 * 24 * 60 * 60 s ~= 3.0e24 FLOP\n\nThe geometric mean of these three estimates is 4.3e24 FLOP."},{"id":"epoch:10085d66eda73731fc36:parameters","entityId":"epoch:10085d66eda73731fc36","metric":"parameters","value":4500000.0,"unit":"parameters","observationPeriod":"2024-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:4cace3c2e98c2f8dc6a7:parameters","entityId":"epoch:4cace3c2e98c2f8dc6a7","metric":"parameters","value":224400000.0,"unit":"parameters","observationPeriod":"2024-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"sam2.1_hiera_large, 224.4M, https://github.com/facebookresearch/sam2"},{"id":"epoch:4cace3c2e98c2f8dc6a7:training-compute","entityId":"epoch:4cace3c2e98c2f8dc6a7","metric":"training-compute","value":1.24e+22,"unit":"FLOP","observationPeriod":"2024-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The released SAM 2 was trained on 256 A100 GPUs for 108 hours\"\n256 * 108 * 3600 * 3.12e14 * 0.40 = 1.24e22"},{"id":"epoch:69bb4be3fa0206c7b297:parameters","entityId":"epoch:69bb4be3fa0206c7b297","metric":"parameters","value":232000000.0,"unit":"parameters","observationPeriod":"2024-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" 232 million"},{"id":"epoch:69bb4be3fa0206c7b297:training-compute","entityId":"epoch:69bb4be3fa0206c7b297","metric":"training-compute","value":5.4310727e+21,"unit":"FLOP","observationPeriod":"2024-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*232000000.00*3901632700000 = 5.4310727e+21"},{"id":"epoch:8c80f157d82ac0ee77f7:parameters","entityId":"epoch:8c80f157d82ac0ee77f7","metric":"parameters","value":771000000.0,"unit":"parameters","observationPeriod":"2024-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"771 million"},{"id":"epoch:8c80f157d82ac0ee77f7:training-compute","entityId":"epoch:8c80f157d82ac0ee77f7","metric":"training-compute","value":1.2406517e+22,"unit":"FLOP","observationPeriod":"2024-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*771000000.00*2681910200000 = 1.2406517e+22"},{"id":"epoch:f2b412338ef652b70169:parameters","entityId":"epoch:f2b412338ef652b70169","metric":"parameters","value":141000000000.0,"unit":"parameters","observationPeriod":"2024-07-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"SaulLM-large is based on Mixtral-141B-Instruct, a Transformer model with a Mixture of Experts that selects only 2 out of 8 experts to process tokens for each layer [1]. \nhttps://arxiv.org/abs/2407.19584"},{"id":"epoch:d3acdf48e72cf04c5590:parameters","entityId":"epoch:d3acdf48e72cf04c5590","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2024-07-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Total Parameters: 3B\" from Supplemental Data Table 2: Prover network architecture hyperparameters."},{"id":"epoch:2c7a221f3c1d99f79d56:parameters","entityId":"epoch:2c7a221f3c1d99f79d56","metric":"parameters","value":123000000000.0,"unit":"parameters","observationPeriod":"2024-07-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f6477f9bef7f308889db:parameters","entityId":"epoch:f6477f9bef7f308889db","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-07-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:d3a0576ff56f01eedb26:parameters","entityId":"epoch:d3a0576ff56f01eedb26","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-07-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:d0a864019739ca4cd295:parameters","entityId":"epoch:d0a864019739ca4cd295","metric":"parameters","value":405000000000.0,"unit":"parameters","observationPeriod":"2024-07-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"405B"},{"id":"epoch:d0a864019739ca4cd295:training-compute","entityId":"epoch:d0a864019739ca4cd295","metric":"training-compute","value":3.8e+25,"unit":"FLOP","observationPeriod":"2024-07-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Stated in paper.\n\nAlso, 6 * 405B * 15.6T training tokens = 3.8e25"},{"id":"epoch:f308d164737f2eb230c4:parameters","entityId":"epoch:f308d164737f2eb230c4","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-07-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:f308d164737f2eb230c4:training-compute","entityId":"epoch:f308d164737f2eb230c4","metric":"training-compute","value":7.929e+24,"unit":"FLOP","observationPeriod":"2024-07-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Huggingface page says 3.1-70B used 7.0M H100 hours and trained over 15T tokens. https://huggingface.co/meta-llama/Llama-3.1-70B\nThe paper also says that 3.1-405B got MFU of between 38-43%; presumably 70B was around the same or a bit higher. I'll assume utilization of 40%\n\n6ND:\n6 * 15T * 70B = 6.3e24 FLOPs\n\nHardware:\n7M * 9.9e14 * 3600 * 0.4 = 9.98e24 FLOPs\n\nGeometric mean: sqrt(6.3e24 * 9.98e24) = 7.929e24\n\nNote that Llama 3-70B also said it used 15T tokens, but only 6.4M H100 hours. This suggests 3.1 might have used a bit more than 15T tokens.\n\nTraining compute upper bound: 7M H100-hours * 989 TFLOPS * 50% utilization = 1.25e25 FLOP"},{"id":"epoch:9e299770191e84ee79fc:parameters","entityId":"epoch:9e299770191e84ee79fc","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-07-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:9e299770191e84ee79fc:training-compute","entityId":"epoch:9e299770191e84ee79fc","metric":"training-compute","value":1.224e+24,"unit":"FLOP","observationPeriod":"2024-07-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Huggingface page says 3.1-8B used 1.46M H100 hours and trained over 15T tokens. https://huggingface.co/meta-llama/Llama-3.1-70B\nThe paper also says that 3.1-405B got MFU of between 38-43%; presumably 8B was around the same or a bit higher. I'll assume utilization of 40%\n\n6ND:\n6 * 15T * 8B = 7.2e23 FLOPs\n\nHardware:\n1.46M * 9.9e14 * 3600 * 0.4 = 2.08e24 FLOPs\n\nGeometric mean: sqrt(7.2e23 * 2.08e24) = 1.224e24\n\nNote that Llama 3-8B also said it used 15T tokens, but only 1.3M H100 hours. This suggests 3.1 might have used a bit more than 15T tokens."},{"id":"epoch:f5f364421191b151862f:parameters","entityId":"epoch:f5f364421191b151862f","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-07-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9f6464d649a5f2faada3:training-compute","entityId":"epoch:9f6464d649a5f2faada3","metric":"training-compute","value":1e+18,"unit":"FLOP","observationPeriod":"2024-07-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware: 1x NVIDIA A100 80GB GPU (3.12e14 FLOP/s fp16)\n2. Training duration: 6 hours (estimated based on dataset size and epochs)\n3. Utilization: 40%\n4. Calculation: 3.12e14 FLOP/s \u00d7 0.4 \u00d7 (6 \u00d7 3600s) = 2.69e18 FLOP \u2248 1.0e18 FLOP"},{"id":"epoch:8465517ed764bc8af410:parameters","entityId":"epoch:8465517ed764bc8af410","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-07-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:8465517ed764bc8af410:training-compute","entityId":"epoch:8465517ed764bc8af410","metric":"training-compute","value":1.05e+23,"unit":"FLOP","observationPeriod":"2024-07-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 7 * 10^9 parameters * 2.5 * 10^12 tokens = 1.05e+23 FLOP"},{"id":"epoch:48ff9ffb9e1d2b253b1f:parameters","entityId":"epoch:48ff9ffb9e1d2b253b1f","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-07-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:0093bfe8f7a33015cd4c:parameters","entityId":"epoch:0093bfe8f7a33015cd4c","metric":"parameters","value":341000000.0,"unit":"parameters","observationPeriod":"2024-07-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 4\n\nSwin-L [32] as an image backbone"},{"id":"epoch:94b529238a70d7b11396:parameters","entityId":"epoch:94b529238a70d7b11396","metric":"parameters","value":12000000000.0,"unit":"parameters","observationPeriod":"2024-07-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12b"},{"id":"epoch:949b11940ffed1db8b31:parameters","entityId":"epoch:949b11940ffed1db8b31","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-07-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2148f44e9122b636e5cb:parameters","entityId":"epoch:2148f44e9122b636e5cb","metric":"parameters","value":7285403648.0,"unit":"parameters","observationPeriod":"2024-07-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"This is an instructed model, with 7,285,403,648 parameters.\""},{"id":"epoch:cb27ffa4225bbb677451:parameters","entityId":"epoch:cb27ffa4225bbb677451","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-07-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:95bda5d63c6522fd78b9:parameters","entityId":"epoch:95bda5d63c6522fd78b9","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2024-07-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f0474b0dfae22df882cc:parameters","entityId":"epoch:f0474b0dfae22df882cc","metric":"parameters","value":1710000000.0,"unit":"parameters","observationPeriod":"2024-07-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The image here, https://huggingface.co/blog/smollm#hyperparameters-choice, shows that SmolLM-1.7B has 1.71B parameters."},{"id":"epoch:f0474b0dfae22df882cc:training-compute","entityId":"epoch:f0474b0dfae22df882cc","metric":"training-compute","value":1.03e+22,"unit":"FLOP","observationPeriod":"2024-07-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Therefore, we decided to train the 1.7B model on 1 trillion tokens\" (https://huggingface.co/blog/smollm#experiments). The model doesn't have any CNNs or RNNs (https://huggingface.co/blog/smollm#hyperparameters-choice), so I will approximate it to be dense. Then assuming the model saw 1 trillion tokens during training, the 6ND approximation yields\nTraining compute\n= # of active parameters / forward pass * # of tokens * 6 FLOPs / token\n= 1.71e9 parameters * 1e12 tokens * 6 FLOPs / token\n~= 1.03e22 FLOPS\n\n\"We also instruction tuned the models using publicly available permissive instruction datasets. We trained all three models for one epoch on the permissive subset of the WebInstructSub dataset, combined with StarCoder2-Self-OSS-Instruct. Following this, we performed DPO (Direct Preference Optimization) for one epoch: using HelpSteer for the 135M and 1.7B models... We followed the training parameters from the Zephyr-Gemma recipe in the alignment handbook, but adjusted the SFT (Supervised Fine-Tuning) learning rate to 3e-4\" (https://huggingface.co/blog/smollm#evaluation)."},{"id":"epoch:d6dbf9b23d4639ed589f:parameters","entityId":"epoch:d6dbf9b23d4639ed589f","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2024-07-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B, 14B, 100B parameter versions"},{"id":"epoch:3b47d3eb77d42c636666:parameters","entityId":"epoch:3b47d3eb77d42c636666","metric":"parameters","value":8200000000.0,"unit":"parameters","observationPeriod":"2024-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The total parameters of Qwen2-Audio is 8.2B parameters.\""},{"id":"epoch:8608e116f163bfbeb85d:parameters","entityId":"epoch:8608e116f163bfbeb85d","metric":"parameters","value":186000000.0,"unit":"parameters","observationPeriod":"2024-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:8608e116f163bfbeb85d:training-compute","entityId":"epoch:8608e116f163bfbeb85d","metric":"training-compute","value":3.3900690258459335e+20,"unit":"FLOP","observationPeriod":"2024-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 8x NVIDIA RTX 4090 GPUs (3.30\u00d710\u00b9\u2074 FLOP/s per GPU)\n\n2. Training duration: 3 weeks (directly provided)\n   = 1,814,400 seconds (3 weeks \u00d7 7 days \u00d7 24 hours \u00d7 3600 seconds)\n\n3. Utilization rate: 40%\n\n4. Final calculation:\n   3.30\u00d710\u00b9\u2074 FLOP/s \u00d7 8 GPUs \u00d7 1,814,400 seconds \u00d7 0.4 = 1.9\u00d710\u00b2\u00b9 FLOPs\n\nAlternative:\n6*54.2B*186M=60487200000000000000\n\nGeometric Mean: 339006902584593356472"},{"id":"epoch:f2eb78a08ca973543ef0:parameters","entityId":"epoch:f2eb78a08ca973543ef0","metric":"parameters","value":76000000000.0,"unit":"parameters","observationPeriod":"2024-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5fd43e8e5da5dac89a0f:training-compute","entityId":"epoch:5fd43e8e5da5dac89a0f","metric":"training-compute","value":9.6585178e+22,"unit":"FLOP","observationPeriod":"2024-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming (!) AMD Instinct MI300, bf16\n\n980600000000000 FLOP / GPU / sec *  91200 GPU-hours * 3600 sec / hour * 0.3 [assumed utilization] = 9.6585178e+22 FLOP \n\n\"Likely\" confidence, because exact AMD Instinct GPU model type is not reported"},{"id":"epoch:da97d42f3e77db7693b7:parameters","entityId":"epoch:da97d42f3e77db7693b7","metric":"parameters","value":2176.0,"unit":"parameters","observationPeriod":"2024-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Firstly, the input layer contains 5 neurons \"\n\"Therefore, we identified a single hidden layer of 128 neurons \nand an output layer of 12 neurons as the appropriate model architecture\"\n\nParameter: 5*128+128*12=2176"},{"id":"epoch:da97d42f3e77db7693b7:training-compute","entityId":"epoch:da97d42f3e77db7693b7","metric":"training-compute","value":39168000000.0,"unit":"FLOP","observationPeriod":"2024-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2*2176*3*15000*200=39168000000"},{"id":"epoch:68c0fdceb701d1ef8326:parameters","entityId":"epoch:68c0fdceb701d1ef8326","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2024-07-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Model Components:\nVision Encoder: SigLIP-So400m, a 400M parameter model pretrained with a contrastive objective.\nLanguage Model: Gemma-2B, a 2B parameter decoder-only model from the Gemma family of LLMs."},{"id":"epoch:68c0fdceb701d1ef8326:training-compute","entityId":"epoch:68c0fdceb701d1ef8326","metric":"training-compute","value":1.0652844e+22,"unit":"FLOP","observationPeriod":"2024-07-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"197000000000000 FLOP/s *256 GPUs *102 hours *3600 s / hour *0.55 [reported utilization] + 123000000000000 FLOP/s *32 GPUs *60 hours [for transfer] *3600 s / hour *0.55 [reported utilization]= 1.0652844e+22 FLOP"},{"id":"epoch:e4f422af5da262addf68:parameters","entityId":"epoch:e4f422af5da262addf68","metric":"parameters","value":150000000.0,"unit":"parameters","observationPeriod":"2024-07-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We conduct various experiments using a model with 150 million parameters on a language modeling task using the C4 dataset (Raffel et al., 2019). \""},{"id":"epoch:e4f422af5da262addf68:training-compute","entityId":"epoch:e4f422af5da262addf68","metric":"training-compute","value":4.152361e+19,"unit":"FLOP","observationPeriod":"2024-07-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 FLOP / token / parameter * 150*10^6 parameters * 46137344000 tokens [see dataset size notes] = 4.152361e+19 FLOP"},{"id":"epoch:2ad410f8d9dfa8dfc153:parameters","entityId":"epoch:2ad410f8d9dfa8dfc153","metric":"parameters","value":1100000000.0,"unit":"parameters","observationPeriod":"2024-07-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We adopt the same hyperparameters as TinyLlama (Zhang et al., 2024), employing a model with 1.1B parameters, a learning rate of 4e\u22124 and a batch size of 2048.\""},{"id":"epoch:2ad410f8d9dfa8dfc153:training-compute","entityId":"epoch:2ad410f8d9dfa8dfc153","metric":"training-compute","value":6.0901294e+20,"unit":"FLOP","observationPeriod":"2024-07-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 FLOP / token / parameter * 1.1*10^9 parameters * 92274688000 tokens [see dataset size notes] = 6.0901294e+20 FLOP\n\n4 replicas of 8xH100 (figure 8) -> 32 H100 GPUs"},{"id":"epoch:90e0170297d6de206b3a:parameters","entityId":"epoch:90e0170297d6de206b3a","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-07-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:90e0170297d6de206b3a:training-compute","entityId":"epoch:90e0170297d6de206b3a","metric":"training-compute","value":1.89e+23,"unit":"FLOP","observationPeriod":"2024-07-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 7 * 10^9 parameters * 4.5 * 10^12 tokens = 1.89e+23 FLOP"},{"id":"epoch:9cc219db09598f76c8c4:parameters","entityId":"epoch:9cc219db09598f76c8c4","metric":"parameters","value":600000000000.0,"unit":"parameters","observationPeriod":"2024-07-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://en.tmtpost.com/post/7159076"},{"id":"epoch:c99f8bce65547e4c13ed:parameters","entityId":"epoch:c99f8bce65547e4c13ed","metric":"parameters","value":89000000.0,"unit":"parameters","observationPeriod":"2024-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:403284f779be1059580f:parameters","entityId":"epoch:403284f779be1059580f","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2024-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:7ad7ad472dac5713f99e:parameters","entityId":"epoch:7ad7ad472dac5713f99e","metric":"parameters","value":25500000000.0,"unit":"parameters","observationPeriod":"2024-07-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"25.5B"},{"id":"epoch:242e287e9014902f026f:parameters","entityId":"epoch:242e287e9014902f026f","metric":"parameters","value":40100000000.0,"unit":"parameters","observationPeriod":"2024-07-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"40.1B"},{"id":"epoch:6b85a1c45a07ce30eda4:parameters","entityId":"epoch:6b85a1c45a07ce30eda4","metric":"parameters","value":82000000.0,"unit":"parameters","observationPeriod":"2024-07-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d58783562ac1b0c76b93:parameters","entityId":"epoch:d58783562ac1b0c76b93","metric":"parameters","value":34000000000.0,"unit":"parameters","observationPeriod":"2024-07-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Smaug-34B is a modified version of the base model Bagel-34B-v0.2 [Durbin, 2024a], which itself is a SFT\nversion of Yi-34B-200k [01.AI, 2024]"},{"id":"epoch:9874c2e43a8bfb79caf9:parameters","entityId":"epoch:9874c2e43a8bfb79caf9","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2024-07-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"For 72B, we start from MoMo-72b-lora-1.8.7-DPO [Moreh, 2024], which itself is a fine-tune of Qwen72B [Bai et al., 2023]. MoMo-72b-lora-1.8.7-DPO has already undergone SFT, so we simply run the DPOP\nroutines as in Smaug-34B."},{"id":"epoch:e9c196f5db0c0c542ba7:parameters","entityId":"epoch:e9c196f5db0c0c542ba7","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We performed instruction tuning in English and also in ASEAN languages such as Indonesian, Thai and Vietnamese on our continued pre-trained Llama-SEA-LION-v2-8B, a decoder model using the Llama3 architecture, to create Llama-SEA-LION-v2-8B-IT\" (https://huggingface.co/aisingapore/Llama-SEA-LION-v2-8B-IT)."},{"id":"epoch:7cc806837331e0f4bb23:parameters","entityId":"epoch:7cc806837331e0f4bb23","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-06-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming as in LLaMA-13 since it is its finetune"},{"id":"epoch:ef9dee46e2ff5db168d4:parameters","entityId":"epoch:ef9dee46e2ff5db168d4","metric":"parameters","value":1900000000.0,"unit":"parameters","observationPeriod":"2024-06-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:06e9981815aa5ca73298:parameters","entityId":"epoch:06e9981815aa5ca73298","metric":"parameters","value":98500000000.0,"unit":"parameters","observationPeriod":"2024-06-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"98.5 billion (Table S1)"},{"id":"epoch:06e9981815aa5ca73298:training-compute","entityId":"epoch:06e9981815aa5ca73298","metric":"training-compute","value":1.07e+24,"unit":"FLOP","observationPeriod":"2024-06-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"ESM3 at its largest scale was trained with 1.07\u00d710^24 FLOPs on 2.78 billion proteins and 771 billion unique tokens, and has 98 billion parameters.\"\n\nper Table 1, trained 98B model on 1.8T training tokens. 98 billion * 1800 billion * 6 = 1.06e24. Likely some rounding, so will go with developer's reported count."},{"id":"epoch:e824bd79778c04a84fff:parameters","entityId":"epoch:e824bd79778c04a84fff","metric":"parameters","value":1400000000.0,"unit":"parameters","observationPeriod":"2024-06-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.4 billion"},{"id":"epoch:e824bd79778c04a84fff:training-compute","entityId":"epoch:e824bd79778c04a84fff","metric":"training-compute","value":2.7e+21,"unit":"FLOP","observationPeriod":"2024-06-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"see Table S1. Not clear which 1.4B model is the one that they released, could be either 6.7e20 or 2.7e21."},{"id":"epoch:eb7009ffb5a3953d2274:training-compute","entityId":"epoch:eb7009ffb5a3953d2274","metric":"training-compute","value":1.26e+21,"unit":"FLOP","observationPeriod":"2024-06-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.26*10^21 from the figure 1\n\nthere is slight disrepancy (within confidence interval) between sigclip/flexi-jest++ training compute ratio if we compare epoch's estimated compute and table 1"},{"id":"epoch:5f6e6bc88e93bce257c7:training-compute","entityId":"epoch:5f6e6bc88e93bce257c7","metric":"training-compute","value":1.9e+21,"unit":"FLOP","observationPeriod":"2024-06-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.9*10^21 from figure 1 and  table 1\n\nthere is slight disrepancy (within confidence interval) between sigclip/jest++ training compute ratio if we compare epoch's estimated compute and table 1"},{"id":"epoch:5317c868d9c8b76a7adb:training-compute","entityId":"epoch:5317c868d9c8b76a7adb","metric":"training-compute","value":2e+21,"unit":"FLOP","observationPeriod":"2024-06-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See the right subfigure of Figure 1 on page 2 of https://arxiv.org/pdf/2406.17711v1."},{"id":"epoch:4115ee4b3408de347f9e:parameters","entityId":"epoch:4115ee4b3408de347f9e","metric":"parameters","value":9000000000.0,"unit":"parameters","observationPeriod":"2024-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:4115ee4b3408de347f9e:training-compute","entityId":"epoch:4115ee4b3408de347f9e","metric":"training-compute","value":4.32e+23,"unit":"FLOP","observationPeriod":"2024-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For the 9B model, we train on an 8x16x32 configuration of TPUv4, totaling 4096 chips\"\n\n6ND = 6 FLOP / token / parameter * 9000000000 parameters * 8000000000000 tokens = 4.32e+23 FLOP"},{"id":"epoch:a879844bbcb73ece0b37:parameters","entityId":"epoch:a879844bbcb73ece0b37","metric":"parameters","value":27000000000.0,"unit":"parameters","observationPeriod":"2024-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a879844bbcb73ece0b37:training-compute","entityId":"epoch:a879844bbcb73ece0b37","metric":"training-compute","value":2.106e+24,"unit":"FLOP","observationPeriod":"2024-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For the 27B model, we train on an 8x24x32 configuration of\nTPUv5p, totaling 6144 chips\"\n\ntrained on 13T tokens\n\n6ND = 6*27000000000*13000000000000=2.106e+24"},{"id":"epoch:67eebbd79742b955fb28:parameters","entityId":"epoch:67eebbd79742b955fb28","metric":"parameters","value":2600000000.0,"unit":"parameters","observationPeriod":"2024-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:67eebbd79742b955fb28:training-compute","entityId":"epoch:67eebbd79742b955fb28","metric":"training-compute","value":3.12e+22,"unit":"FLOP","observationPeriod":"2024-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For the 2.6B model, we train on a 2x16x16 configuration of TPUv5e, totaling 512 chips\"\n\n6ND = 6 FLOP / token / parameter * 2600000000 parameters * 2000000000000 tokens = 3.12e+22 FLOP"},{"id":"epoch:d88a799ca43e5fece941:parameters","entityId":"epoch:d88a799ca43e5fece941","metric":"parameters","value":2900000.0,"unit":"parameters","observationPeriod":"2024-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.9M transformer model"},{"id":"epoch:cf48e846282bfa0bf7d7:parameters","entityId":"epoch:cf48e846282bfa0bf7d7","metric":"parameters","value":34000000000.0,"unit":"parameters","observationPeriod":"2024-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:dfbdfd6147a0a6346033:parameters","entityId":"epoch:dfbdfd6147a0a6346033","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:797fcffdf8fc96881232:parameters","entityId":"epoch:797fcffdf8fc96881232","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:38c172e384f0f7e9c387:training-compute","entityId":"epoch:38c172e384f0f7e9c387","metric":"training-compute","value":2.700000000000001e+25,"unit":"FLOP","observationPeriod":"2024-06-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Blog post by Dario Amodei includes some info on 3.5 Sonnet compute: https://darioamodei.com/on-deepseek-and-export-controls\n\"Claude 3.5 Sonnet is a mid-sized model that cost a few $10M's to train (I won't give an exact number). Also, 3.5 Sonnet was not trained in any way that involved a larger or more expensive model (contrary to some rumors).\"\n\nUsing assumptions about GPU pricing, this lets us estimate compute. https://docs.google.com/spreadsheets/d/1-p-ab6t6dkUM6T7GwnFp85ePTMpZMW7LFY7fW2t8POs/"},{"id":"epoch:69113460452a40758ce8:parameters","entityId":"epoch:69113460452a40758ce8","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-06-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70b"},{"id":"epoch:37d97a65b1a32866c108:parameters","entityId":"epoch:37d97a65b1a32866c108","metric":"parameters","value":16800000.0,"unit":"parameters","observationPeriod":"2024-06-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:37d97a65b1a32866c108:training-compute","entityId":"epoch:37d97a65b1a32866c108","metric":"training-compute","value":3.6889344e+18,"unit":"FLOP","observationPeriod":"2024-06-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4*35580000000000*0.4*18hours=3688934400000000000"},{"id":"epoch:0c8ac953acd51bc62f5b:parameters","entityId":"epoch:0c8ac953acd51bc62f5b","metric":"parameters","value":9000000000.0,"unit":"parameters","observationPeriod":"2024-06-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9B"},{"id":"epoch:c6baaf34f247982d6a21:parameters","entityId":"epoch:c6baaf34f247982d6a21","metric":"parameters","value":236000000000.0,"unit":"parameters","observationPeriod":"2024-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Mixture of experts model. 21B parameters activated per token."},{"id":"epoch:c6baaf34f247982d6a21:training-compute","entityId":"epoch:c6baaf34f247982d6a21","metric":"training-compute","value":1.2852e+24,"unit":"FLOP","observationPeriod":"2024-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Trained on a total of 10.2T tokens\n6NC: 6 * 10.2T * 21B active parameters = 1.285e24"},{"id":"epoch:129eec9c479f29a37420:parameters","entityId":"epoch:129eec9c479f29a37420","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:129eec9c479f29a37420:training-compute","entityId":"epoch:129eec9c479f29a37420","metric":"training-compute","value":1.7e+23,"unit":"FLOP","observationPeriod":"2024-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Fine tune: 989500000000000*128*0.3*15*60*60=2.0518272e+21\nQWEN 1.5 training FLOP: 1.68e+23\nTotal: 1.7e+23"},{"id":"epoch:40087617b34e7bf99226:parameters","entityId":"epoch:40087617b34e7bf99226","metric":"parameters","value":38800000000.0,"unit":"parameters","observationPeriod":"2024-06-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"38.8B - total parameters\n13B - activated parameters"},{"id":"epoch:40087617b34e7bf99226:training-compute","entityId":"epoch:40087617b34e7bf99226","metric":"training-compute","value":4.3596131e+23,"unit":"FLOP","observationPeriod":"2024-06-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 13 * 10^9 activated parameters * 5 * 10^12 tokens = 3.9e+23 FLOP\n\nNvidia A800 and Ascend 910B chips were used\n\n(376000000000000 FLOP / sec / GPU + 312000000000000 FLOP / sec / GPU) * 0.5 [assuming A800/910B 50/50] * 1464 hours [see training time notes] * 3600 sec / hour * 896 GPUs * 0.3 [assumed utilization] = 4.8733913e+23 FLOP\n\nsqrt(3.9e+23 * 4.8733913e+23) = 4.3596131e+23 FLOP"},{"id":"epoch:5561b57a82d844471b14:parameters","entityId":"epoch:5561b57a82d844471b14","metric":"parameters","value":340000000000.0,"unit":"parameters","observationPeriod":"2024-06-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"340B"},{"id":"epoch:5561b57a82d844471b14:training-compute","entityId":"epoch:5561b57a82d844471b14","metric":"training-compute","value":1.8e+25,"unit":"FLOP","observationPeriod":"2024-06-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9 trillion tokens for training\n6 * 340B * 9T = 1.8E25\n\nalternatively, can do a hardware estimate with a few extra steps:\n\nAccording to the technical report, Nemotron-4 340B was trained using up to 6144 H100 GPUs. Helpfully, they also report the model FLOP utilization (MFU), which was 41-42% (Table 2). This is the ratio of the actual output of their GPUs, in FLOP used for training, relative to their theoretical max of 989 teraFLOP/s per GPU. \nUnfortunately, the report omits the last ingredient, which is the duration of the training run. However, in Table 2 they report some relevant data that we can use to infer the training time. \nNemotron-4 was trained in several stages, but the largest stage used all 6144 GPUs with a batch size of 2304 and an iteration time (time per batch) of 8.0 seconds. This stage involved 7.6T tokens, so it makes up the majority of training. \nA batch size of 2304 means that each batch consists of 2304 sequences, and they report that the sequence length used for training was 4096 tokens. This means that each batch contained 4096 * 2304 = 9,437,184 tokens. \nSo, during this stage, it took 8 seconds to train the model on 9.4m tokens. Extrapolating to the entire 9T token dataset, this implies the training run would have taken 7,659,574 seconds, or 89 days. (it actually took longer because they didn't use all their GPUs for the whole run) \nMultiplying 7,659,574 seconds by 41% MFU, 989 peak teraFLOP/s for each H100, and 6144 H100s, we get ~1.9e25 FLOP. This is very close to our first estimate. \n"},{"id":"epoch:530a5c15c8c3d55d43e0:parameters","entityId":"epoch:530a5c15c8c3d55d43e0","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2024-06-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:530a5c15c8c3d55d43e0:training-compute","entityId":"epoch:530a5c15c8c3d55d43e0","metric":"training-compute","value":1.2e+24,"unit":"FLOP","observationPeriod":"2024-06-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*100B*2T=1.2e24"},{"id":"epoch:54712f3cd3f2e280e9b6:parameters","entityId":"epoch:54712f3cd3f2e280e9b6","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2024-06-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9f392e3e736f95211984:parameters","entityId":"epoch:9f392e3e736f95211984","metric":"parameters","value":7188100000.0,"unit":"parameters","observationPeriod":"2024-06-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Based on a Prismatic-7B VLM backbone, which itself is comprised of 600M parameter vision encoder (DinoV2 + SigLIP) plus Llama-2 7B. Table 1 indicates 7.1881 billion trainable parameters"},{"id":"epoch:9f392e3e736f95211984:training-compute","entityId":"epoch:9f392e3e736f95211984","metric":"training-compute","value":1.1e+23,"unit":"FLOP","observationPeriod":"2024-06-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Majority of compute is from VLA pre-training embedded in Prismatic-7B and it's constituent models. \n\nThe fine-tuning compute used in this paper is \"64 A100 GPUs for 14 days, or a total of 21,500 A100-hours\"\n21500 * 3600 * 3.12e14 * 0.4 = 9.66e21\n\nPrismatic-7B training took \"less than 9 hours\" on 8 A100s: 9 * 3600 * 8 * 3.12e14 * 0.4 = 3.23e19\n\nAdd in the pre-trained components:\n- DinoV2 = 7.42e21, per our database\n- The SigLIP model in question is SoViT-400m/14 from the cited Alabdulmohsin et al., 2023) and \"is pretrained on 40 billion examples, which amounts to 9T GFLOPs and 230K TPUv3 core-hours\" = 9e21\n- Llama 2-7B = 8.4e22, per our database\n\nTotal\n9.66e21 + 3.23e19 + 7.42e21 + 9e21 + 8.4e22 = 1.10e23"},{"id":"epoch:0328b6fb4ba2e8f6c88e:parameters","entityId":"epoch:0328b6fb4ba2e8f6c88e","metric":"parameters","value":8660000000.0,"unit":"parameters","observationPeriod":"2024-06-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 6"},{"id":"epoch:0328b6fb4ba2e8f6c88e:training-compute","entityId":"epoch:0328b6fb4ba2e8f6c88e","metric":"training-compute","value":1.8186e+23,"unit":"FLOP","observationPeriod":"2024-06-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*8660000000.00 parameters * 3500000000000 tokens = 1.8186 \u00d7 10^23\n\n"},{"id":"epoch:0a994c5a92980fb94d40:parameters","entityId":"epoch:0a994c5a92980fb94d40","metric":"parameters","value":2500000000.0,"unit":"parameters","observationPeriod":"2024-06-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.5B"},{"id":"epoch:b2802f66e0be7a875780:parameters","entityId":"epoch:b2802f66e0be7a875780","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2024-06-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b2802f66e0be7a875780:training-compute","entityId":"epoch:b2802f66e0be7a875780","metric":"training-compute","value":3.6e+22,"unit":"FLOP","observationPeriod":"2024-06-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 3000000000 parameters * 2000000000000 tokens = 3.6e+22 FLOP"},{"id":"epoch:1f6e0d0e31a614e74d08:parameters","entityId":"epoch:1f6e0d0e31a614e74d08","metric":"parameters","value":3800000000.0,"unit":"parameters","observationPeriod":"2024-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.8B, table 10"},{"id":"epoch:1f6e0d0e31a614e74d08:training-compute","entityId":"epoch:1f6e0d0e31a614e74d08","metric":"training-compute","value":7.3e+22,"unit":"FLOP","observationPeriod":"2024-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We scale SAMBA with 421M, 1.3B, 1.7B and up to 3.8B parameters. In particular, the largest 3.8B base model pre-trained with 3.2T tokens achieves a 71.2 score for MMLU [HBB+21], 54.9 for HumanEval [CTJ+21], and 69.6 for GSM8K\"\n\n3.8B * 3.2T * 6 = 7.3e22"},{"id":"epoch:4e8aa0d9c7f69a4e78a6:parameters","entityId":"epoch:4e8aa0d9c7f69a4e78a6","metric":"parameters","value":307000000.0,"unit":"parameters","observationPeriod":"2024-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For TiTok variants, we primarily investigate three model sizes\u2014small, base, and large (i.e., TiTok-S, TiTok-B, TiTok-L)\u2014comprising 22M, 86M, and 307M parameters for encoder and decoder, respectively\""},{"id":"epoch:4e8aa0d9c7f69a4e78a6:training-compute","entityId":"epoch:4e8aa0d9c7f69a4e78a6","metric":"training-compute","value":1.7252352e+21,"unit":"FLOP","observationPeriod":"2024-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000*5120*3600*0.3 = 1.7252352e+21"},{"id":"epoch:9e2268b103d2c804cb77:parameters","entityId":"epoch:9e2268b103d2c804cb77","metric":"parameters","value":72710000000.0,"unit":"parameters","observationPeriod":"2024-06-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"72.71B parameters in total, of which 70.21B are non-embedding parameters"},{"id":"epoch:9e2268b103d2c804cb77:training-compute","entityId":"epoch:9e2268b103d2c804cb77","metric":"training-compute","value":3.02e+24,"unit":"FLOP","observationPeriod":"2024-06-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"72 billion params, 7 trillion tokens\n\n6 * 72 billion * 7 trillion ~= 3.02e24"},{"id":"epoch:2b92ee16180462a83967:parameters","entityId":"epoch:2b92ee16180462a83967","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-06-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B parameters (table 1)"},{"id":"epoch:2b92ee16180462a83967:training-compute","entityId":"epoch:2b92ee16180462a83967","metric":"training-compute","value":2.9400000000001e+23,"unit":"FLOP","observationPeriod":"2024-06-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7 billion params, 7 trillion tokens\n\n6 FLOP * 7 billion * 7 trillion ~= 2.94e23 FLOP"},{"id":"epoch:3268f05e4cd25c85ec7a:parameters","entityId":"epoch:3268f05e4cd25c85ec7a","metric":"parameters","value":57000000000.0,"unit":"parameters","observationPeriod":"2024-06-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"57B parameters (table 1)"},{"id":"epoch:3268f05e4cd25c85ec7a:training-compute","entityId":"epoch:3268f05e4cd25c85ec7a","metric":"training-compute","value":3.7800000000001e+23,"unit":"FLOP","observationPeriod":"2024-06-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For MoE models, 57B-A14B denotes that the model has 57B parameters in total and for each token 14B parameters are active\" (page 5)\n\nC ~= 6 FLOP * 14e9 * 4.5e12 = 3.78e23"},{"id":"epoch:33642105678947520beb:parameters","entityId":"epoch:33642105678947520beb","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2024-06-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.5B"},{"id":"epoch:33642105678947520beb:training-compute","entityId":"epoch:33642105678947520beb","metric":"training-compute","value":6.3e+22,"unit":"FLOP","observationPeriod":"2024-06-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 1.5 * 10^9  parameters * 7 * 10^12 tokens = 6.3e+22 FLOP"},{"id":"epoch:c08aafaa690c142ed67d:parameters","entityId":"epoch:c08aafaa690c142ed67d","metric":"parameters","value":500000000.0,"unit":"parameters","observationPeriod":"2024-06-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c08aafaa690c142ed67d:training-compute","entityId":"epoch:c08aafaa690c142ed67d","metric":"training-compute","value":3.6e+22,"unit":"FLOP","observationPeriod":"2024-06-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 500000000 parameters * 12000000000000 tokens = 3.6e+22 FLOP"},{"id":"epoch:45951890b1480a7acc85:parameters","entityId":"epoch:45951890b1480a7acc85","metric":"parameters","value":930000000.0,"unit":"parameters","observationPeriod":"2024-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:4ff67455ff8ca364a4a0:parameters","entityId":"epoch:4ff67455ff8ca364a4a0","metric":"parameters","value":650000000.0,"unit":"parameters","observationPeriod":"2024-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:8bf55fe38e501adbb73d:parameters","entityId":"epoch:8bf55fe38e501adbb73d","metric":"parameters","value":2442057984.0,"unit":"parameters","observationPeriod":"2024-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:8bf55fe38e501adbb73d:training-compute","entityId":"epoch:8bf55fe38e501adbb73d","metric":"training-compute","value":1.584e+22,"unit":"FLOP","observationPeriod":"2024-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 2.4*10^9 parameters * 1.1*10^12 tokens = 1.584e+22 FLOP"},{"id":"epoch:03ba73083c16dd0c008e:parameters","entityId":"epoch:03ba73083c16dd0c008e","metric":"parameters","value":4000000000.0,"unit":"parameters","observationPeriod":"2024-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4b"},{"id":"epoch:27c827ca771331a0059a:parameters","entityId":"epoch:27c827ca771331a0059a","metric":"parameters","value":1247442432.0,"unit":"parameters","observationPeriod":"2024-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:27c827ca771331a0059a:training-compute","entityId":"epoch:27c827ca771331a0059a","metric":"training-compute","value":7.92e+21,"unit":"FLOP","observationPeriod":"2024-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 1.2*10^9 parameters * 1.1*10^12 tokens = 7.92e+21 FLOP"},{"id":"epoch:6b66973f0e07d9fe7225:parameters","entityId":"epoch:6b66973f0e07d9fe7225","metric":"parameters","value":35000000.0,"unit":"parameters","observationPeriod":"2024-06-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6b66973f0e07d9fe7225:training-compute","entityId":"epoch:6b66973f0e07d9fe7225","metric":"training-compute","value":5.1777408e+20,"unit":"FLOP","observationPeriod":"2024-06-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Finetune: 989500000000000*9days*0.4=307774080000000000000\n\nBase model: 209999999999999970000.00\n\nTotal: 209999999999999970000.00+307774080000000000000=517774080000000000000"},{"id":"epoch:5edf297098b9df41f955:training-compute","entityId":"epoch:5edf297098b9df41f955","metric":"training-compute","value":6.469e+19,"unit":"FLOP","observationPeriod":"2024-06-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"48*60*60*4*312000000000000*0.3=6.469632e+19"},{"id":"epoch:100d46dce9458e5e71a4:parameters","entityId":"epoch:100d46dce9458e5e71a4","metric":"parameters","value":300000000.0,"unit":"parameters","observationPeriod":"2024-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:14966091e08dfa23e5c1:parameters","entityId":"epoch:14966091e08dfa23e5c1","metric":"parameters","value":2600000000.0,"unit":"parameters","observationPeriod":"2024-05-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.6B (Table 4)"},{"id":"epoch:c9ac5cd6e3ceba89e724:parameters","entityId":"epoch:c9ac5cd6e3ceba89e724","metric":"parameters","value":2700000000.0,"unit":"parameters","observationPeriod":"2024-05-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.7B"},{"id":"epoch:c9ac5cd6e3ceba89e724:training-compute","entityId":"epoch:c9ac5cd6e3ceba89e724","metric":"training-compute","value":4.86e+21,"unit":"FLOP","observationPeriod":"2024-05-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*2.7B*300B = 4.86\u00d710^21"},{"id":"epoch:e9f7c978022374661a45:parameters","entityId":"epoch:e9f7c978022374661a45","metric":"parameters","value":20000000000.0,"unit":"parameters","observationPeriod":"2024-05-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e9f7c978022374661a45:training-compute","entityId":"epoch:e9f7c978022374661a45","metric":"training-compute","value":3.0000000000001e+23,"unit":"FLOP","observationPeriod":"2024-05-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*2500000000000*20000000000=3e+23"},{"id":"epoch:293f3b572f232d12a719:parameters","entityId":"epoch:293f3b572f232d12a719","metric":"parameters","value":14158450.0,"unit":"parameters","observationPeriod":"2024-05-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\" with our model comprising 8 layers and 12 heads. \"\n\"The vocabulary of sgRNAGen encompasses383\nseven tokens: A, T, C, G, and special tokens such as [PAD], [MASK], and [UNK]\"\n\"The hidden size386\nwas set as 384\"\n\n8 layers, 12 heads, vocab of 7, d_model of 384\nd_head=384/12=32\nAttention parameters: 12*(384*(2*32+32)+32*384)=589824\nMLP = 384*4*384*2= 1179648\n\nTotal: 7*382+8*(589824+1179648)=14158450\n\n"},{"id":"epoch:1344ef0634478db5b3af:training-compute","entityId":"epoch:1344ef0634478db5b3af","metric":"training-compute","value":7.646000000001e+21,"unit":"FLOP","observationPeriod":"2024-05-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 2x NVIDIA A100 40GB GPUs (3.12 x 10^14 FLOP/s per GPU)\n2. Training duration: 4 days (directly provided) = 345,600 seconds\n3. Utilization rate: 40%\n4. Calculation: 2 GPUs \u00d7 3.12\u00d710^14 FLOP/s \u00d7 345,600s \u00d7 0.4 = 8.6\u00d710^19 FLOPs\n\nBase model: 7.560000000001e+21\nTotal: 7646000000001000000000"},{"id":"epoch:86cea1d2f4e491505cd7:parameters","entityId":"epoch:86cea1d2f4e491505cd7","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2024-05-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:86cea1d2f4e491505cd7:training-compute","entityId":"epoch:86cea1d2f4e491505cd7","metric":"training-compute","value":3.1054234e+22,"unit":"FLOP","observationPeriod":"2024-05-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000 FLOP / GPU / sec [A800 reported] * 256 GPUs * 360 hours * 3600 sec / hour * 0.3 [assumed utilization] = 3.1054234e+22 FLOP"},{"id":"epoch:4d87bd15040f2ee2f06f:parameters","entityId":"epoch:4d87bd15040f2ee2f06f","metric":"parameters","value":22200000000.0,"unit":"parameters","observationPeriod":"2024-05-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"22.2B from hugging face model card"},{"id":"epoch:a501b2f9bbecfbf071ed:parameters","entityId":"epoch:a501b2f9bbecfbf071ed","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-05-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"same as base model"},{"id":"epoch:24aacfb5486a3b0bb6d6:parameters","entityId":"epoch:24aacfb5486a3b0bb6d6","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2024-05-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:24aacfb5486a3b0bb6d6:training-compute","entityId":"epoch:24aacfb5486a3b0bb6d6","metric":"training-compute","value":4.5287424e+21,"unit":"FLOP","observationPeriod":"2024-05-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The model is trained using bf16 mixed\nprecision.\"\n\n312000000000000*420*32*3600*.3 =  4528742400000000000000 FLOP\n\n"},{"id":"epoch:5172362981ecfc6ace20:parameters","entityId":"epoch:5172362981ecfc6ace20","metric":"parameters","value":15800000000.0,"unit":"parameters","observationPeriod":"2024-05-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/Nanbeige/Nanbeige2-16B-Chat"},{"id":"epoch:5172362981ecfc6ace20:training-compute","entityId":"epoch:5172362981ecfc6ace20","metric":"training-compute","value":4.05e+23,"unit":"FLOP","observationPeriod":"2024-05-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The model has 15.8B parameters and was trained on 4.5T tokens during the training phrase (https://huggingface.co/Nanbeige/Nanbeige2-16B-Chat). It appears to be entirely transformer-based (https://huggingface.co/Nanbeige/Nanbeige2-16B-Chat/blob/main/modeling_nanbeige.py).\n\nAssuming training was done for 1 epoch, the 6ND approximation yields\nTraining compute\n= # of active parameters / forward pass * # of tokens * 6 FLOPS / token\n= 1.5e10 parameters * 4.5e12 tokens * 6 FLOPS / token\n= 4.05e23 FLOPS"},{"id":"epoch:0b9a18593f01845eecbe:parameters","entityId":"epoch:0b9a18593f01845eecbe","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-05-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:60e10b7e6b997b9005a0:parameters","entityId":"epoch:60e10b7e6b997b9005a0","metric":"parameters","value":21400000000.0,"unit":"parameters","observationPeriod":"2024-05-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"21.4B "},{"id":"epoch:b4e6463c49a07cf64de6:parameters","entityId":"epoch:b4e6463c49a07cf64de6","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-05-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b4e6463c49a07cf64de6:training-compute","entityId":"epoch:b4e6463c49a07cf64de6","metric":"training-compute","value":8.82e+22,"unit":"FLOP","observationPeriod":"2024-05-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 7 * 10^9 parameters * 2100000000000 tokens = 8.82e+22 FLOP\n\n[assuming 1 epoch]"},{"id":"epoch:be93db6c037addcf2354:parameters","entityId":"epoch:be93db6c037addcf2354","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-05-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2161666627f3cbb3ee2e:parameters","entityId":"epoch:2161666627f3cbb3ee2e","metric":"parameters","value":15700000.0,"unit":"parameters","observationPeriod":"2024-05-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:2161666627f3cbb3ee2e:training-compute","entityId":"epoch:2161666627f3cbb3ee2e","metric":"training-compute","value":3.234816e+20,"unit":"FLOP","observationPeriod":"2024-05-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train Genie 2 using data parallelism on 8 Nvidia A100 GPUs with an effective batch size of 48. We train the model for 40 epochs (\u223c5 days) for a total of \u223c960 GPU hours.\"\n\n312000000000000 peak FLOP/s * 960 GPU hours * 3600 s * 0.3 assumed utilization = 3.234816e+20 "},{"id":"epoch:639c2fd452f46bbb45c0:parameters","entityId":"epoch:639c2fd452f46bbb45c0","metric":"parameters","value":344000000.0,"unit":"parameters","observationPeriod":"2024-05-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f89279ea82f774aabd19:training-compute","entityId":"epoch:f89279ea82f774aabd19","metric":"training-compute","value":2.3e+17,"unit":"FLOP","observationPeriod":"2024-05-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3600 * 2 * (91.1 * 10^12) * 0.35 = 2.3e17"},{"id":"epoch:48f2101fc827e01a5167:parameters","entityId":"epoch:48f2101fc827e01a5167","metric":"parameters","value":29500000.0,"unit":"parameters","observationPeriod":"2024-05-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"29.5M"},{"id":"epoch:48f2101fc827e01a5167:training-compute","entityId":"epoch:48f2101fc827e01a5167","metric":"training-compute","value":1.478888e+17,"unit":"FLOP","observationPeriod":"2024-05-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*29500000.00*118000*500 = 1.0443e+16 (likely an underestimation)\n\n(speculative confidence because I don't know the amount of tokens per image)\n\nassuming(!) batch size 64 -> number of updates per epoch = 118000/64 = 1844\n\n160.4 Gigaflops * 1844 * 500 = 1.478888e+17"},{"id":"epoch:a3663118e97600ccce82:parameters","entityId":"epoch:a3663118e97600ccce82","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-05-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a3663118e97600ccce82:training-compute","entityId":"epoch:a3663118e97600ccce82","metric":"training-compute","value":1.428e+23,"unit":"FLOP","observationPeriod":"2024-05-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*3400000000000*7000000000=1.428e+23"},{"id":"epoch:32d5860d65aa706ae301:parameters","entityId":"epoch:32d5860d65aa706ae301","metric":"parameters","value":1600000000.0,"unit":"parameters","observationPeriod":"2024-05-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 12"},{"id":"epoch:1c38d62490e334e1ac42:parameters","entityId":"epoch:1c38d62490e334e1ac42","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-05-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1c38d62490e334e1ac42:training-compute","entityId":"epoch:1c38d62490e334e1ac42","metric":"training-compute","value":1.716e+23,"unit":"FLOP","observationPeriod":"2024-05-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Finetuning compute: \n6*13000000000*1200000000000=9.36e+22\n\nLlama 13B training FLOP: 7.8e+22\n\nTotal: 1.716e+23"},{"id":"epoch:afe652b54e155cd4e6aa:parameters","entityId":"epoch:afe652b54e155cd4e6aa","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-05-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:afe652b54e155cd4e6aa:training-compute","entityId":"epoch:afe652b54e155cd4e6aa","metric":"training-compute","value":1.062e+24,"unit":"FLOP","observationPeriod":"2024-05-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Llama 70B: 8.1e+23\nFinetune: 6*70000000000*600000000000=252000000000000000000000\nTotal: 1062000000000000000000000"},{"id":"epoch:962de66904b2b69234b4:parameters","entityId":"epoch:962de66904b2b69234b4","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-05-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:962de66904b2b69234b4:training-compute","entityId":"epoch:962de66904b2b69234b4","metric":"training-compute","value":9.04e+22,"unit":"FLOP","observationPeriod":"2024-05-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"FT: 6*7000000000*1200000000000=5.04e+22\n\nLlama 7B compute: 4e+22\n\nTotal: 9.04e+22"},{"id":"epoch:20a8bfa68c133071778c:parameters","entityId":"epoch:20a8bfa68c133071778c","metric":"parameters","value":34000000000.0,"unit":"parameters","observationPeriod":"2024-05-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:20a8bfa68c133071778c:training-compute","entityId":"epoch:20a8bfa68c133071778c","metric":"training-compute","value":1.0608e+24,"unit":"FLOP","observationPeriod":"2024-05-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*34000000000*5200000000000=1.060800e+24"},{"id":"epoch:416a71ad18b22253c962:training-compute","entityId":"epoch:416a71ad18b22253c962","metric":"training-compute","value":8.0490594e+20,"unit":"FLOP","observationPeriod":"2024-05-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"82600000000000*9022.8*3600*0.3=8.0490594e+20"},{"id":"epoch:51fb11ed966e521cd916:parameters","entityId":"epoch:51fb11ed966e521cd916","metric":"parameters","value":93000000.0,"unit":"parameters","observationPeriod":"2024-05-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Source: https://arxiv.org/abs/2405.12213 "},{"id":"epoch:51fb11ed966e521cd916:training-compute","entityId":"epoch:51fb11ed966e521cd916","metric":"training-compute","value":5.85e+20,"unit":"FLOP","observationPeriod":"2024-05-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training compute \n= peak FLOPs * utilization rate * training time\n~= 128 TPUs * 275e12 FLOPs / TPU * 0.33 * 14 hours * 3600 s / hour\n~= 585446400e12 FLOPs\n= 5.85e20 FLOPs,\nassuming utilization rate = 0.33."},{"id":"epoch:8ccd94f195afc962371d:parameters","entityId":"epoch:8ccd94f195afc962371d","metric":"parameters","value":27000000.0,"unit":"parameters","observationPeriod":"2024-05-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Source: https://arxiv.org/abs/2405.12213 "},{"id":"epoch:da19134bf1c2b7ff17cf:parameters","entityId":"epoch:da19134bf1c2b7ff17cf","metric":"parameters","value":110000000.0,"unit":"parameters","observationPeriod":"2024-05-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Explicitly denoted in Table 2"},{"id":"epoch:da19134bf1c2b7ff17cf:training-compute","entityId":"epoch:da19134bf1c2b7ff17cf","metric":"training-compute","value":6.46714368e+20,"unit":"FLOP","observationPeriod":"2024-05-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"All ProSST models is trained on a DGX-A800 GPU (8\u00d780G) server in BF16 precision for about a month.\" \n\n8*77970000000000*0.4*1month=646714368000000000000"},{"id":"epoch:2b145991182e463413ca:parameters","entityId":"epoch:2b145991182e463413ca","metric":"parameters","value":34000000000.0,"unit":"parameters","observationPeriod":"2024-05-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2b145991182e463413ca:training-compute","entityId":"epoch:2b145991182e463413ca","metric":"training-compute","value":1.6453571041e+24,"unit":"FLOP","observationPeriod":"2024-05-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"GPU method:\nTable 2 shows that 34B model pre-training uses 4282407 GPU-hours, trained across 3072 A100s.\n3.12e14 * 4282407 * 3600 * 0.3 =  1.44e24\n\nParameter-token method:\nPre-training goes over 9.2T tokens, post-training only goes over 1.1B tokens (sum of tokens column in Table 3).\n6 * 34B *  9.2T = 1.88e24\n\nGeometric mean: sqrt(1.44e24 * 1.88e24) = 1.65e24"},{"id":"epoch:eb2ac1b6a5f622db31c8:parameters","entityId":"epoch:eb2ac1b6a5f622db31c8","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-05-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:eb2ac1b6a5f622db31c8:training-compute","entityId":"epoch:eb2ac1b6a5f622db31c8","metric":"training-compute","value":3.3399700602e+23,"unit":"FLOP","observationPeriod":"2024-05-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"GPU method:\nTable 2 shows that 7B model pre-training uses 856481 GPU-hours, trained across 1024 A100s\n3.12e14 * 856481 * 3600 * 0.3 =  2.89e23\n\nParameter-token method:\nPre-training goes over 9.2T tokens, post-training only goes over 1.1B tokens (sum of tokens column in Table 3)\n6 * 7B *  9.2T = 3.86e23\n\nGeometric mean: sqrt(2.89e23 * 3.86e23) = 3.34e23"},{"id":"epoch:90a08227dc488f1d8070:parameters","entityId":"epoch:90a08227dc488f1d8070","metric":"parameters","value":67000000.0,"unit":"parameters","observationPeriod":"2024-05-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we are able to train a 67 million parameter model\""},{"id":"epoch:90a08227dc488f1d8070:training-compute","entityId":"epoch:90a08227dc488f1d8070","metric":"training-compute","value":1.078272e+19,"unit":"FLOP","observationPeriod":"2024-05-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1 day on 1 A100\n\n312000000000000*1day*0.4=10782720000000000000"},{"id":"epoch:e6ca428bc55ed4190812:parameters","entityId":"epoch:e6ca428bc55ed4190812","metric":"parameters","value":40000000000.0,"unit":"parameters","observationPeriod":"2024-05-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9aa991cd14840353064c:parameters","entityId":"epoch:9aa991cd14840353064c","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-05-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d75cc7f6804655a1ced2:parameters","entityId":"epoch:d75cc7f6804655a1ced2","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2024-05-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.5B"},{"id":"epoch:67a77e65807451753432:parameters","entityId":"epoch:67a77e65807451753432","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2024-05-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Yi-Large is a software over-the-air-driven closed-source large model with a parameter of over 100 billion tokens.\" from https://www.chinadaily.com.cn/a/202405/13/WS6641abd1a31082fc043c6ccd.html"},{"id":"epoch:67a77e65807451753432:training-compute","entityId":"epoch:67a77e65807451753432","metric":"training-compute","value":1.8e+24,"unit":"FLOP","observationPeriod":"2024-05-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*100000000000*3000000000000=1.8e+24\n\n(speculative confidence because training dataset size is very uncertain)"},{"id":"epoch:0b7ac999e3e78f32e6c5:parameters","entityId":"epoch:0b7ac999e3e78f32e6c5","metric":"parameters","value":34000000000.0,"unit":"parameters","observationPeriod":"2024-05-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"34b"},{"id":"epoch:0b7ac999e3e78f32e6c5:training-compute","entityId":"epoch:0b7ac999e3e78f32e6c5","metric":"training-compute","value":7.344e+23,"unit":"FLOP","observationPeriod":"2024-05-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 34*10^9 parameters * 3.6*10^12 tokens = 7.344e+23 FLOP"},{"id":"epoch:a9529e18ed91ab61ad1c:parameters","entityId":"epoch:a9529e18ed91ab61ad1c","metric":"parameters","value":8830000000.0,"unit":"parameters","observationPeriod":"2024-05-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8.83B (safetensors)"},{"id":"epoch:a9529e18ed91ab61ad1c:training-compute","entityId":"epoch:a9529e18ed91ab61ad1c","metric":"training-compute","value":1.90728e+23,"unit":"FLOP","observationPeriod":"2024-05-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 8.83*10^9 parameters * 3.6*10^12 tokens = 1.90728e+23 FLOP"},{"id":"epoch:e76b2dc0dbb71dfcbd98:parameters","entityId":"epoch:e76b2dc0dbb71dfcbd98","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2024-05-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d2819ee4421e21cc71ff:parameters","entityId":"epoch:d2819ee4421e21cc71ff","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-05-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"same as llama 3 70B"},{"id":"epoch:c1e14dcad20d9d1fb003:training-compute","entityId":"epoch:c1e14dcad20d9d1fb003","metric":"training-compute","value":2.1062592e+18,"unit":"FLOP","observationPeriod":"2024-05-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"training speed 95.2 molecules/sec (Table 1)\n\n1.5*10^6 molecules / 95.2 = 15756 seconds = 4 hours (1 epoch)\n\n\"a few GPU days\" - let's assume (!) it was 10 GPU-days \n\n10*24*3600*8126000000000*0.3 = 2.1062592e+18"},{"id":"epoch:2629d02abd9d444960ca:parameters","entityId":"epoch:2629d02abd9d444960ca","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-05-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Fugaku-LLM has 13 billion parameters (2)\""},{"id":"epoch:2629d02abd9d444960ca:training-compute","entityId":"epoch:2629d02abd9d444960ca","metric":"training-compute","value":2.9640000000001e+22,"unit":"FLOP","observationPeriod":"2024-05-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.wolframalpha.com/input?i=6+FLOP+*+13+billion+*+380+billion"},{"id":"epoch:36064600213ea5f84a57:parameters","entityId":"epoch:36064600213ea5f84a57","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-05-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:17dfa2d65dc460a29225:parameters","entityId":"epoch:17dfa2d65dc460a29225","metric":"parameters","value":4500000.0,"unit":"parameters","observationPeriod":"2024-05-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"As the training data size increases up to 3M, the the total number of parameters in M3GNet increase accordingly from 880K to 4.5M.\""},{"id":"epoch:17dfa2d65dc460a29225:training-compute","entityId":"epoch:17dfa2d65dc460a29225","metric":"training-compute","value":1.62e+16,"unit":"FLOP","observationPeriod":"2024-05-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Speculative confidence because I am unsure how to calculate gradient updates/tokens\n\n6ND = 6*4500000*3000000*200 = 1.62e+16"},{"id":"epoch:a106c8c5c6accdd7286f:parameters","entityId":"epoch:a106c8c5c6accdd7286f","metric":"parameters","value":182000000.0,"unit":"parameters","observationPeriod":"2024-05-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The total parameters of Graphormer is 182M.\""},{"id":"epoch:a106c8c5c6accdd7286f:training-compute","entityId":"epoch:a106c8c5c6accdd7286f","metric":"training-compute","value":1.118208e+20,"unit":"FLOP","observationPeriod":"2024-05-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Speculative confidence because I am unsure how to calculate gradient updates/tokens\n\n6*1,562,500*256*256*182000000=1.118208e+20"},{"id":"epoch:f38990d602e256866130:parameters","entityId":"epoch:f38990d602e256866130","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2024-05-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"11B"},{"id":"epoch:f38990d602e256866130:training-compute","entityId":"epoch:f38990d602e256866130","metric":"training-compute","value":3.6e+23,"unit":"FLOP","observationPeriod":"2024-05-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"trained on 5.5T tokens\n\n6 * 11B * 5.5T = 3.6e23"},{"id":"epoch:f28451d126a7d1faafaa:training-compute","entityId":"epoch:f28451d126a7d1faafaa","metric":"training-compute","value":4.1405645e+22,"unit":"FLOP","observationPeriod":"2024-05-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"256 GPUs * 480 hours [see training time notes] * 3600 sec / hour * 312000000000000 FLOP / GPU / sec * 0.3 [assumed utilization] = 4.1405645e+22 FLOP"},{"id":"epoch:61221de5286ff36d7318:parameters","entityId":"epoch:61221de5286ff36d7318","metric":"parameters","value":37000000000.0,"unit":"parameters","observationPeriod":"2024-05-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"37B\n\n\"we leverage pretrained EVA-02-CLIP-E-plus [70], LLaMA-33B [73] and SDXL [58] to initialize the Visual Encoder, Multimodal Modeling, and Visual Decoder, respectively\""},{"id":"epoch:3adbe9eef9c755279e8a:parameters","entityId":"epoch:3adbe9eef9c755279e8a","metric":"parameters","value":236000000000.0,"unit":"parameters","observationPeriod":"2024-05-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"21B active params, 236B total"},{"id":"epoch:3adbe9eef9c755279e8a:training-compute","entityId":"epoch:3adbe9eef9c755279e8a","metric":"training-compute","value":1.02e+24,"unit":"FLOP","observationPeriod":"2024-05-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"21b active params * 8.1 trillion * 6 = 1.02e24"},{"id":"epoch:8ad4a0740a999fb9d958:parameters","entityId":"epoch:8ad4a0740a999fb9d958","metric":"parameters","value":1422600000.0,"unit":"parameters","observationPeriod":"2024-05-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:8ad4a0740a999fb9d958:training-compute","entityId":"epoch:8ad4a0740a999fb9d958","metric":"training-compute","value":2.56e+18,"unit":"FLOP","observationPeriod":"2024-05-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We developed and trained all our models and baselines over the course of\nthree months on a cluster with 128 nodes of eight NVIDIA A100 GPUs each.\" 1024 \n\n6*300000000*1422600000=2.56068e+18"},{"id":"epoch:009268648039e2427396:parameters","entityId":"epoch:009268648039e2427396","metric":"parameters","value":500000000000.0,"unit":"parameters","observationPeriod":"2024-05-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"500B "},{"id":"epoch:009268648039e2427396:training-compute","entityId":"epoch:009268648039e2427396","metric":"training-compute","value":1.602828e+22,"unit":"FLOP","observationPeriod":"2024-05-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"989400000000000 FLOP/GPU/sec * 15000 GPU-hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.602828e+22 FLOP"},{"id":"epoch:957d9101431d2fe08bee:parameters","entityId":"epoch:957d9101431d2fe08bee","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:82b2e2c68674a1f7fb2e:parameters","entityId":"epoch:82b2e2c68674a1f7fb2e","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:172df0a33112173f268d:parameters","entityId":"epoch:172df0a33112173f268d","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:f437b226238c45924497:parameters","entityId":"epoch:f437b226238c45924497","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:bc4cfaba0e6012492743:parameters","entityId":"epoch:bc4cfaba0e6012492743","metric":"parameters","value":13493916736.0,"unit":"parameters","observationPeriod":"2024-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/Efficient-Large-Model/VILA1.5-13b/blob/main/llm/model.safetensors.index.json\nhttps://huggingface.co/Efficient-Large-Model/VILA1.5-13b/tree/main/vision_tower\nhttps://huggingface.co/Efficient-Large-Model/VILA1.5-13b/blob/main/mm_projector/model.safetensors\n\nllm: 13015864320\nvision_tower: 428225600\nmm_projector: 49826816\n\ntotal: 13493916736\n"},{"id":"epoch:bc4cfaba0e6012492743:training-compute","entityId":"epoch:bc4cfaba0e6012492743","metric":"training-compute","value":2.3003136e+21,"unit":"FLOP","observationPeriod":"2024-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"not directly reported (arXiv preprint is for VILA-13B), but assumed to be similar to VILA-13B (given similar size/architecture)"},{"id":"epoch:503c94173a7fce0cb744:parameters","entityId":"epoch:503c94173a7fce0cb744","metric":"parameters","value":1080000000.0,"unit":"parameters","observationPeriod":"2024-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.08B (Table 4a)"},{"id":"epoch:503c94173a7fce0cb744:training-compute","entityId":"epoch:503c94173a7fce0cb744","metric":"training-compute","value":1.0520327e+22,"unit":"FLOP","observationPeriod":"2024-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"details from Table 4 and Table 9\n\n6 FLOP / token / parameter * 1.08*10^9 parameters * 1.5*10^12 tokens = 9.72e+21 FLOP\n\n312000000000000 FLOP / sec / GPU [bf16 assumed] * 128 GPUs * 11 days * 24 hours / day * 3600 sec / hour * 0.3 [assumed utilization] = 1.1386552e+22 FLOP\n\nsqrt(9.72e+21*1.1386552e+22) = 1.0520327e+22"},{"id":"epoch:fe69cac0693fedb73330:parameters","entityId":"epoch:fe69cac0693fedb73330","metric":"parameters","value":3040000000.0,"unit":"parameters","observationPeriod":"2024-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.04B (Table 4a)"},{"id":"epoch:fe69cac0693fedb73330:training-compute","entityId":"epoch:fe69cac0693fedb73330","metric":"training-compute","value":3.417119e+22,"unit":"FLOP","observationPeriod":"2024-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"details from Table 4 and Table 9\n\n6 FLOP / token / parameter * 3.04*10^9 parameters * 1.5*10^12 tokens = 2.736e+22 FLOP\n\n989500000000000 FLOP / sec / GPU [bf16 assumed] *  128 GPUs * 13 days * 24 hours / day * 3600 sec / hour * 0.3 [assumed utilization] = 4.2678006e+22 FLOP\n\nsqrt(2.736e+22*4.2678006e+22) = 3.417119e+22"},{"id":"epoch:b1226f614784c78c1ede:parameters","entityId":"epoch:b1226f614784c78c1ede","metric":"parameters","value":450000000.0,"unit":"parameters","observationPeriod":"2024-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"0.45B (Table 4a)"},{"id":"epoch:b1226f614784c78c1ede:training-compute","entityId":"epoch:b1226f614784c78c1ede","metric":"training-compute","value":6.3156568e+21,"unit":"FLOP","observationPeriod":"2024-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"details from Table 4 and Table 9\n\n6 FLOP / token / parameter * 0.45*10^9 parameters * 1.5*10^12 tokens = 4.05e+21 FLOP\n\n989500000000000 FLOP / sec / GPU [bf16 assumed] * 128 GPUs * 3 days * 24 hours / day * 3600 sec / hour * 0.3 [assumed utilization] = 9.8487706e+21 FLOP\n\nsqrt(4.05e+21*9.8487706e+21) = 6.3156568e+21"},{"id":"epoch:8641500b4b71e7a368d7:parameters","entityId":"epoch:8641500b4b71e7a368d7","metric":"parameters","value":270000000.0,"unit":"parameters","observationPeriod":"2024-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"0.27B (Table 4a)"},{"id":"epoch:8641500b4b71e7a368d7:training-compute","entityId":"epoch:8641500b4b71e7a368d7","metric":"training-compute","value":2.7470309e+21,"unit":"FLOP","observationPeriod":"2024-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"details from Table 4 and Table 9\n\n6 FLOP / token / parameter * 0.27*10^9 parameters * 1.5*10^12 tokens = 2.43e+21 FLOP\n\n312000000000000 FLOP / sec / GPU [bf16 assumed] * 128 GPUs * 3 days * 24 hours / day * 3600 sec / hour * 0.3 [assumed utilization]  = 3.1054234e+21 FLOP\n\nsqrt(2.43e+21*3.1054234e+21) = 2.7470309e+21"},{"id":"epoch:f7638766907296688826:training-compute","entityId":"epoch:f7638766907296688826","metric":"training-compute","value":8.169984e+20,"unit":"FLOP","observationPeriod":"2024-05-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The model was trained for 120 hours on 32 TPUv5. \nAssuming it was TPUv5e, bf16 precision:\n\n197000000000000 FLOP/s * 120 hours * 3600 s/hour * 32 instances * 0.3 [assumed utilization] = 8.169984e+20 FLOP"},{"id":"epoch:070662fcaaf5b46ace76:parameters","entityId":"epoch:070662fcaaf5b46ace76","metric":"parameters","value":6700000000.0,"unit":"parameters","observationPeriod":"2024-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6.7B (\u201c7B\u201d)"},{"id":"epoch:070662fcaaf5b46ace76:training-compute","entityId":"epoch:070662fcaaf5b46ace76","metric":"training-compute","value":3.841092e+23,"unit":"FLOP","observationPeriod":"2024-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"training all models reported in the paper required around 500K GPU hours of computation on hardware of type A100-80GB and H100.\"\nA100-80 GB peak FLOP/s [assumed fp16 precision]: 77970000000000\nH100 peak FLOP/s [assumed SXM5 TensorCore]: 989000000000000\nassuming 50/50 usage:\n\n(77970000000000+989000000000000)*0.5*500000hours*3600s*0.3=2.880819e+23 for ALL models in the paper\n\nassuming this model has taken around 40% of all used compute (https://docs.google.com/spreadsheets/d/1Yc-HAdYgn6e9SUIliMaQtvQscIsEZomLdXbWyYl_jfQ/edit?usp=sharing)\n\nthen it's assumed compute is 3.841092e+23 FLOPs"},{"id":"epoch:ccbb6f2a69181720325a:parameters","entityId":"epoch:ccbb6f2a69181720325a","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B (Figure 1)"},{"id":"epoch:ccbb6f2a69181720325a:training-compute","entityId":"epoch:ccbb6f2a69181720325a","metric":"training-compute","value":1.5364368e+23,"unit":"FLOP","observationPeriod":"2024-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"training all models reported in the paper required around 500K GPU hours of computation on hardware of type A100-80GB and H100.\"\nA100-80 GB peak FLOP/s [assumed fp16 precision]: 77970000000000\nH100 peak FLOP/s [assumed SXM5 TensorCore]: 989000000000000\nassuming 50/50 usage:\n\n(77970000000000+989000000000000)*0.5*500000hours*3600s*0.3=2.880819e+23 for ALL models in the paper\n\nassuming this model has taken around 16% of all used compute (https://docs.google.com/spreadsheets/d/1Yc-HAdYgn6e9SUIliMaQtvQscIsEZomLdXbWyYl_jfQ/edit?usp=sharing)\n\nthen it's assumed compute is 1.5364368e+23 FLOPs"},{"id":"epoch:522e125fd8c2ac76ae49:parameters","entityId":"epoch:522e125fd8c2ac76ae49","metric":"parameters","value":104000000.0,"unit":"parameters","observationPeriod":"2024-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:522e125fd8c2ac76ae49:training-compute","entityId":"epoch:522e125fd8c2ac76ae49","metric":"training-compute","value":7.667785728001e+21,"unit":"FLOP","observationPeriod":"2024-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware: 8x NVIDIA A800 80GB GPUs (7.80e13 FLOP/s per GPU)\n2. Training duration: 5 days = 432,000 seconds (directly provided)\n3. Utilization: 40%\n4. Calculation: \n   7.80e13 FLOP/s/GPU \u00d7 8 GPUs \u00d7 432,000s \u00d7 0.40 = 1.08e20 FLOP\n   (7.80e13 \u00d7 8 = 6.24e14 FLOP/s total \u2192 \u00d7 432,000s = 2.70e20 \u2192 \u00d7 0.40 = 1.08e20)\n\nBase model: 7.560000000001e+21\n\nTotal: 7667785728001000000000"},{"id":"epoch:40f77fa57be2913046c9:parameters","entityId":"epoch:40f77fa57be2913046c9","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"To address these challenges, we introduce SeedLLM\u00b7Rice (SeedLLM), a 7-billion-parameter model trained on 1.4 million rice-related publications, representing nearly 98.24% of global rice research output."},{"id":"epoch:dfc4bac8e44a77095f18:parameters","entityId":"epoch:dfc4bac8e44a77095f18","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:61275bfaed3bbbe7f360:parameters","entityId":"epoch:61275bfaed3bbbe7f360","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ea9640f82146c615c591:parameters","entityId":"epoch:ea9640f82146c615c591","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:3b7899a6e60fe431bbde:parameters","entityId":"epoch:3b7899a6e60fe431bbde","metric":"parameters","value":25500000000.0,"unit":"parameters","observationPeriod":"2024-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"25.5B"},{"id":"epoch:c74d589353b9d6453d99:parameters","entityId":"epoch:c74d589353b9d6453d99","metric":"parameters","value":70000000.0,"unit":"parameters","observationPeriod":"2024-04-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:af5a04b97f5f90d26709:parameters","entityId":"epoch:af5a04b97f5f90d26709","metric":"parameters","value":110000000000.0,"unit":"parameters","observationPeriod":"2024-04-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"110B\n"},{"id":"epoch:5992f728aa010c969a66:parameters","entityId":"epoch:5992f728aa010c969a66","metric":"parameters","value":35800000.0,"unit":"parameters","observationPeriod":"2024-04-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b51e5efcecf62745535f:parameters","entityId":"epoch:b51e5efcecf62745535f","metric":"parameters","value":480000000000.0,"unit":"parameters","observationPeriod":"2024-04-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\" It combines a 10B dense transformer model with a residual 128x3.66B MoE MLP resulting in 480B total and 17B active parameters chosen using a top-2 gating.\""},{"id":"epoch:b51e5efcecf62745535f:training-compute","entityId":"epoch:b51e5efcecf62745535f","metric":"training-compute","value":3.8347175e+23,"unit":"FLOP","observationPeriod":"2024-04-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from the graph:\n1x - Arctic\n1.9X - Llama 3 8B (7.2\u00d710^23) ~ = Yi 34B (6.1e23) -> x = 3.2105263e+23\n3X - Code Llama 70B (1.26e+24) -> x = 4.2e+23\n17.5X - Llama 3 70B (7.861e24) -> x =4.492e+23\n\n= 3.7975893e+23\n\nOperation counting (17B active parameters): \n6ND = 6 FLOP / parameter / token * 17*10^9 parameters * 3.5*10^12 tokens = 3.57e+23 FLOP\n\ngeometric mean:(3.2105263e+23*4.492e+23*4.2e+23*3.57e+23)^(1/4) = 3.8347175e+23"},{"id":"epoch:7012de0eaa272c475bbc:parameters","entityId":"epoch:7012de0eaa272c475bbc","metric":"parameters","value":3800000000.0,"unit":"parameters","observationPeriod":"2024-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.8B"},{"id":"epoch:7012de0eaa272c475bbc:training-compute","entityId":"epoch:7012de0eaa272c475bbc","metric":"training-compute","value":7.524e+22,"unit":"FLOP","observationPeriod":"2024-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"counting operations: 6\u00d73.3\u00d710^12 tokens \u00d73.8\u00d710^9 parameters \u22487.524\u00d710^22 FLOPS\nhardware estimate: 7 days \u00d724 hours / day\u00d73600 sec / hour *989,000,000,000,000 FLOP/s*512 GPUs*0.3 [assumed utilization]=9.187540992\u00d710^22"},{"id":"epoch:598f2e4d5ad19e2f8598:parameters","entityId":"epoch:598f2e4d5ad19e2f8598","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2024-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B"},{"id":"epoch:598f2e4d5ad19e2f8598:training-compute","entityId":"epoch:598f2e4d5ad19e2f8598","metric":"training-compute","value":4.032e+23,"unit":"FLOP","observationPeriod":"2024-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"counting operations: 6\u00d74.8\u00d710^12 tokens \u00d7 14\u00d710^9 parameters \u2248 4.032\u00d710^23 FLOPS"},{"id":"epoch:46b055bc0aa1d828a41f:parameters","entityId":"epoch:46b055bc0aa1d828a41f","metric":"parameters","value":600000000000.0,"unit":"parameters","observationPeriod":"2024-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"This article claims the model is a 600B MoE:\nhttps://www.sensetime.com/cn/news-detail/51168158?categoryId=72"},{"id":"epoch:3ce713ec571f93ca5d91:parameters","entityId":"epoch:3ce713ec571f93ca5d91","metric":"parameters","value":7400000000.0,"unit":"parameters","observationPeriod":"2024-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.4B"},{"id":"epoch:3ce713ec571f93ca5d91:training-compute","entityId":"epoch:3ce713ec571f93ca5d91","metric":"training-compute","value":2.1312e+23,"unit":"FLOP","observationPeriod":"2024-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 FLOP / parameter / token * 7.4B parameters * 4.8T tokens = 2.1312e+23 FLOP"},{"id":"epoch:9cd329e091e6956e53c4:parameters","entityId":"epoch:9cd329e091e6956e53c4","metric":"parameters","value":3800000000.0,"unit":"parameters","observationPeriod":"2024-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.8B"},{"id":"epoch:9cd329e091e6956e53c4:training-compute","entityId":"epoch:9cd329e091e6956e53c4","metric":"training-compute","value":3.7101154e+22,"unit":"FLOP","observationPeriod":"2024-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*3800000000.00 parameters *3400000000000 tokens  = 7.752e+22\n\n512 GPUs *133800000000000 FLOP/s *240 hours *3600 sec/hour *0.3 [assumed utilization] = 1.7756652e+22\n\ngeometric mean sqrt (7.752e+22*1.7756652e+22) = 3.7101154e+22"},{"id":"epoch:20bb1ffc37ad54d13904:parameters","entityId":"epoch:20bb1ffc37ad54d13904","metric":"parameters","value":4200000000.0,"unit":"parameters","observationPeriod":"2024-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4.2B"},{"id":"epoch:20bb1ffc37ad54d13904:training-compute","entityId":"epoch:20bb1ffc37ad54d13904","metric":"training-compute","value":8.784e+22,"unit":"FLOP","observationPeriod":"2024-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Base model: 7.524e+22\nFinetune: 1.260000e+22\nTotal: 87840000000000000000000"},{"id":"epoch:79b2843f5a0bf2de4ca0:parameters","entityId":"epoch:79b2843f5a0bf2de4ca0","metric":"parameters","value":60800000000.0,"unit":"parameters","observationPeriod":"2024-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Phi-3.5-MoE has 16x3.8B parameters with 6.6B active parameters when using 2 experts. The model is a mixture-of-expert decoder-only Transformer model using the tokenizer with vocabulary size of 32,064.\""},{"id":"epoch:79b2843f5a0bf2de4ca0:training-compute","entityId":"epoch:79b2843f5a0bf2de4ca0","metric":"training-compute","value":3.0202896e+23,"unit":"FLOP","observationPeriod":"2024-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"512 GPUs * 989500000000000 FLOP / sec * 552 hours * 3600 sec / hour * 0.3 [assumed utilization] = 3.0202896e+23 FLOP\n\n6 FLOP / token / parameter * 4900000000000 tokens * 6.6*10^9 active parameters = 1.9404e+23 FLOP (slightly less confidence than hardware estimation since the 6ND formula is less accurate for MoE)"},{"id":"epoch:7f1e033e5c2ee4b51265:parameters","entityId":"epoch:7f1e033e5c2ee4b51265","metric":"parameters","value":100000000.0,"unit":"parameters","observationPeriod":"2024-04-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"VISTA-2D has a network architecture with ~100 million training hyperparameters\"\nPresumably the quote is incorrect and they meant 100 million parameters."},{"id":"epoch:4283f3d7d3f3149c7fa0:parameters","entityId":"epoch:4283f3d7d3f3149c7fa0","metric":"parameters","value":89100000.0,"unit":"parameters","observationPeriod":"2024-04-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:7a594f311f73e7f622aa:parameters","entityId":"epoch:7a594f311f73e7f622aa","metric":"parameters","value":650000000.0,"unit":"parameters","observationPeriod":"2024-04-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"To our knowledge, SaProt stands out as the PLM currently trained with the largest number of protein structures, containing 650 million parameters\" Assume 40% utilization, FP16 tensor precision"},{"id":"epoch:7a594f311f73e7f622aa:training-compute","entityId":"epoch:7a594f311f73e7f622aa","metric":"training-compute","value":6.21084672e+22,"unit":"FLOP","observationPeriod":"2024-04-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Its training lasted 3 months and utilized 64 NVIDIA 80G A100 GPUs,\"\n64*312000000000000*0.4*3months=6.210847e+22"},{"id":"epoch:fc7355ed1ffd3f0790db:parameters","entityId":"epoch:fc7355ed1ffd3f0790db","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-04-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:fc7355ed1ffd3f0790db:training-compute","entityId":"epoch:fc7355ed1ffd3f0790db","metric":"training-compute","value":7.861e+24,"unit":"FLOP","observationPeriod":"2024-04-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Arithmetic calculation:\n6 * 15T tokens * 70B parameters = 6.3e24\n\nGPU calculation:\nhttps://huggingface.co/meta-llama/Meta-Llama-3-70B indicates training took 6.4M GPU-hours\nWe also know their larger scale training runs for 405B were getting between 0.38-0.41 MFU. Presumably the 70B model gets at least 0.43 utilization (405B has to be split across two nodes, while 70B should fit on one).\n990 TFLOPS per GPU * 6.4 million GPU hours * 3600s * 0.43 = 9.808e24\n\nGeometric mean: sqrt(6.3e24 * 9.808e24) = 7.861e24"},{"id":"epoch:346c553c9a7a695b22b1:parameters","entityId":"epoch:346c553c9a7a695b22b1","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-04-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:346c553c9a7a695b22b1:training-compute","entityId":"epoch:346c553c9a7a695b22b1","metric":"training-compute","value":7.2e+23,"unit":"FLOP","observationPeriod":"2024-04-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Counting operations\n15000000000000 tokens*8000000000.00 parameters*6 FLOP / token / parameter = 7.2\u00d710^23 FLOP\n\nGPU calculation\n400 TFLOPS per GPU * 1.3M GPU hours * 3600s=1.872\u00d710^24 \n(it is not confident that 400 TFLOPs applies to the Llama 3-8B training run)"},{"id":"epoch:118701cac09e656a087f:parameters","entityId":"epoch:118701cac09e656a087f","metric":"parameters","value":1740000000.0,"unit":"parameters","observationPeriod":"2024-04-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.74B"},{"id":"epoch:118701cac09e656a087f:training-compute","entityId":"epoch:118701cac09e656a087f","metric":"training-compute","value":2.5264222803e+22,"unit":"FLOP","observationPeriod":"2024-04-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 1500000000000 tokens * 1740000000 parameters = 1.566e+22 FLOP\nor\n312000000000000 FLOPs/s/GPU * 112 GPUs * 45 days * 24 h * 3600 s * 0.3  = 4.076e22 FLOP\ngeometric mean: 2.5e22"},{"id":"epoch:2e4ce51ec793ae73df0a:parameters","entityId":"epoch:2e4ce51ec793ae73df0a","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-04-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:bdcc465a1db37e705d3b:parameters","entityId":"epoch:bdcc465a1db37e705d3b","metric":"parameters","value":1100000000.0,"unit":"parameters","observationPeriod":"2024-04-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.1B"},{"id":"epoch:114756b34e117940871c:parameters","entityId":"epoch:114756b34e117940871c","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-04-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B dense"},{"id":"epoch:114756b34e117940871c:training-compute","entityId":"epoch:114756b34e117940871c","metric":"training-compute","value":1.89e+23,"unit":"FLOP","observationPeriod":"2024-04-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 7 * 10^9 parameters * 4.5 * 10^12 text tokens = 1.89e+23 FLOP \n\nthe calculation above doesn't account for multimodal training, total training compute could be 2-5 times higher\n\n\"Reka Flash and Edge were trained on several hundreds of H100s across a period of several weeks.\"\n\n989400000000000 FLOP / GPU / sec * 300 GPUs [assumption] * 3 weeks [assumption] * 168 hours / week * 3600 sec / hour * 0.3 [assumed utilization] = 1.6156506e+23 FLOP (same OOM)"},{"id":"epoch:950765a791afd962dc7c:parameters","entityId":"epoch:950765a791afd962dc7c","metric":"parameters","value":141000000000.0,"unit":"parameters","observationPeriod":"2024-04-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"141B params, 39B active: https://mistral.ai/news/mixtral-8x22b/ "},{"id":"epoch:950765a791afd962dc7c:training-compute","entityId":"epoch:950765a791afd962dc7c","metric":"training-compute","value":2.34e+24,"unit":"FLOP","observationPeriod":"2024-04-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming the model was trained on ~1-10 trillions of tokens (same OOM as the models from the comparison in Figure 1. Llama 2 was trained on 2T tokens) + Mistral Small 3 was trained on 8T of tokens, we can estimate training compute with \"speculative\" confidence:\n\n6 FLOP / token / parameter * 39 * 10^9 active parameters * 10*10^12 tokens [speculatively] = 2.34e+24 FLOP"},{"id":"epoch:49ea8f7dd617a1bac8bb:parameters","entityId":"epoch:49ea8f7dd617a1bac8bb","metric":"parameters","value":50000000.0,"unit":"parameters","observationPeriod":"2024-04-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e03b9bb05e89da23b8f8:parameters","entityId":"epoch:e03b9bb05e89da23b8f8","metric":"parameters","value":7240000000.0,"unit":"parameters","observationPeriod":"2024-04-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.24B"},{"id":"epoch:60d8976b4ee18f2344a4:parameters","entityId":"epoch:60d8976b4ee18f2344a4","metric":"parameters","value":46700000000.0,"unit":"parameters","observationPeriod":"2024-04-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"46.7B"},{"id":"epoch:33a4a31bdc1c7a61ecb5:parameters","entityId":"epoch:33a4a31bdc1c7a61ecb5","metric":"parameters","value":400000000000.0,"unit":"parameters","observationPeriod":"2024-04-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"\u300c\u5929\u5de53.0\u300d\u62e5\u67094000\u4ebf\u53c2\u6570\uff0c\u8d85\u8d8a\u4e863140\u4ebf\u53c2\u6570\u7684Grok-1\uff0c\u662f\u5168\u7403\u6700\u5927\u7684\u5f00\u6e90MoE\u5927\u6a21\u578b\""},{"id":"epoch:c57383da057136789495:parameters","entityId":"epoch:c57383da057136789495","metric":"parameters","value":1000000000000.0,"unit":"parameters","observationPeriod":"2024-04-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:31443c5f88444ca23ebc:parameters","entityId":"epoch:31443c5f88444ca23ebc","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-04-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:cfb72a7e93fd82b693d9:parameters","entityId":"epoch:cfb72a7e93fd82b693d9","metric":"parameters","value":67000000000.0,"unit":"parameters","observationPeriod":"2024-04-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:cfb72a7e93fd82b693d9:training-compute","entityId":"epoch:cfb72a7e93fd82b693d9","metric":"training-compute","value":8.400010000000001e+24,"unit":"FLOP","observationPeriod":"2024-04-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"No direct information about Reka Core model (\"Reka Core has not finished training and is still improving.\")\n\nThe smaller dense model Reka Flash has 21B parameters and was trained on 5 trillion language tokens.\n\nThere is information about compute: \"Our setup comprises of clusters from a mixture of vendors with our peak compute being approximately\n2.5K H100s and 2.5K A100s.\"\n\nIf we assume 2 months of training with 2.5k H100s and 2.5k A100s at utilization 0.5 we get 8.4e24 FLOP (2500*9.9e14+2500*3.12e14)*60*60*24*60*0.5."},{"id":"epoch:3e224b22526065e27291:parameters","entityId":"epoch:3e224b22526065e27291","metric":"parameters","value":21000000000.0,"unit":"parameters","observationPeriod":"2024-04-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"21B dense"},{"id":"epoch:3e224b22526065e27291:training-compute","entityId":"epoch:3e224b22526065e27291","metric":"training-compute","value":6.3e+23,"unit":"FLOP","observationPeriod":"2024-04-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Reka Flash has 21B parameters and was trained on 5 trillion language tokens (+ unknown amount of multimodal data)\n\n6 FLOP / token / parameter * 21B parameters * 5 trillion text tokens = 6.3 \u00d7 10^23 FLOP\n\nOOM agrees with GPU details: \"Reka Flash and Edge were trained on several hundreds of H100s across a period of several weeks.\"\n\n3 weeks * 300 H100s * 7 day/week * 24 hour/day * 3600 s/day * 9.9e14 FLOP/GPU-s * 0.3 [assumed utilization]=  1.6156506e+23 FLOP\n\nNot enough info to estimate SFT and RLHF post-training FLOPs."},{"id":"epoch:4900665233e636b07bd7:parameters","entityId":"epoch:4900665233e636b07bd7","metric":"parameters","value":141000000000.0,"unit":"parameters","observationPeriod":"2024-04-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Parameters: 141B"},{"id":"epoch:85da130cfc024ec4dcbf:parameters","entityId":"epoch:85da130cfc024ec4dcbf","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-04-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70b"},{"id":"epoch:d55002838db4ed3f6a73:parameters","entityId":"epoch:d55002838db4ed3f6a73","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-04-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7b"},{"id":"epoch:cf0abc62c1ecb36fdda1:training-compute","entityId":"epoch:cf0abc62c1ecb36fdda1","metric":"training-compute","value":9.000000000000006e+17,"unit":"FLOP","observationPeriod":"2024-04-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 1x NVIDIA A40 GPU (1.50\u00d710\u00b9\u2074 FLOP/s per GPU)\n\n2. Training duration: 15,000 seconds (directly provided - sum of TCGA training: 3,780s and GTEx training: 11,220s)\n\n3. Utilization rate: 40%\n\n4. Calculation: 1.50\u00d710\u00b9\u2074 FLOP/s \u00d7 1 GPU \u00d7 15,000s \u00d7 0.4 = 9.0\u00d710\u00b9\u2077 FLOPs"},{"id":"epoch:9e59dd65c95547e30133:training-compute","entityId":"epoch:9e59dd65c95547e30133","metric":"training-compute","value":9.000000000000006e+17,"unit":"FLOP","observationPeriod":"2024-04-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 1x NVIDIA A40 GPU (1.50\u00d710\u00b9\u2074 FLOP/s per GPU)\n\n2. Training duration: 15,000 seconds (directly provided - sum of TCGA training: 3,780s and GTEx training: 11,220s)\n\n3. Utilization rate: 40%\n\n4. Calculation: 1.50\u00d710\u00b9\u2074 FLOP/s \u00d7 1 GPU \u00d7 15,000s \u00d7 0.4 = 9.0\u00d710\u00b9\u2077 FLOPs"},{"id":"epoch:6e461c9c93f9b90e7a25:parameters","entityId":"epoch:6e461c9c93f9b90e7a25","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2024-04-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"multi-billion (putting billion, then speculative)"},{"id":"epoch:5210e60e6160e693aa89:parameters","entityId":"epoch:5210e60e6160e693aa89","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-04-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:5210e60e6160e693aa89:training-compute","entityId":"epoch:5210e60e6160e693aa89","metric":"training-compute","value":4.2e+22,"unit":"FLOP","observationPeriod":"2024-04-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 * 7000000000 * 1000000000000 = 4.2e+22\n(unknown number of epochs -> 'likely' confidence)"},{"id":"epoch:52935a2a9fa3ace1e1ba:parameters","entityId":"epoch:52935a2a9fa3ace1e1ba","metric":"parameters","value":2900000000.0,"unit":"parameters","observationPeriod":"2024-04-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.9B"},{"id":"epoch:52935a2a9fa3ace1e1ba:training-compute","entityId":"epoch:52935a2a9fa3ace1e1ba","metric":"training-compute","value":1.74e+21,"unit":"FLOP","observationPeriod":"2024-04-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6* 2.9*10^9*100*10^9 = 1.74e+21"},{"id":"epoch:b206df62aed0a856cd50:parameters","entityId":"epoch:b206df62aed0a856cd50","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2024-04-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1B"},{"id":"epoch:b206df62aed0a856cd50:training-compute","entityId":"epoch:b206df62aed0a856cd50","metric":"training-compute","value":6.000000001e+21,"unit":"FLOP","observationPeriod":"2024-04-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*10^9 * 100*10^9 = 6*10^21"},{"id":"epoch:f713d98c900ae2a24640:parameters","entityId":"epoch:f713d98c900ae2a24640","metric":"parameters","value":141000000000.0,"unit":"parameters","observationPeriod":"2024-04-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"A Mixture of Experts (MoE) model with 141B total parameters and 39B active parameters.\""},{"id":"epoch:d4b0dbdedd23b31ea142:training-compute","entityId":"epoch:d4b0dbdedd23b31ea142","metric":"training-compute","value":4.000000000000001e+20,"unit":"FLOP","observationPeriod":"2024-04-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup:\n- Main model: 8\u00d7 NVIDIA A800 GPUs (7.80\u00d710\u00b9\u00b3 FLOP/s per GPU)\n- MDN model: 4\u00d7 NVIDIA Tesla V100-SXM2 GPUs (1.25\u00d710\u00b9\u2074 FLOP/s per GPU)\n\n2. Training duration (estimated from steps and step time):\n- Main model: 262,000 steps \u00d7 5s = 1.31\u00d710\u2076 seconds (~15 days)\n- MDN model: 65,000 steps \u00d7 5s = 3.25\u00d710\u2075 seconds (~3.76 days)\n\n3. Utilization rate: 40%\n\n4. Final calculation:\nMain: 8 GPUs \u00d7 7.80\u00d710\u00b9\u00b3 FLOP/s \u00d7 1.31\u00d710\u2076 s \u00d7 0.4 = 3.31\u00d710\u00b2\u2070 FLOPs\nMDN: 4 GPUs \u00d7 1.25\u00d710\u00b9\u2074 FLOP/s \u00d7 3.25\u00d710\u2075 s \u00d7 0.4 = 6.50\u00d710\u00b9\u2079 FLOPs\nTotal: 3.31\u00d710\u00b2\u2070 + 6.50\u00d710\u00b9\u2079 \u2248 4.0\u00d710\u00b2\u2070 FLOPs"},{"id":"epoch:da1c372220f3071a39e1:parameters","entityId":"epoch:da1c372220f3071a39e1","metric":"parameters","value":15000000000.0,"unit":"parameters","observationPeriod":"2024-04-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"15B"},{"id":"epoch:ac9d7731ffa1ac2ee141:parameters","entityId":"epoch:ac9d7731ffa1ac2ee141","metric":"parameters","value":12143605760.0,"unit":"parameters","observationPeriod":"2024-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Precise number given in HF model card"},{"id":"epoch:ac9d7731ffa1ac2ee141:training-compute","entityId":"epoch:ac9d7731ffa1ac2ee141","metric":"training-compute","value":2.91e+23,"unit":"FLOP","observationPeriod":"2024-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2* 12143605760 params * 3* 2T tokens * 2 epochs = 2.91e23. \nTrained on 384 H100s (AWS P5 instances)."},{"id":"epoch:aaab24b81343bf459e26:parameters","entityId":"epoch:aaab24b81343bf459e26","metric":"parameters","value":2300000000.0,"unit":"parameters","observationPeriod":"2024-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:aaab24b81343bf459e26:training-compute","entityId":"epoch:aaab24b81343bf459e26","metric":"training-compute","value":8.21376e+19,"unit":"FLOP","observationPeriod":"2024-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"There is cascade of 3 models:\n\nGEN64\nParameters: 2.3 billion\nInput: Textual prompt\nOutput: 64x64 image\nBatch Size: 4800\nIterations: 1.1 \u00d7 10^6 \n\nCompute = 6 * 2.3 billion * 4800 * 1.1 \u00d7 10^6 = 7.2864e+19\n\nSR256\nParameters: 700 million\nInput: 64x64 image\nOutput: 256x256 image\nBatch Size: 960\nIterations: 1.5 \u00d7 10^6 \n\nCompute = 6 * 700 million * 960 * 1.5 \u00d7 10^6 = 6.048e+18\n\nSR1024\nParameters: 700 million\nInput: 256x256 image\nOutput: 1024x1024 image\nBatch Size: 512\nIterations: 1.5 \u00d7 10^6 \n\nCompute = 6 * 700 million * 512 * 1.5 \u00d7 10^6 = 3.2256e+18\n\nTotal compute = 7.2864e+19 + 6.048e+18 + 3.2256e+18 = 8.21376e+19"},{"id":"epoch:d5fc658f7b797aeaba6a:parameters","entityId":"epoch:d5fc658f7b797aeaba6a","metric":"parameters","value":13100000000.0,"unit":"parameters","observationPeriod":"2024-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/openthaigpt/openthaigpt-1.0.0-13b-chat"},{"id":"epoch:d5fc658f7b797aeaba6a:training-compute","entityId":"epoch:d5fc658f7b797aeaba6a","metric":"training-compute","value":5.11e+21,"unit":"FLOP","observationPeriod":"2024-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Pretrained upon a foundation of more than 65 billion Thai language words and meticulously fine-tuned with over 1 million Thai instruction examples\". The model has 13.1B parameters (https://huggingface.co/openthaigpt/openthaigpt-1.0.0-13b-chat).\n\nEvery Asian language in https://docs.google.com/document/d/1XWLyMzcVfDv4eFQX3yPgM8MZ3_Q1phtIFz9GKv4_KaM/edit?tab=t.0#heading=h.ieihc08p8dn0 has a word per token ratio of 1, so I will assume the same for Thai. Since Llama has dense architecture, assuming pretraining was done for 1 epoch, the 6ND approximation yields\nPre-training compute\n= # of active parameters / forward pass * # of tokens * 6 FLOPS / token\n~= 1.31e10 parameters * 6.5e10 words * 1 token / word * 6 FLOPS / token\n~= 5.11e21 FLOPS"},{"id":"epoch:d85ec97af9a0d430e360:parameters","entityId":"epoch:d85ec97af9a0d430e360","metric":"parameters","value":6810000000.0,"unit":"parameters","observationPeriod":"2024-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/openthaigpt/openthaigpt-1.0.0-7b-chat"},{"id":"epoch:d85ec97af9a0d430e360:training-compute","entityId":"epoch:d85ec97af9a0d430e360","metric":"training-compute","value":2.66e+21,"unit":"FLOP","observationPeriod":"2024-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Pretrained upon a foundation of more than 65 billion Thai language words and meticulously fine-tuned with over 1 million Thai instruction examples\". The model has 6.81B parameters (https://huggingface.co/openthaigpt/openthaigpt-1.0.0-7b-chat).\n\nEvery Asian language in https://docs.google.com/document/d/1XWLyMzcVfDv4eFQX3yPgM8MZ3_Q1phtIFz9GKv4_KaM/edit?tab=t.0#heading=h.ieihc08p8dn0 has a word per token ratio of 1, so I will assume the same for Thai. Since Llama has dense architecture, assuming pretraining was done for 1 epoch, the 6ND approximation yields\nPre-training compute\n= # of active parameters / forward pass * # of tokens * 6 FLOPS / token\n~= 6.81e9 parameters * 6.5e10 words * 1 token / word * 6 FLOPS / token\n~= 2.66e21 FLOPS"},{"id":"epoch:2e63e4a6baa84026b036:parameters","entityId":"epoch:2e63e4a6baa84026b036","metric":"parameters","value":6950000000.0,"unit":"parameters","observationPeriod":"2024-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/sambanovasystems/SambaLingo-Thai-Chat"},{"id":"epoch:2e63e4a6baa84026b036:training-compute","entityId":"epoch:2e63e4a6baa84026b036","metric":"training-compute","value":8.569999999999999e+22,"unit":"FLOP","observationPeriod":"2024-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Continuous Pre-training: We pack the pretraining mixture into sequences of length 4096 and pretrain with document attention as described in Section 3.2 of Iyer et al. (2022) to ensure we only attend to tokens in the context of the corresponding text document. We train with a global batch size of 1024, sequence length of 4096, maximum learning rate of 1e-4 with cosine decay, warm-up ratio of 0.01 and a weight decay of 0.1. Each expert is trained for a maximum of 4 epochs, following (Muennighoff et al., 2023). Notably, we train all model parameters, foregoing use of PEFT methods such as LoRA (Hu et al., 2022), which are known to be inferior to\nfull parameter training (Zhao et al., 2024a)(Sun et al., 2023)\" (https://arxiv.org/pdf/2404.05829, page 16). \n\n\"The base model adapts Llama-2-7b to Thai by training on 38 billion tokens from the Thai split of the Cultura-X dataset\" (https://huggingface.co/sambanovasystems/SambaLingo-Thai-Chat).  \n\n\nFine-tune: 7.0e+9*4.0e+10*6=1.7e+21\nBase model: 8.4e+22\nTotal: 8.4e+22+1.7e+21=8.6e+22"},{"id":"epoch:7d50637972633676ba76:parameters","entityId":"epoch:7d50637972633676ba76","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/sambanovasystems/SambaLingo-Thai-Chat-70B"},{"id":"epoch:7d50637972633676ba76:training-compute","entityId":"epoch:7d50637972633676ba76","metric":"training-compute","value":8.1168e+23,"unit":"FLOP","observationPeriod":"2024-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Continuous Pre-training: We pack the pretraining mixture into sequences of length 4096 and pretrain with document attention as described in Section 3.2 of Iyer et al. (2022) to ensure we only attend to tokens in the context of the corresponding text document. We train with a global batch size of 1024, sequence length of 4096, maximum learning rate of 1e-4 with cosine decay, warm-up ratio of 0.01 and a weight decay of 0.1. Each expert is trained for a maximum of 4 epochs, following (Muennighoff et al., 2023). Notably, we train all model parameters, foregoing use of PEFT methods such as LoRA (Hu et al., 2022), which are known to be inferior to\nfull parameter training (Zhao et al., 2024a)(Sun et al., 2023)\" (https://arxiv.org/pdf/2404.05829, page 16). \n\n\"The base model adapts Llama-2-70b to Thai by training on 26 billion tokens from the Thai split of the Cultura-X dataset\" (https://huggingface.co/sambanovasystems/SambaLingo-Thai-Chat).  \n\nSince Llama is a dense model, assuming SambaLingo-Thai-Chat was trained for 4 epochs, the 6ND approximation yields\nPre-training compute\n= # of active parameters / forward pass * # of tokens * 6 FLOPs / token * # of epochs\n~= 7e10 parameters * 2.6e10 tokens * 6 FLOPs / token * 4 epochs\n~= 4.37e22 FLOPs\n\nFinetune: 6*7.0e+10*4.0e+9=1.7e+21\nBase model: 8.1e+23\nTotal: 8.116800e+23"},{"id":"epoch:59a900ab96c416ec3a80:parameters","entityId":"epoch:59a900ab96c416ec3a80","metric":"parameters","value":35000000.0,"unit":"parameters","observationPeriod":"2024-04-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:59a900ab96c416ec3a80:training-compute","entityId":"epoch:59a900ab96c416ec3a80","metric":"training-compute","value":7.28e+20,"unit":"FLOP","observationPeriod":"2024-04-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware: 16x NVIDIA A100 GPUs (3.12e14 FLOP/s per GPU)\n2. Training duration: 3 days directly reported = 259,200 seconds\n3. Utilization: 40% assumed\n4. Calculation: 3.12e14 FLOP/s \u00d7 16 GPUs \u00d7 259,200s \u00d7 0.40 = 5.18e20 FLOPs\n\nBase model: 209999999999999970000\nTotal: 209999999999999970000+5.18e20=728000000000000000000"},{"id":"epoch:b35f017dac59f153b0af:parameters","entityId":"epoch:b35f017dac59f153b0af","metric":"parameters","value":104000000000.0,"unit":"parameters","observationPeriod":"2024-04-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ba945dc03c6b9a73cafd:parameters","entityId":"epoch:ba945dc03c6b9a73cafd","metric":"parameters","value":33000000000.0,"unit":"parameters","observationPeriod":"2024-04-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"33B"},{"id":"epoch:ba945dc03c6b9a73cafd:training-compute","entityId":"epoch:ba945dc03c6b9a73cafd","metric":"training-compute","value":2.574e+23,"unit":"FLOP","observationPeriod":"2024-04-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Plan is to train on 2 trillion tokens, but most recent release is at 1.3T\n6 * 33B * 1.3 trillion = 2.574E23"},{"id":"epoch:c1ffee709d13dfeca461:parameters","entityId":"epoch:c1ffee709d13dfeca461","metric":"parameters","value":7720000000.0,"unit":"parameters","observationPeriod":"2024-04-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/sail/Sailor-7B-Chat"},{"id":"epoch:c1ffee709d13dfeca461:training-compute","entityId":"epoch:c1ffee709d13dfeca461","metric":"training-compute","value":1.7726e+23,"unit":"FLOP","observationPeriod":"2024-04-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The final pre-training corpus, SailCraft, is composed of approximately 200B tokens, integrating both SEA tokens and replay tokens, as elaborated in Section 4.4. We use the batch size of 4M tokens and the learning rate of 1e-4. Following a warmup period of 500 steps, the learning rate remains constant. This scheduling strategy encourages more transferable conclusions from simulations and allows for easier recovery from interrupted training\nsessions. Generally Sailor models consume around 200B tokens, completing a full pass through the SailCraft corpus once\" (https://arxiv.org/pdf/2404.03608, page 15).\n\nQwen1.5 is a dense model (https://huggingface.co/Qwen/Qwen1.5-7B), so assuming Sailor-7B-Chat was trained on SailCraft for 1 epoch, the 6ND approximation yields\nPre-training compute\n= # of active parameters / forward pass * # of tokens * 6 FLOPs / token\n~= 7.72e9 parameters * 2e11 tokens * 6 FLOPs / token\n~= 9.26e21 FLOPS\n\nBase model: 1.68e23\nTotal: 9.26e21+1.68e23=1.8e+23\n"},{"id":"epoch:98cdecc3771119bf2148:parameters","entityId":"epoch:98cdecc3771119bf2148","metric":"parameters","value":600000000.0,"unit":"parameters","observationPeriod":"2024-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Training on more than 12.5 million hours of diverse multilingual audio data, 600M-parameter Conformer RNN-T based Universal-1 achieves remarkable robustness\"\n"},{"id":"epoch:06d8ea929d4092e1f14b:parameters","entityId":"epoch:06d8ea929d4092e1f14b","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2024-04-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 4: \"We used the 12.5% capacity MoD variant to perform an isoFLOP analysis for 6e18, 2e19, and 1e20 FLOPs, training models varying in size from 60M to 3B parameters\""},{"id":"epoch:06d8ea929d4092e1f14b:training-compute","entityId":"epoch:06d8ea929d4092e1f14b","metric":"training-compute","value":1e+20,"unit":"FLOP","observationPeriod":"2024-04-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 4: \"We used the 12.5% capacity MoD variant to perform an isoFLOP analysis for 6e18, 2e19, and 1e20 FLOPs, training models varying in size from 60M to 3B parameters\""},{"id":"epoch:bc4bbd882d1c6effae8d:parameters","entityId":"epoch:bc4bbd882d1c6effae8d","metric":"parameters","value":4200000000.0,"unit":"parameters","observationPeriod":"2024-04-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"It uses a mixed expert model (MoE, Mixture-of-experts) architecture. The total parameter scale of the model is 25.8 billion, and the actual number of activated parameters is 4.2 billion."},{"id":"epoch:2f02e7a7ac429efbf2b0:parameters","entityId":"epoch:2f02e7a7ac429efbf2b0","metric":"parameters","value":136000000000.0,"unit":"parameters","observationPeriod":"2024-04-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"According to reports, APUS-xDAN-4.0 (MoE) is the first open source large model in China with a MoE architecture of more than 100 billion parameters, with a parameter scale of 136 billion. This is also the largest open source model in China. Among the previous large open source models in China, the largest parameter scale was Alibaba's Qianwen 72B with 72 billion parameters, with a parameter scale of 72 billion."},{"id":"epoch:a1f00024f09494be185c:parameters","entityId":"epoch:a1f00024f09494be185c","metric":"parameters","value":149700000.0,"unit":"parameters","observationPeriod":"2024-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 7 \nParams (M) (img+txt)\n86.3 + 63.4"},{"id":"epoch:a1f00024f09494be185c:training-compute","entityId":"epoch:a1f00024f09494be185c","metric":"training-compute","value":1.3423599e+22,"unit":"FLOP","observationPeriod":"2024-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 490*10^9 tokens [see dataset size notes] * 30.50 epochs * 149700000.00 parameters = 1.3423599e+22 FLOP\n\n'Likely' confidence because I am not sure about dataset size"},{"id":"epoch:d80cf6849c2aa00c3dbd:parameters","entityId":"epoch:d80cf6849c2aa00c3dbd","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d80cf6849c2aa00c3dbd:training-compute","entityId":"epoch:d80cf6849c2aa00c3dbd","metric":"training-compute","value":4.2e+22,"unit":"FLOP","observationPeriod":"2024-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"80 nodes, each having 8 Nvidia A100 Sxm 40GB GPUs\n\n6 FLOP / token / parameter * 7*10^9 parameters * 1 * 10^12 tokens = 4.2e+22 FLOP"},{"id":"epoch:128fc9151537793e2ef1:parameters","entityId":"epoch:128fc9151537793e2ef1","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2024-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:128fc9151537793e2ef1:training-compute","entityId":"epoch:128fc9151537793e2ef1","metric":"training-compute","value":1.44e+22,"unit":"FLOP","observationPeriod":"2024-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"80 nodes, each having 8 Nvidia A100 Sxm 40GB GPUs\n\n6 FLOP / token / parameter * 3*10^9 parameters * 0.8 * 10^12 tokens = 1.44e+22 FLOP"},{"id":"epoch:f70af275eead2d4c1a4c:parameters","entityId":"epoch:f70af275eead2d4c1a4c","metric":"parameters","value":12000000000.0,"unit":"parameters","observationPeriod":"2024-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f70af275eead2d4c1a4c:training-compute","entityId":"epoch:f70af275eead2d4c1a4c","metric":"training-compute","value":8.64e+22,"unit":"FLOP","observationPeriod":"2024-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"80 nodes, each having 8 Nvidia A100 Sxm 40GB GPUs\n\n6 FLOP / token / parameter * 12*10^9 parameters * 1.2 * 10^12 tokens = 8.64e+22 FLOP"},{"id":"epoch:cf9daaea454c3f845814:parameters","entityId":"epoch:cf9daaea454c3f845814","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2024-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:48caa336b046b95e093d:parameters","entityId":"epoch:48caa336b046b95e093d","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2024-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a1dd7635c81cdf59460e:parameters","entityId":"epoch:a1dd7635c81cdf59460e","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2024-03-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Fine-tuned FLAN-T5 models ranging from 80M to 3B"},{"id":"epoch:a16fc094864779573173:training-compute","entityId":"epoch:a16fc094864779573173","metric":"training-compute","value":9.26e+24,"unit":"FLOP","observationPeriod":"2024-03-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Lower bound is taken from Grok-1 estimation\nUpper bound is taken from Grok-2 estimation\n\ngeometric mean: \nsqrt(2.90000000001*29.6)*10^24 = 9.26e+24"},{"id":"epoch:1be9bd28f6cc874e32f0:parameters","entityId":"epoch:1be9bd28f6cc874e32f0","metric":"parameters","value":51600000000.0,"unit":"parameters","observationPeriod":"2024-03-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"51.6B from https://huggingface.co/ai21labs/Jamba-v0.1\n(12B active parameters, 52B total available parameters)"},{"id":"epoch:dfffe42cdba410c69966:parameters","entityId":"epoch:dfffe42cdba410c69966","metric":"parameters","value":132000000000.0,"unit":"parameters","observationPeriod":"2024-03-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"132B mixture of experts. 36B parameters active per inference"},{"id":"epoch:dfffe42cdba410c69966:training-compute","entityId":"epoch:dfffe42cdba410c69966","metric":"training-compute","value":2.6e+24,"unit":"FLOP","observationPeriod":"2024-03-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Mixture of Experts (MoE)\n\n36 billion active params * 12 trillion tokens * 6 ~= 2.6e24\nhttps://www.wolframalpha.com/input?i=6+FLOP+*+36+billion+*+12+trillion\n\nalso, it was trained on 3072 NVIDIA H100s, but with an unclear timeframe (end-end process was three months, including evals and red-teaming)."},{"id":"epoch:85bd3634557e4e1c1dbe:parameters","entityId":"epoch:85bd3634557e4e1c1dbe","metric":"parameters","value":75000000000.0,"unit":"parameters","observationPeriod":"2024-03-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:539125d231543a7e88a6:parameters","entityId":"epoch:539125d231543a7e88a6","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2024-03-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6176d4e099235c58cc4f:parameters","entityId":"epoch:6176d4e099235c58cc4f","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2024-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6176d4e099235c58cc4f:training-compute","entityId":"epoch:6176d4e099235c58cc4f","metric":"training-compute","value":3.09999999999999e+21,"unit":"FLOP","observationPeriod":"2024-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 8x NVIDIA A100 GPUs (3.12e14 FLOP/s per GPU)\n\n2. Training duration: Directly provided - 36 days total (10 days pre-training + 26 days translation training) = 3,110,400 seconds\n\n3. Utilization rate: 40%\n\n4. Final calculation:\n3.12e14 FLOP/s \u00d7 8 GPUs \u00d7 3,110,400s \u00d7 0.4 = 3.1e21 FLOPs"},{"id":"epoch:83c493365cfd696e5b5f:parameters","entityId":"epoch:83c493365cfd696e5b5f","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2024-03-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" The Xuanji Yuheng language interaction big model boasts enhanced industry capabilities, security and reliability, and full-stack localization support. It has developed multiple parameter sizes, including 9 billion, 13.9 billion, 57 billion, and 100 billion"},{"id":"epoch:8104170c6d3df3cdf639:parameters","entityId":"epoch:8104170c6d3df3cdf639","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-03-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B (Table 3)"},{"id":"epoch:3ea1d105e77ed1ad805f:parameters","entityId":"epoch:3ea1d105e77ed1ad805f","metric":"parameters","value":774000000.0,"unit":"parameters","observationPeriod":"2024-03-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 4."},{"id":"epoch:7b54e3d3d61e6ef54c23:training-compute","entityId":"epoch:7b54e3d3d61e6ef54c23","metric":"training-compute","value":5.1757056e+20,"unit":"FLOP","observationPeriod":"2024-03-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4 nodes * 8 GPUs * 144 hours * 3600 sec / hour * 312000000000000 FLOP / GPU / hour * 0.3 [assumed utilization] / 3 models = 5.1757056e+20 FLOP"},{"id":"epoch:14847087a9475207e804:parameters","entityId":"epoch:14847087a9475207e804","metric":"parameters","value":86000000.0,"unit":"parameters","observationPeriod":"2024-03-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:14847087a9475207e804:training-compute","entityId":"epoch:14847087a9475207e804","metric":"training-compute","value":2.1000000000000013e+21,"unit":"FLOP","observationPeriod":"2024-03-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware: 24x V100 32GB GPUs (1.30e+14 FLOPs/s per GPU)\n2. Training duration: 20 days (directly provided) = 1.728e+6 seconds\n3. Utilization: 40%\n4. Calculation: \n   24 GPUs \u00d7 1.30e+14 FLOPs/s \u00d7 1.728e+6 seconds \u00d7 0.4 utilization = 2.1e+21 FLOPs"},{"id":"epoch:a29845109130b50f89b2:parameters","entityId":"epoch:a29845109130b50f89b2","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2024-03-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"30B"},{"id":"epoch:a29845109130b50f89b2:training-compute","entityId":"epoch:a29845109130b50f89b2","metric":"training-compute","value":4.86e+23,"unit":"FLOP","observationPeriod":"2024-03-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Pre-trained on ~2B image-text pairs and 2T tokens (Table 2). Each image is 144 tokens, so the images are ~300B tokens.\nThen additional multimodal training for 400B tokens, for a total of ~2.7T tokens.\n\nThis is the final training recipe: \"We initialize both the image encoder and the underlying LLM decoder weights for MM1 from in-house pre-trained models2. We then perform multimodal pre-training on the above data mix for 200k steps (approx. 400B tokens).\"\n\nCompute  = 6ND = 6 * 2.7 trillion * 30 billion = 4.86e23\n\nmaybe the size of the visual connector is relevant"},{"id":"epoch:ac9b2b9aba7013c5b6a6:parameters","entityId":"epoch:ac9b2b9aba7013c5b6a6","metric":"parameters","value":20000000000.0,"unit":"parameters","observationPeriod":"2024-03-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"20B"},{"id":"epoch:1aa88a065e9435027de3:parameters","entityId":"epoch:1aa88a065e9435027de3","metric":"parameters","value":35000000000.0,"unit":"parameters","observationPeriod":"2024-03-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/CohereForAI/c4ai-command-r-v01"},{"id":"epoch:eab82cc564a9b063b465:parameters","entityId":"epoch:eab82cc564a9b063b465","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-03-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8b"},{"id":"epoch:eab82cc564a9b063b465:training-compute","entityId":"epoch:eab82cc564a9b063b465","metric":"training-compute","value":2.4e+20,"unit":"FLOP","observationPeriod":"2024-03-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 5000000000 tokens * 8*10^9 parameters = 2.4e+20 FLOP\n\nI am not confident about amount of epochs and whether the model is dense"},{"id":"epoch:ef0baeb61d53a1104b5f:parameters","entityId":"epoch:ef0baeb61d53a1104b5f","metric":"parameters","value":800000000.0,"unit":"parameters","observationPeriod":"2024-03-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:7460cd0200b291e6fb5a:parameters","entityId":"epoch:7460cd0200b291e6fb5a","metric":"parameters","value":928000000.0,"unit":"parameters","observationPeriod":"2024-03-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The model is a BiT-M ResNet101x3\"\n\nBiT-M has 928M parameters, I assume same here"},{"id":"epoch:ae22cbad757cb3c6e541:parameters","entityId":"epoch:ae22cbad757cb3c6e541","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-03-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ae22cbad757cb3c6e541:training-compute","entityId":"epoch:ae22cbad757cb3c6e541","metric":"training-compute","value":1.0264959e+23,"unit":"FLOP","observationPeriod":"2024-03-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8.4e+22 FLOP [base model] + 1.8649595e+22 FLOP [finetune compute] = 1.0264959e+23 FLOP"},{"id":"epoch:a01e9e4611785aa6de79:parameters","entityId":"epoch:a01e9e4611785aa6de79","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2024-03-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a01e9e4611785aa6de79:training-compute","entityId":"epoch:a01e9e4611785aa6de79","metric":"training-compute","value":8.6547326e+21,"unit":"FLOP","observationPeriod":"2024-03-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.9e+21 FLOP [base model] + 4.7547326e+21 FLOP [finetune compute] = 8.6547326e+21 FLOP"},{"id":"epoch:0b2c2443e355f20f4f7d:training-compute","entityId":"epoch:0b2c2443e355f20f4f7d","metric":"training-compute","value":8.000001e+24,"unit":"FLOP","observationPeriod":"2024-03-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Inflection-1 used approximately 4% the training FLOPs of GPT-4 and, on average, performed at approximately 72% GPT-4 level on a diverse range of IQ-oriented tasks. Inflection-2.5, now powering Pi, achieves more than 94% the average performance of GPT-4 despite using only 40% the training FLOPs.\"\n\nThis is a weird one - we estimated GPT-4 at 2.1e25 FLOP (which could be off somewhat, or Inflection could believe a different number). 40% of that is ~8e24. \n"},{"id":"epoch:b18d08e2051f0126b972:parameters","entityId":"epoch:b18d08e2051f0126b972","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-03-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:95697efde30d58472433:parameters","entityId":"epoch:95697efde30d58472433","metric":"parameters","value":250000000000.0,"unit":"parameters","observationPeriod":"2024-03-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"It has 250 billion parameters that are adjustable during training to generate outputs or make predictions.\""},{"id":"epoch:95697efde30d58472433:training-compute","entityId":"epoch:95697efde30d58472433","metric":"training-compute","value":1.05e+25,"unit":"FLOP","observationPeriod":"2024-03-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*250B*7T=1.05e+25"},{"id":"epoch:8a4dc34c2cbc616f6a7e:training-compute","entityId":"epoch:8a4dc34c2cbc616f6a7e","metric":"training-compute","value":8.76096e+20,"unit":"FLOP","observationPeriod":"2024-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000*2600*3600*0.3 = 8.76096e+20"},{"id":"epoch:12f35dbe530034fa54a0:parameters","entityId":"epoch:12f35dbe530034fa54a0","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2024-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:62ee67241c475de817c1:parameters","entityId":"epoch:62ee67241c475de817c1","metric":"parameters","value":1000000000000.0,"unit":"parameters","observationPeriod":"2024-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The company has already launched the Step series of foundation models, which includes Step-2, a cutting-edge trillion-parameter Mixture of Experts (MoE) language model\" according to the organization card on HF.\nhttps://huggingface.co/stepfun-ai"},{"id":"epoch:91a7195fcea94a33bb55:parameters","entityId":"epoch:91a7195fcea94a33bb55","metric":"parameters","value":16000000000.0,"unit":"parameters","observationPeriod":"2024-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"16B"},{"id":"epoch:86dc77c82c1b3e41f0aa:parameters","entityId":"epoch:86dc77c82c1b3e41f0aa","metric":"parameters","value":34000000000.0,"unit":"parameters","observationPeriod":"2024-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"34B"},{"id":"epoch:308c93994118cb2def8e:parameters","entityId":"epoch:308c93994118cb2def8e","metric":"parameters","value":15000000000.0,"unit":"parameters","observationPeriod":"2024-02-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"15B"},{"id":"epoch:308c93994118cb2def8e:training-compute","entityId":"epoch:308c93994118cb2def8e","metric":"training-compute","value":3.87e+23,"unit":"FLOP","observationPeriod":"2024-02-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"estimation is given in Table 6 "},{"id":"epoch:fa29a5f910c3e3d69948:parameters","entityId":"epoch:fa29a5f910c3e3d69948","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-02-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:fa29a5f910c3e3d69948:training-compute","entityId":"epoch:fa29a5f910c3e3d69948","metric":"training-compute","value":1.55e+23,"unit":"FLOP","observationPeriod":"2024-02-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"estimation is given in Table 6 "},{"id":"epoch:2e93250b8e57225ed56d:parameters","entityId":"epoch:2e93250b8e57225ed56d","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2024-02-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3B"},{"id":"epoch:2e93250b8e57225ed56d:training-compute","entityId":"epoch:2e93250b8e57225ed56d","metric":"training-compute","value":5.94e+22,"unit":"FLOP","observationPeriod":"2024-02-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"estimation is given in Table 6 "},{"id":"epoch:16d2793addd6bbab9dad:parameters","entityId":"epoch:16d2793addd6bbab9dad","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2024-02-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:16d2793addd6bbab9dad:training-compute","entityId":"epoch:16d2793addd6bbab9dad","metric":"training-compute","value":1.5848931924611135e+22,"unit":"FLOP","observationPeriod":"2024-02-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 1.a\n10^(1/5)*10^22 = 15848931924611134852021.0137339150701326944213382503906831629"},{"id":"epoch:fc40dac258fadfdf8914:parameters","entityId":"epoch:fc40dac258fadfdf8914","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-02-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"All three model families are trained at a range of model scales from 100M to 7B parameters, with an additional Griffin model with 14 billion parameters.\""},{"id":"epoch:fc40dac258fadfdf8914:training-compute","entityId":"epoch:fc40dac258fadfdf8914","metric":"training-compute","value":3.95e+21,"unit":"FLOP","observationPeriod":"2024-02-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 1.a\nDigitized and found 3.95e21 for biggest Hawk model"},{"id":"epoch:865ec38971a6f8511964:parameters","entityId":"epoch:865ec38971a6f8511964","metric":"parameters","value":8000000.0,"unit":"parameters","observationPeriod":"2024-02-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Largest model trained has 8M params. Note actual model used in real-world experiments appears to be 2M, to improve latency. See sections 5.1 and 5.9."},{"id":"epoch:c1f963a38523c81860a1:parameters","entityId":"epoch:c1f963a38523c81860a1","metric":"parameters","value":57300000.0,"unit":"parameters","observationPeriod":"2024-02-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"57.3M"},{"id":"epoch:c1f963a38523c81860a1:training-compute","entityId":"epoch:c1f963a38523c81860a1","metric":"training-compute","value":1.74258e+17,"unit":"FLOP","observationPeriod":"2024-02-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*57300000.00 parameters *118000 images *500 epochs = 2.02842e+16 (likely an underestimation)\n\n(speculative confidence because I don't know the amount of tokens per image)\n\nassuming(!) batch size 64 -> number of updates per epoch = 118000/64 = 1844\n\n 189.0 Gigaflops * 1844 * 500 = 1.74258e+17"},{"id":"epoch:6e259aadaa272d5d2632:parameters","entityId":"epoch:6e259aadaa272d5d2632","metric":"parameters","value":650000000.0,"unit":"parameters","observationPeriod":"2024-02-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6e259aadaa272d5d2632:training-compute","entityId":"epoch:6e259aadaa272d5d2632","metric":"training-compute","value":1.0500000000000005e+21,"unit":"FLOP","observationPeriod":"2024-02-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware: 7x NVIDIA A100 GPUs (3.12e14 FLOP/s per GPU)\n2. Training duration: 2 weeks (directly provided) = 1,209,600 seconds\n3. Utilization: 40% (0.4)\n4. Calculation: 3.12e14 FLOP/s \u00d7 7 GPUs \u00d7 1,209,600s \u00d7 0.4 = 1.05e21 FLOPs"},{"id":"epoch:1680d89ac81d60a5b628:parameters","entityId":"epoch:1680d89ac81d60a5b628","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-02-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e4f6afe9829b5831ec42:parameters","entityId":"epoch:e4f6afe9829b5831ec42","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-02-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Based on a StripedHyena architecture, context length of 131 kilobases"},{"id":"epoch:e4f6afe9829b5831ec42:training-compute","entityId":"epoch:e4f6afe9829b5831ec42","metric":"training-compute","value":2.0000000001e+22,"unit":"FLOP","observationPeriod":"2024-02-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In total, Evo was trained on approximately 340B tokens, using approximately 2e22 FLOPS\""},{"id":"epoch:67156e3e1c7005a3a1d7:parameters","entityId":"epoch:67156e3e1c7005a3a1d7","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-02-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:67156e3e1c7005a3a1d7:training-compute","entityId":"epoch:67156e3e1c7005a3a1d7","metric":"training-compute","value":2.8735486e+22,"unit":"FLOP","observationPeriod":"2024-02-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*70*10^9*100*10^9 = 42000000000000000000000 (4.2e+22)\n\nFigure 3 suggests it 41.2 times more energy efficient than LLAMA 70B (which is estimated 8.1e+23 FLOPS  -> 1.9660194e+22 FLOPs \n\ngeometric mean -> 2.8735486e+22"},{"id":"epoch:f65808cf94af24b9203a:parameters","entityId":"epoch:f65808cf94af24b9203a","metric":"parameters","value":15000000000.0,"unit":"parameters","observationPeriod":"2024-02-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"15b"},{"id":"epoch:f65808cf94af24b9203a:training-compute","entityId":"epoch:f65808cf94af24b9203a","metric":"training-compute","value":7.5005116e+23,"unit":"FLOP","observationPeriod":"2024-02-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 FLOP/token/parameter * 15*10^9 parameters * 8*10^12 tokens = 7.2e+23 FLOP\n\n\"Nemotron-4 was trained using 384 DGX H100 nodes; each node contains 8 H100 80GB SXM5 GPUs based on the NVIDIA Hopper architecture (NVIDIA, 2022). Each H100 GPU has a peak throughput of 989 teraFLOP/s when doing 16-bit floating point (bfloat16) arithmetic without sparsity.\n\nTable 2 reports more detailed training schedule:\n\n989*10^12 FLOP/sec * 3600 sec/hour * 24 hours * (768 gpus * 0.343 [reported utilization] * 0.8 days +  1536 gpus * 0.333 [reported utilization] * 0.4 days + 2304 gpus * 0.305 [reported utilization] * 11.9 days) = 7.5005116e+23 FLOP"},{"id":"epoch:707cb31a2b6f209c5da4:parameters","entityId":"epoch:707cb31a2b6f209c5da4","metric":"parameters","value":3500000000.0,"unit":"parameters","observationPeriod":"2024-02-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The Model Description section of [1] says that the model \u201cfollows the same architecture as Stable Diffusion XL.\u201d It also says that the model is a latent diffusion model that uses OpenCLIP-ViT/G and CLIP-ViT/L as text encoders, which seemingly would make it identical to Stable Diffusion XL [2]. Therefore, I will assume that Playground v2.5 has the same number of parameters as the Stable Diffusion XL 1.0 base model, which has 3.5B model parameters [3].\n\n1. https://huggingface.co/playgroundai/playground-v2.5-1024px-aesthetic\n2. https://arxiv.org/pdf/2307.01952\n3. https://stability.ai/news/stable-diffusion-sdxl-1-announcement"},{"id":"epoch:e7d85b3c6b7c725c295d:parameters","entityId":"epoch:e7d85b3c6b7c725c295d","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-02-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e7d85b3c6b7c725c295d:training-compute","entityId":"epoch:e7d85b3c6b7c725c295d","metric":"training-compute","value":8.412e+22,"unit":"FLOP","observationPeriod":"2024-02-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 8x NVIDIA RTX A6000 GPUs (3.87e13 FLOP/s per GPU)\n\n2. Training duration: Provided directly - Stage 1: 6 days, Stage 2: 5 days (Total: 11 days = 950,400 seconds)\n\n3. Utilization rate: 40%\n\n4. Calculation:\n(8 GPUs \u00d7 3.87e13 FLOP/s/GPU) \u00d7 950,400 seconds \u00d7 0.4 utilization = 1.2e20 FLOPs\n(Stage 1: 1.603e20 + Stage 2: 1.338e20) \u00d7 0.4 = 1.2e20 FLOPs\n\nBase model: 8.4e+22\ntotal : 84120000000000000000000"},{"id":"epoch:37dc3f4987a6aed0ea23:training-compute","entityId":"epoch:37dc3f4987a6aed0ea23","metric":"training-compute","value":1.8999999999999996e+19,"unit":"FLOP","observationPeriod":"2024-02-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 1x NVIDIA A100 GPU (3.12e14 FLOPs/s for FP16)\n2. Training duration: 41.7 hours (provided directly) = 150,120 seconds\n3. Utilization rate: 40% (assumed)\n4. Calculation: 3.12e14 FLOPs/s \u00d7 1 GPU \u00d7 150,120s \u00d7 0.4 = 1.873e19 FLOPs"},{"id":"epoch:72831bfc1a231b5c30c9:parameters","entityId":"epoch:72831bfc1a231b5c30c9","metric":"parameters","value":8540000000.0,"unit":"parameters","observationPeriod":"2024-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Safetensors \nModel size 8.54B params"},{"id":"epoch:72831bfc1a231b5c30c9:training-compute","entityId":"epoch:72831bfc1a231b5c30c9","metric":"training-compute","value":3.0744e+23,"unit":"FLOP","observationPeriod":"2024-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*6000000000000*8540000000=3.0744e+23"},{"id":"epoch:7263dfbb2c06bcb681f8:parameters","entityId":"epoch:7263dfbb2c06bcb681f8","metric":"parameters","value":175000000000.0,"unit":"parameters","observationPeriod":"2024-02-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Two models are trained for epoch to evaluate the MegaScale training system; one model with 175B and another with 530B parameters. This entry reports the 175B model.\n\nThere is a third production model mentioned, with fewer details."},{"id":"epoch:7263dfbb2c06bcb681f8:training-compute","entityId":"epoch:7263dfbb2c06bcb681f8","metric":"training-compute","value":2.7385671436e+23,"unit":"FLOP","observationPeriod":"2024-02-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2 gives details for the 175B model. Looking at the largest 1 epoch run with 12288 GPUs:\n2166.3 aggregate PFlops/sec * 1.75 days * 24 hours/day * 3600 seconds/hour = 3.275e23\n\nThis is consistent with the theoretical computation counting estimate, if they factor MFU rate into their 2166.3 figure:\n2 \u00d7 175B params \u00d7 3 \u00d7 300B tokens \u00d7 1 epoch = 2.29e23\n\nI use the geometric mean of these two:\n(3.275e23 + 2.29e23) / 2 = 2.74e23"},{"id":"epoch:e10681a6329bcbfb2c59:parameters","entityId":"epoch:e10681a6329bcbfb2c59","metric":"parameters","value":530000000000.0,"unit":"parameters","observationPeriod":"2024-02-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Two models are trained for epoch to evaluate the MegaScale training system; one model with 175B and another with 530B parameters. This entry reports the 530B model.\n\nThere is a third production model mentioned, with fewer details."},{"id":"epoch:e10681a6329bcbfb2c59:training-compute","entityId":"epoch:e10681a6329bcbfb2c59","metric":"training-compute","value":9.6910000000001e+23,"unit":"FLOP","observationPeriod":"2024-02-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"175B models uses 3.2e23 FLOPs (Table 2, bottom row)\n\nWith constant dataset size and utilization, FLOPs should scale linearly in # parameters, so: 3.2e23 * (530/175) = 9.7e23"},{"id":"epoch:424234e6c06fb709aac7:parameters","entityId":"epoch:424234e6c06fb709aac7","metric":"parameters","value":530000000000.0,"unit":"parameters","observationPeriod":"2024-02-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Production run is stated to have \"hundreds of billions of parameters\". Since the authors also do a number of experiments with a 530B model, I speculate they've used 530B for the production model."},{"id":"epoch:424234e6c06fb709aac7:training-compute","entityId":"epoch:424234e6c06fb709aac7","metric":"training-compute","value":3.9e+24,"unit":"FLOP","observationPeriod":"2024-02-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Speculative. The model is stated to have trained for \"several weeks\". Assuming 530B parameters and \"several\" = 3, compute can be estimated from the 175B model's stated PFLOP/sec:\n2166.3 aggregate PFlops/sec * 3 weeks * 7 days/week * 24 hours/day * 3600 seconds/hour = 3.9e+24.\nAs an upper bound, say 8e+24. "},{"id":"epoch:3135a6a49fde8a56077b:parameters","entityId":"epoch:3135a6a49fde8a56077b","metric":"parameters","value":10700000000.0,"unit":"parameters","observationPeriod":"2024-02-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"When combined with the tokenizer and action\nmodel this brings the total to 10.7B parameters,\ntrained on 942B tokens, which we refer to as the\nGenie model. "},{"id":"epoch:3135a6a49fde8a56077b:training-compute","entityId":"epoch:3135a6a49fde8a56077b","metric":"training-compute","value":6.6e+22,"unit":"FLOP","observationPeriod":"2024-02-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 12"},{"id":"epoch:621ac470f7eb92be08a2:parameters","entityId":"epoch:621ac470f7eb92be08a2","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-02-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:621ac470f7eb92be08a2:training-compute","entityId":"epoch:621ac470f7eb92be08a2","metric":"training-compute","value":5.0000000000000004e+22,"unit":"FLOP","observationPeriod":"2024-02-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Finally, we performed\na scaling study of this combination up to a model size of\n8B parameters and 5 \u00d7 1022 training FLOPs."},{"id":"epoch:7137fa50b350ad1c0431:parameters","entityId":"epoch:7137fa50b350ad1c0431","metric":"parameters","value":8538074112.0,"unit":"parameters","observationPeriod":"2024-02-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2, sum of embedding and non-embedding parameters"},{"id":"epoch:7137fa50b350ad1c0431:training-compute","entityId":"epoch:7137fa50b350ad1c0431","metric":"training-compute","value":3.07e+23,"unit":"FLOP","observationPeriod":"2024-02-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND aproximation 6*8.54B*6T = 3.07e23\n\"Gemma 2B and 7B are trained on 2T and 6T tokens respectively of primarily-English data from web documents, mathematics, and code.\"\n\nAs confirmation: \"We estimate the carbon emissions from pretraining the Gemma models to be \u223c 131 \ud835\udc61\ud835\udc36\ud835\udc422\ud835\udc52\ud835\udc5e. \"\n\nU.S. avg CO2 per kWh is ~0.87lbs \n131 tCO2 * 2000 lb/t * (1 kWh/0.87lb) = 3.01e5 kWh\n\nPer SemiAnalysis TPU v5e uses ~ 5x less power than H100, so ~140 W TDP\n3.01e5 kWh * 1000 W/kW * 1 TPUv5e/140 W  = 2.15e6 TPUv5e-hours\n\nIn bf16 precision, TPUv5e has peak performance of 197 TF/s, so:\n2.15e6 * 3600 * 197e12 * 0.3 = 4.57e23"},{"id":"epoch:00951c50780e5067026d:training-compute","entityId":"epoch:00951c50780e5067026d","metric":"training-compute","value":7.500000000000002e+19,"unit":"FLOP","observationPeriod":"2024-02-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 1x NVIDIA A100 GPU (3.12e14 FLOP/s)\n\n2. Training duration: 7 days (directly provided) = 604,800 seconds\n\n3. Utilization rate: 40%\n\n4. Final calculation:\n1 GPU \u00d7 3.12e14 FLOP/s \u00d7 604,800s \u00d7 0.4 = 7.55e19 FLOPs"},{"id":"epoch:d4a5cb67acceb9350685:parameters","entityId":"epoch:d4a5cb67acceb9350685","metric":"parameters","value":2506434560.0,"unit":"parameters","observationPeriod":"2024-02-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2, sum of embedding and non-embedding parameters:\n\n2B 524,550,144 + 1,981,884,416 = 2506434560"},{"id":"epoch:d4a5cb67acceb9350685:training-compute","entityId":"epoch:d4a5cb67acceb9350685","metric":"training-compute","value":4.5115822e+22,"unit":"FLOP","observationPeriod":"2024-02-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*2506434560.00 parameters * 3*10^12 training tokens = 4.5115822e+22\n\n(assuming 1 epoch)"},{"id":"epoch:f6937647536de426f140:parameters","entityId":"epoch:f6937647536de426f140","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-02-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:5333a412228815691bb4:parameters","entityId":"epoch:5333a412228815691bb4","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-02-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:b9eddb3414c8f2534b50:parameters","entityId":"epoch:b9eddb3414c8f2534b50","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2024-02-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b9eddb3414c8f2534b50:training-compute","entityId":"epoch:b9eddb3414c8f2534b50","metric":"training-compute","value":7.200353201209069e+23,"unit":"FLOP","observationPeriod":"2024-02-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"GPU hours: 86400s*9.9e+14*2*0.4=6.8e+19\n\n6ND: 6*8000000000*380000001=18240000048000000000\n\nGMean(1.8e+19,6.8e+19)=3.5e+19\n\nBase model: 7.2e+23\n\nTotal: 720035320120906900000000"},{"id":"epoch:e865c6a4fb05941d29cc:parameters","entityId":"epoch:e865c6a4fb05941d29cc","metric":"parameters","value":630000000.0,"unit":"parameters","observationPeriod":"2024-02-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"They train 200M and 630M models (see e.g. table 6)"},{"id":"epoch:e865c6a4fb05941d29cc:training-compute","entityId":"epoch:e865c6a4fb05941d29cc","metric":"training-compute","value":1.6387080192e+21,"unit":"FLOP","observationPeriod":"2024-02-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 7 shows that each example (16 frames of 224*224 resoultion) is represented with 1568 tokens. \n\nTable 8 specifies (for ViT-H/16_224):\nBatch size = 3072\nIterations = 90000\n\nTotal training tokens: 90000 iterations *3072 batch size * 1568 tokens =433520640000\n\nTraining compute using 6ND (ignoring fine-tuning): 6*630000000*433520640000 = 1638708019200000000000"},{"id":"epoch:d374038cce7bbddf8b3a:parameters","entityId":"epoch:d374038cce7bbddf8b3a","metric":"parameters","value":175000000000.0,"unit":"parameters","observationPeriod":"2024-02-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"175B"},{"id":"epoch:7a04d98913a1acc2adfb:parameters","entityId":"epoch:7a04d98913a1acc2adfb","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-02-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B  - fine tune of mT5 - from last page - model card "},{"id":"epoch:c0958a29a22228c60bc6:parameters","entityId":"epoch:c0958a29a22228c60bc6","metric":"parameters","value":1474624.0,"unit":"parameters","observationPeriod":"2024-02-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Figure 3: The prediction module takes a 1792x1 feature vector as its input, which is then partitioned into two streams: The first stream processes the first 1024 indices of the feature vector through a 512-node protein-specific linear layer, followed by a ReLU activation function. Concurrently, the second stream processes the latter 768 feature indices through a similar 512-node moleculespecific linear layer, also followed by a ReLU activation.  Outputs from both streams are concatenated into a single vector of 1024 elements. This combined vector is passed through a batch normalization layer with a momentum of 0.9 and epsilon of 0.001. The vector is then fed through the 512-node Linear Layer 1 and ReLU activation. A dropout layer with a probability of 20% is then applied to mitigate overfitting.  Following the dropout layer, the prediction module continues to reduce the feature space with the 64-node Linear Layer 2 with ReLU activation, before reaching the single node Linear Layer 3. This layer outputs the predicted scalar value representing the normalized negative log10 affinity value, which is then un-normalized.\"  (512 * 1024) + (512 * 768) + (512 * 1024) + (64 * 512) + 64 = 1474624"},{"id":"epoch:c0958a29a22228c60bc6:training-compute","entityId":"epoch:c0958a29a22228c60bc6","metric":"training-compute","value":3.900846548437782e+22,"unit":"FLOP","observationPeriod":"2024-02-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1474624 connections; 90,000 training examples; table a4 mentions 60 training rounds\n\n6*1474624*90000*60=47777817600000\n\nIncluding base models:\n8465436600000000000 + 3.9e+22 + 47777817600000 = 39008465484377820000000"},{"id":"epoch:aee96601b7293a27663d:parameters","entityId":"epoch:aee96601b7293a27663d","metric":"parameters","value":5120000000.0,"unit":"parameters","observationPeriod":"2024-02-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we are providing two checkpoints for Stage C, two for Stage B and one for Stage A. Stage C comes with a 1 billion and 3.6 billion parameter version, but we highly recommend using the 3.6 billion version, as most work was put into its finetuning. The two versions for Stage B amount to 700 million and 1.5 billion parameters. Both achieve great results, however the 1.5 billion excels at reconstructing small and fine details. Therefore, you will achieve the best results if you use the larger variant of each. Lastly, Stage A contains 20 million parameters and is fixed due to its small size.\"\n\n3.6B + 1.5B + 20M = 5.12B"},{"id":"epoch:9a31bb1d5dc17f059050:training-compute","entityId":"epoch:9a31bb1d5dc17f059050","metric":"training-compute","value":3.399999999999988e+19,"unit":"FLOP","observationPeriod":"2024-02-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup:\n- VAE stage: 1x NVIDIA RTX A6000 (3.87e13 FLOP/s)\n- UNet stage: 1x NVIDIA A100 40GB (3.12e14 FLOP/s)\n\n2. Training duration (provided directly):\n- VAE: 24 GPU-hours (86,400 seconds)\n- UNet: 72 GPU-hours (259,200 seconds)\n\n3. Utilization rate: 40% for both stages\n\n4. Calculation:\nVAE: 3.87e13 FLOP/s \u00d7 86,400s \u00d7 0.4 = 1.34e18 FLOPs\nUNet: 3.12e14 FLOP/s \u00d7 259,200s \u00d7 0.4 = 3.24e19 FLOPs\nTotal = 1.34e18 + 3.24e19 = 3.4e19 FLOPs"},{"id":"epoch:6f566d7985ea14064e7b:parameters","entityId":"epoch:6f566d7985ea14064e7b","metric":"parameters","value":270000000.0,"unit":"parameters","observationPeriod":"2024-02-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our largest model has roughly 270 million parameters.\""},{"id":"epoch:6f566d7985ea14064e7b:training-compute","entityId":"epoch:6f566d7985ea14064e7b","metric":"training-compute","value":1.035671832e+22,"unit":"FLOP","observationPeriod":"2024-02-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"10356718320000000065536 FLOP\n\"Board states \ud835\udc60 are encoded as FEN strings which we convert to fixed-length strings of 77 characters where the ASCII-code of each character is one token.\" so 77 tokens for board + 1 token for action \"For our largest training dataset, based on 10M games, this results in 15.32B action-value estimates\"\nso input is 78 tokens for each action-value\nnumber of tokens = 1194960000000.0\nThe model is dense transformer\n\" We train for 10 million\nsteps, which corresponds to 2.67 epochs for a batch\nsize of 4096 with 15.32B data points \", but in appendix A.2 there is mention of 5.35 epochs\nI have used higher value from 5.35 and 2.67,\nProbably final they trained model for 5.35 epochs and used checkpoint from 2.67 as final model.\naproximation 6ND for 5.35 epochs = 6*270e6*1194960000000.0 * 5.35 =  10356718320000000065536\n"},{"id":"epoch:e242e0b8ee81c0864d3a:parameters","entityId":"epoch:e242e0b8ee81c0864d3a","metric":"parameters","value":650000000.0,"unit":"parameters","observationPeriod":"2024-02-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:880f131a1249452f8a40:parameters","entityId":"epoch:880f131a1249452f8a40","metric":"parameters","value":50000000000.0,"unit":"parameters","observationPeriod":"2024-02-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"tens of billions of parameters\"\n\n~50B parameters with \"Likely\" confidence"},{"id":"epoch:2b8030ce3ab41eb4fb9b:parameters","entityId":"epoch:2b8030ce3ab41eb4fb9b","metric":"parameters","value":643000000.0,"unit":"parameters","observationPeriod":"2024-02-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"643M from Table S1"},{"id":"epoch:2b8030ce3ab41eb4fb9b:training-compute","entityId":"epoch:2b8030ce3ab41eb4fb9b","metric":"training-compute","value":1.0193977e+22,"unit":"FLOP","observationPeriod":"2024-02-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 128 NVIDIA V100 GPUs (1.25e14 FLOP/s per GPU)\n\n2. Training duration: 56 days (directly provided)\n   - Converted to seconds: 56 \u00d7 24 \u00d7 3600 = 4.8384e6 seconds\n\n3. Utilization rate: 40%\n\n4. Final calculation:\n   1.25e14 FLOP/s \u00d7 128 GPUs \u00d7 4.8384e6 seconds \u00d7 0.4 = 3.1e22 FLOPs\n\n\n6 FLOP / parameter / token * 640 *10^6 parameters * 11000 tokens per GPU per batch * 128 GPUs * 620000 updates = 3.3521664e+21 FLOP\n\nsqrt(3.1e22*3.3521664e+21) = 1.0193977e+22 FLOP"},{"id":"epoch:547b4a8160ae16193ea0:parameters","entityId":"epoch:547b4a8160ae16193ea0","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2024-02-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:3f14b945c897c04bf42c:parameters","entityId":"epoch:3f14b945c897c04bf42c","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2024-02-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"30B"},{"id":"epoch:cb8a786c330c385f6d3a:parameters","entityId":"epoch:cb8a786c330c385f6d3a","metric":"parameters","value":34400000000.0,"unit":"parameters","observationPeriod":"2024-02-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:7f4423d89208b66b45c3:parameters","entityId":"epoch:7f4423d89208b66b45c3","metric":"parameters","value":32000000000.0,"unit":"parameters","observationPeriod":"2024-02-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32B"},{"id":"epoch:ba57690b8a54ddd2bd60:parameters","entityId":"epoch:ba57690b8a54ddd2bd60","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-02-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our model is initialized with DeepSeek-Coder-Base-v1.5 7B (Guo et al., 2024) and trained for 500B tokens.\""},{"id":"epoch:ba57690b8a54ddd2bd60:training-compute","entityId":"epoch:ba57690b8a54ddd2bd60","metric":"training-compute","value":1.014e+23,"unit":"FLOP","observationPeriod":"2024-02-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8.04e+22 (base model) + 2.1e+22 (fine-tune) = 1.014e+23"},{"id":"epoch:2bf6310ba51ccab63c4a:parameters","entityId":"epoch:2bf6310ba51ccab63c4a","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2024-02-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"72B"},{"id":"epoch:2bf6310ba51ccab63c4a:training-compute","entityId":"epoch:2bf6310ba51ccab63c4a","metric":"training-compute","value":1.3e+24,"unit":"FLOP","observationPeriod":"2024-02-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3T training tokens: https://github.com/QwenLM/Qwen2/issues/97 \n\n6 * 72 billion * 3 trillion = ~1.3e24"},{"id":"epoch:eacf21ea341ec97e0ad0:parameters","entityId":"epoch:eacf21ea341ec97e0ad0","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-02-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:eacf21ea341ec97e0ad0:training-compute","entityId":"epoch:eacf21ea341ec97e0ad0","metric":"training-compute","value":1.68e+23,"unit":"FLOP","observationPeriod":"2024-02-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 7*10^9 parameters * 4*10^12 tokens =  1.68e+23 FLOP"},{"id":"epoch:8f8d7f8b36998576dce8:parameters","entityId":"epoch:8f8d7f8b36998576dce8","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2024-02-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B"},{"id":"epoch:8f8d7f8b36998576dce8:training-compute","entityId":"epoch:8f8d7f8b36998576dce8","metric":"training-compute","value":3.36e+23,"unit":"FLOP","observationPeriod":"2024-02-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 14*10^9 parameters * 4*10^12 tokens =  3.36e+23 FLOP"},{"id":"epoch:2d6f8146895ed06cc80d:parameters","entityId":"epoch:2d6f8146895ed06cc80d","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-02-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2d6f8146895ed06cc80d:training-compute","entityId":"epoch:2d6f8146895ed06cc80d","metric":"training-compute","value":1.0332e+23,"unit":"FLOP","observationPeriod":"2024-02-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"direct calculation:\n6*7B*2.46trillion=1.0332 \u00d7 10^23\n\n(calculation also reporoduced by the developers in https://arxiv.org/pdf/2501.00656)"},{"id":"epoch:9cf614f3c8a76d4678d1:parameters","entityId":"epoch:9cf614f3c8a76d4678d1","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2024-02-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9cf614f3c8a76d4678d1:training-compute","entityId":"epoch:9cf614f3c8a76d4678d1","metric":"training-compute","value":1.2e+22,"unit":"FLOP","observationPeriod":"2024-02-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"direct calculation:\n6*1B*2trillion=1.2e22\n\n(calculation also reporoduced by the developers in https://arxiv.org/pdf/2501.00656)"},{"id":"epoch:fd664a5d6b68b48b4fff:parameters","entityId":"epoch:fd664a5d6b68b48b4fff","metric":"parameters","value":34750000000.0,"unit":"parameters","observationPeriod":"2024-01-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"34.75B"},{"id":"epoch:fd664a5d6b68b48b4fff:training-compute","entityId":"epoch:fd664a5d6b68b48b4fff","metric":"training-compute","value":2.5878528e+20,"unit":"FLOP","observationPeriod":"2024-01-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.6e20 = 32 * 312e12 * 0.3 * 24* 3600 = num gpus * peak flops * assumed utilization rate * time in seconds\n\"The largest 34B variant finishes training in ~1 day with 32 A100s.\""},{"id":"epoch:3a80f456b888a697eb51:parameters","entityId":"epoch:3a80f456b888a697eb51","metric":"parameters","value":335000000.0,"unit":"parameters","observationPeriod":"2024-01-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"335M params (from Hugging Face model card)"},{"id":"epoch:07a22fd054e672c8d8b0:parameters","entityId":"epoch:07a22fd054e672c8d8b0","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-01-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:07a22fd054e672c8d8b0:training-compute","entityId":"epoch:07a22fd054e672c8d8b0","metric":"training-compute","value":1.26e+24,"unit":"FLOP","observationPeriod":"2024-01-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Base model saw 2T tokens, Code Llama-70B was trained on an additional 1T. 6NC:\n6 * 3T * 70B = 1.26e24"},{"id":"epoch:8035120ffb523e8fd5f8:parameters","entityId":"epoch:8035120ffb523e8fd5f8","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2024-01-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Baichuan Intelligence releases Baichuan 3, a large-scale model with over 100 billion data points, with Chinese evaluation level exceeding GPT-4\"\n\n*different sources report diffrenetly: \"over 100 billion\" or \"1 trillion\" parameters"},{"id":"epoch:fb513609a60f9f4f2d02:parameters","entityId":"epoch:fb513609a60f9f4f2d02","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2024-01-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:782fff3df82207e1a061:parameters","entityId":"epoch:782fff3df82207e1a061","metric":"parameters","value":1137000000.0,"unit":"parameters","observationPeriod":"2024-01-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:782fff3df82207e1a061:training-compute","entityId":"epoch:782fff3df82207e1a061","metric":"training-compute","value":7.616995200001001e+21,"unit":"FLOP","observationPeriod":"2024-01-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"GPU hour estimate\n4*36000s*9.9e+14*0.4=5.7e+19\n\nBase model: 7.560000000001e+21\nCombined: 7616995200001001000000"},{"id":"epoch:4e059048bb942acfd9c1:parameters","entityId":"epoch:4e059048bb942acfd9c1","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-01-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Not stated. Qwen-VL (less capable, presumably smaller version) is 9.6B\n\nUpd: 7B parameters mentioned here\nhttps://github.com/QwenLM/Qwen-VL#qwen-vl-plus"},{"id":"epoch:2a8b1e38a978bbb52b26:parameters","entityId":"epoch:2a8b1e38a978bbb52b26","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2024-01-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Fuyu-Heavy is the world\u2019s third-most-capable multimodal model, behind only GPT4-V and Gemini Ultra, which are 10-20 times bigger\"\n\nSo possibly around ~100B params, though GPT-4/Gemini params aren't public"},{"id":"epoch:1815dee60460143e6ce1:parameters","entityId":"epoch:1815dee60460143e6ce1","metric":"parameters","value":34000000000.0,"unit":"parameters","observationPeriod":"2024-01-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"34b"},{"id":"epoch:1815dee60460143e6ce1:training-compute","entityId":"epoch:1815dee60460143e6ce1","metric":"training-compute","value":1.85174e+22,"unit":"FLOP","observationPeriod":"2024-01-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"989500000000000*240*3600*128*0.3 = 3.2829235e+22\n\n6*34B*(100*10^6*224\u00d7224/14\u00d714 + 25*10^6*448\u00d7448/14\u00d714) =  1.04448 \u00d7 10^22 \n\nsqrt(3.2829235e+22*1.04448 \u00d7 10^22 ) = 1.85174... \u00d7 10^22"},{"id":"epoch:6dc92ebadbd6c2ad8f07:parameters","entityId":"epoch:6dc92ebadbd6c2ad8f07","metric":"parameters","value":4300000.0,"unit":"parameters","observationPeriod":"2024-01-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"This model employs a BPE tokenizer with\na vocabulary size of 10k, which consequentially expands the\nmodel parameters to 4.3 million\""},{"id":"epoch:417b2946dc8b899b0e88:parameters","entityId":"epoch:417b2946dc8b899b0e88","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2024-01-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:139e17426651600c637b:parameters","entityId":"epoch:139e17426651600c637b","metric":"parameters","value":1644417024.0,"unit":"parameters","observationPeriod":"2024-01-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table under Model Architecture gives exact parameter count"},{"id":"epoch:139e17426651600c637b:training-compute","entityId":"epoch:139e17426651600c637b","metric":"training-compute","value":1.92e+22,"unit":"FLOP","observationPeriod":"2024-01-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 * 1.6B * 2T = 19200000000000000000000\n"},{"id":"epoch:724db5e88df0ecf53c96:parameters","entityId":"epoch:724db5e88df0ecf53c96","metric":"parameters","value":151000000.0,"unit":"parameters","observationPeriod":"2024-01-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Overall, the transformer has 151 million parameters, excluding embedding layers at its input and output heads.\""},{"id":"epoch:60e127f729cacecd254c:parameters","entityId":"epoch:60e127f729cacecd254c","metric":"parameters","value":540000000000.0,"unit":"parameters","observationPeriod":"2024-01-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:30b4d019b8b0ec184ddd:parameters","entityId":"epoch:30b4d019b8b0ec184ddd","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-01-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:3b7613339ecc8ff81a9d:parameters","entityId":"epoch:3b7613339ecc8ff81a9d","metric":"parameters","value":6000000000.0,"unit":"parameters","observationPeriod":"2024-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6B"},{"id":"epoch:3b7613339ecc8ff81a9d:training-compute","entityId":"epoch:3b7613339ecc8ff81a9d","metric":"training-compute","value":7.56e+21,"unit":"FLOP","observationPeriod":"2024-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Python, Java, Javascript, Rust, C++, C, and C# subset of Starcoder Training Dataset\n\nFrom the Starcoder paper, that's 7.9% + 11.3% + 8.4% + 1.2% + 6.4% + 7% + 5.8% of the total, so 48% of 815 GB, say 410 GB of code. If we assume they trained for 2 epochs, 820 GB.\n\n820 GB * 200M word per GB = 1.6e11 words\n1.6e11 / 0.75 = 2.1e11 tokens\n\nC = 6ND = 6 * 2.1e11 * 6e9 = 7.6e21"},{"id":"epoch:bd60872cf0586c1a74e1:parameters","entityId":"epoch:bd60872cf0586c1a74e1","metric":"parameters","value":1700000000.0,"unit":"parameters","observationPeriod":"2024-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:bd60872cf0586c1a74e1:training-compute","entityId":"epoch:bd60872cf0586c1a74e1","metric":"training-compute","value":2.51092992e+21,"unit":"FLOP","observationPeriod":"2024-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"200k steps, 2048batch, 601 sequence length\n\n200k*2048*601=246169600000\n\nCompute: 6*1.7e+9*2.5e+11=2.5e+21\n\nEpoch: 246169600000/1076462666666=0.22"},{"id":"epoch:e1a13c3df57cb5647b3c:parameters","entityId":"epoch:e1a13c3df57cb5647b3c","metric":"parameters","value":6000000000.0,"unit":"parameters","observationPeriod":"2024-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6B"},{"id":"epoch:14fbd426a1f6f58c3a96:parameters","entityId":"epoch:14fbd426a1f6f58c3a96","metric":"parameters","value":700000000.0,"unit":"parameters","observationPeriod":"2024-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"0.7B"},{"id":"epoch:80351e9110f8d33811aa:parameters","entityId":"epoch:80351e9110f8d33811aa","metric":"parameters","value":45000000000.0,"unit":"parameters","observationPeriod":"2024-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"45B"},{"id":"epoch:7d91b56cbc0eaf56ba20:parameters","entityId":"epoch:7d91b56cbc0eaf56ba20","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2024-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B\n"},{"id":"epoch:7d91b56cbc0eaf56ba20:training-compute","entityId":"epoch:7d91b56cbc0eaf56ba20","metric":"training-compute","value":1.744956e+23,"unit":"FLOP","observationPeriod":"2024-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"trainable / total parameters \nstage 1: 13B / 13B \nstage 2: 1B / 14B\n\ntraining tokens: \nstage 1: (28.7-0.5)*0.5*(196/16)^2 + 0.5*(224/16)^2  = 2213B\nstage 2: 1.6*(224/16)^2 = 313.6 B\n\n6*13B*2213B + 6*1B*313.6 B = 174495.6 *10^18 = 1.744956 \u00d7 1023"},{"id":"epoch:006d5d33a95ade7a4de9:parameters","entityId":"epoch:006d5d33a95ade7a4de9","metric":"parameters","value":20000000000.0,"unit":"parameters","observationPeriod":"2024-01-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"20B"},{"id":"epoch:006d5d33a95ade7a4de9:training-compute","entityId":"epoch:006d5d33a95ade7a4de9","metric":"training-compute","value":3.12e+23,"unit":"FLOP","observationPeriod":"2024-01-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 * 2600000000000 * 20000000000 = 3.12e+23"},{"id":"epoch:a9b1f5f0b34545d7c47a:parameters","entityId":"epoch:a9b1f5f0b34545d7c47a","metric":"parameters","value":16000000000.0,"unit":"parameters","observationPeriod":"2024-01-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"16B (total, but it's sparse)"},{"id":"epoch:a9b1f5f0b34545d7c47a:training-compute","entityId":"epoch:a9b1f5f0b34545d7c47a","metric":"training-compute","value":3.4e+22,"unit":"FLOP","observationPeriod":"2024-01-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"With the DeepSeekMoE architecture, we scale up our MoE model to a larger scale with 16B total parameters and train it on 2T tokens\"\n\n\"Evaluation results reveal that with only about 40% of computations, DeepSeekMoE 16B achieves comparable performance with DeepSeek 7B (DeepSeek-AI, 2024), a dense model trained on the same 2T corpus\"\n\n40% * 7B = 2.8B, so 2.8B effective parameters (see also Table 4)\n\n2.8B * 2T * 6 ~= 3.4e22\n\nTable 4 shows 74.4T FLOPs per 4k tokens. Over 2T tokens, this is 3.72e22 FLOPs (unclear whether this is FLOPs for training or inference)"},{"id":"epoch:0686b1e0cd8349c1080f:parameters","entityId":"epoch:0686b1e0cd8349c1080f","metric":"parameters","value":68200000.0,"unit":"parameters","observationPeriod":"2024-01-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(the biggest model from YOLOv8 family trained on COCO dataset for object detection tasks)"},{"id":"epoch:0eba14f7812284bd18a8:parameters","entityId":"epoch:0eba14f7812284bd18a8","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-01-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Honor releases its self-developed 7 billion parameter device-side platform-level AI big model \"Magic Big Model\""},{"id":"epoch:a6dbc157a2ddf7ace88f:parameters","entityId":"epoch:a6dbc157a2ddf7ace88f","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2024-01-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train non-autoregressive transformer models using 300M (MAGNET-small) and 1.5B\n(MAGNET-large) parameters.\""},{"id":"epoch:4a6f2b6c5796c8da502f:parameters","entityId":"epoch:4a6f2b6c5796c8da502f","metric":"parameters","value":2796431360.0,"unit":"parameters","observationPeriod":"2024-01-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2796431360 from https://huggingface.co/stabilityai/stable-code-3b#model-architecture\n\"stable-code-3b is a 2.7B billion parameter decoder-only language model pre-trained on 1.3 trillion tokens of diverse textual and code datasets. \""},{"id":"epoch:4a6f2b6c5796c8da502f:training-compute","entityId":"epoch:4a6f2b6c5796c8da502f","metric":"training-compute","value":2.106e+22,"unit":"FLOP","observationPeriod":"2024-01-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 2.7e9 * 1.3e12 * 6 = 2,106E+22\n\"stable-code-3b is a 2.7B billion parameter decoder-only language model pre-trained on 1.3 trillion tokens of diverse textual and code datasets. \""},{"id":"epoch:58e3d20f3c4ecaf31bea:parameters","entityId":"epoch:58e3d20f3c4ecaf31bea","metric":"parameters","value":16800000.0,"unit":"parameters","observationPeriod":"2024-01-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:58e3d20f3c4ecaf31bea:training-compute","entityId":"epoch:58e3d20f3c4ecaf31bea","metric":"training-compute","value":1.6000000000000008e+19,"unit":"FLOP","observationPeriod":"2024-01-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware: 2x NVIDIA RTX A6000 (3.87e13 FLOP/s per GPU)\n2. Training duration: 6 days = 518,400 seconds (directly provided)\n3. Utilization: 40%\n4. Calculation: \n   2 GPUs \u00d7 3.87e13 FLOP/s \u00d7 518,400s \u00d7 0.40 = 1.6e19 FLOPs"},{"id":"epoch:059ad8901f5703cf5226:parameters","entityId":"epoch:059ad8901f5703cf5226","metric":"parameters","value":67000000000.0,"unit":"parameters","observationPeriod":"2024-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"67B\n"},{"id":"epoch:059ad8901f5703cf5226:training-compute","entityId":"epoch:059ad8901f5703cf5226","metric":"training-compute","value":8.04e+23,"unit":"FLOP","observationPeriod":"2024-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"67B parameters * 2T tokens * 6 FLOP / token / parameter = 8.04e23 FLOP"},{"id":"epoch:7e8605f049f4cbeb1486:parameters","entityId":"epoch:7e8605f049f4cbeb1486","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2024-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.3b"},{"id":"epoch:7e8605f049f4cbeb1486:training-compute","entityId":"epoch:7e8605f049f4cbeb1486","metric":"training-compute","value":3.9e+21,"unit":"FLOP","observationPeriod":"2024-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 1.3*10^9 parameters * 500*10^9 tokens = 3.9\u00d710^21 FLOP"},{"id":"epoch:ea4524f43e1a68a3e7b4:parameters","entityId":"epoch:ea4524f43e1a68a3e7b4","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2024-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:ea4524f43e1a68a3e7b4:training-compute","entityId":"epoch:ea4524f43e1a68a3e7b4","metric":"training-compute","value":8.4e+22,"unit":"FLOP","observationPeriod":"2024-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B parameters * 2T tokens * 6 FLOP / parameter / token = 84*10^21 FLOP"},{"id":"epoch:76c7351a8a9b12263d50:parameters","entityId":"epoch:76c7351a8a9b12263d50","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2024-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"over 100B parameters"},{"id":"epoch:1f66ed8b32ca1c183292:parameters","entityId":"epoch:1f66ed8b32ca1c183292","metric":"parameters","value":35000000.0,"unit":"parameters","observationPeriod":"2024-01-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from RT-1"},{"id":"epoch:a5d842fc1158d4984bba:parameters","entityId":"epoch:a5d842fc1158d4984bba","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2024-01-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a5d842fc1158d4984bba:training-compute","entityId":"epoch:a5d842fc1158d4984bba","metric":"training-compute","value":1.60209723904e+23,"unit":"FLOP","observationPeriod":"2024-01-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"57 hours pretraining on 8 A100, 2.7 hours finteuning on 4 A100\n57 hours * 3.12E14 * 0.4 * 8 + 2.7 hours * 3.12E14 * 0.4 * 4 = 209723904000000000000\n\nBase model: 1.6e+23\nTotal: 160209723904000000000000"},{"id":"epoch:ed160b038850305e04d0:parameters","entityId":"epoch:ed160b038850305e04d0","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2024-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b25ae12190a3539642e2:parameters","entityId":"epoch:b25ae12190a3539642e2","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2023-12-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B"},{"id":"epoch:14d719a2e29eed7ea0c0:parameters","entityId":"epoch:14d719a2e29eed7ea0c0","metric":"parameters","value":12400000000.0,"unit":"parameters","observationPeriod":"2023-12-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Since the default setting consists of two GPT-J (6B) and a DeBERTa-large (0.4B), we note our backbone as \u201cGPT-J 12B\u201d, which implies around 12.4 billion parameters in total. \""},{"id":"epoch:794d5b04c57a86a28240:parameters","entityId":"epoch:794d5b04c57a86a28240","metric":"parameters","value":6700000000.0,"unit":"parameters","observationPeriod":"2023-12-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:bf865ae7db728e0c0e96:parameters","entityId":"epoch:bf865ae7db728e0c0e96","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-12-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:4111fd19615c8458eef3:parameters","entityId":"epoch:4111fd19615c8458eef3","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-12-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:03b5232e9316c6d6799d:parameters","entityId":"epoch:03b5232e9316c6d6799d","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-12-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1c1a0d8e1bb1063d8279:parameters","entityId":"epoch:1c1a0d8e1bb1063d8279","metric":"parameters","value":34000000000.0,"unit":"parameters","observationPeriod":"2023-12-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"34B"},{"id":"epoch:421bb0eb1d1d7b3eca24:parameters","entityId":"epoch:421bb0eb1d1d7b3eca24","metric":"parameters","value":10700000000.0,"unit":"parameters","observationPeriod":"2023-12-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"10.7B"},{"id":"epoch:70129f96dc1e27d47ae1:parameters","entityId":"epoch:70129f96dc1e27d47ae1","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2023-12-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"same as base model"},{"id":"epoch:70129f96dc1e27d47ae1:training-compute","entityId":"epoch:70129f96dc1e27d47ae1","metric":"training-compute","value":3.4912896e+22,"unit":"FLOP","observationPeriod":"2023-12-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.3e+22 FLOP [base model] + 1.912896e+21 FLOP = 3.4912896e+22 FLOP"},{"id":"epoch:aeb6016a710629598e21:parameters","entityId":"epoch:aeb6016a710629598e21","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2023-12-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:aeb6016a710629598e21:training-compute","entityId":"epoch:aeb6016a710629598e21","metric":"training-compute","value":4.77e+23,"unit":"FLOP","observationPeriod":"2023-12-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1000 A800 GPUs"},{"id":"epoch:1ddfd8b4098b634dcc42:parameters","entityId":"epoch:1ddfd8b4098b634dcc42","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2023-12-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Biggest model has 8B parameters; there are also experiments with 300M and 1B models."},{"id":"epoch:7e2debc8a10324b85e27:parameters","entityId":"epoch:7e2debc8a10324b85e27","metric":"parameters","value":14200000000.0,"unit":"parameters","observationPeriod":"2023-12-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Source: https://huggingface.co/rinna/nekomata-14b"},{"id":"epoch:7e2debc8a10324b85e27:training-compute","entityId":"epoch:7e2debc8a10324b85e27","metric":"training-compute","value":2.5562e+23,"unit":"FLOP","observationPeriod":"2023-12-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Begins from Qwen-14B (2.5e23 FLOP). They continue pretraining on a mix of Japanese and English text for 66B tokens.\n(assuming 1 epoch, and using the C=6ND approximation)\n= # of active parameters / forward pass * # of tokens * 6 FLOPs / token\n~= 14.2e9 active parameters * 66e9 tokens * 6 FLOPs / token\n~= 5623.2e18 FLOPs\n~= 5.62e21 FLOPs\n\nIn total, 2.5562e23"},{"id":"epoch:76b4b8e0c89524198761:parameters","entityId":"epoch:76b4b8e0c89524198761","metric":"parameters","value":3250000000.0,"unit":"parameters","observationPeriod":"2023-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.25B"},{"id":"epoch:0996e407e6d9e17c69e4:parameters","entityId":"epoch:0996e407e6d9e17c69e4","metric":"parameters","value":1800000000.0,"unit":"parameters","observationPeriod":"2023-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.8B"},{"id":"epoch:bb3d65fe95ce1e8818cf:parameters","entityId":"epoch:bb3d65fe95ce1e8818cf","metric":"parameters","value":3000000.0,"unit":"parameters","observationPeriod":"2023-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The model boasts over 3M parameters.\""},{"id":"epoch:3628bbfe658c311359bb:parameters","entityId":"epoch:3628bbfe658c311359bb","metric":"parameters","value":7800000.0,"unit":"parameters","observationPeriod":"2023-12-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"EfficientNet-B1 architecture (7.8M parameters) - I assume this model has the same amount\nembedding size: 1280"},{"id":"epoch:d991138e7b937c792e7f:parameters","entityId":"epoch:d991138e7b937c792e7f","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2023-12-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"10B"},{"id":"epoch:0217bce51ea9131953e4:parameters","entityId":"epoch:0217bce51ea9131953e4","metric":"parameters","value":41000000000.0,"unit":"parameters","observationPeriod":"2023-12-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0217bce51ea9131953e4:training-compute","entityId":"epoch:0217bce51ea9131953e4","metric":"training-compute","value":1.722e+23,"unit":"FLOP","observationPeriod":"2023-12-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"=41000000000 parameters * 700000000000 tokens [see dataset size notes]  * 6 FLOP / token / parameter =1.722 \u00d7 10^23 FLOP"},{"id":"epoch:1bea31a54d1d62dfec21:parameters","entityId":"epoch:1bea31a54d1d62dfec21","metric":"parameters","value":34200000000.0,"unit":"parameters","observationPeriod":"2023-12-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/LumiOpen/Poro-34B"},{"id":"epoch:1bea31a54d1d62dfec21:training-compute","entityId":"epoch:1bea31a54d1d62dfec21","metric":"training-compute","value":2.052e+23,"unit":"FLOP","observationPeriod":"2023-12-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*1T*34.2B= 2.04e+23\n\n\"This allowed total training cycle throughput of 49618 TFLOPs and 174378 tokens/second.\"\n\nthe training took around 18 months (https://hplt-project.org/deliverables)\n49618*18*30*24*3600*10^12=2.3149774e+24\n\n"},{"id":"epoch:73129301d76282ccf14c:parameters","entityId":"epoch:73129301d76282ccf14c","metric":"parameters","value":18000000000.0,"unit":"parameters","observationPeriod":"2023-12-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:73129301d76282ccf14c:training-compute","entityId":"epoch:73129301d76282ccf14c","metric":"training-compute","value":6.707e+22,"unit":"FLOP","observationPeriod":"2023-12-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"States 12.6 TFLOP per 1120x1120 image forward pass. Trained 60k steps with 4608 batch size, and then 10k with 1024 batch size.\n12.6 TFLOP * (60000*4608 + 10000*1024) = 3.76e21\n\nUses pretrained CogVLM as base (6.331e22 FLOP), along with EVA2-CLIP-L. EVA2-CLIP-L's FLOPs are potentially estimable, but based on details about EVA2-CLIP-g/14 (a larger model), they likely contribute negligibly to CogAgent.\n\nSum: 6.707e22"},{"id":"epoch:4404f100f1e3249dd19f:parameters","entityId":"epoch:4404f100f1e3249dd19f","metric":"parameters","value":15000000000.0,"unit":"parameters","observationPeriod":"2023-12-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From the section called \"Pretrained LLM\": \"We use Codey, an LLM built on top of the PaLM2 model family... Because FunSearch relies on sampling from an LLM extensively, an important performance-defining tradeoff is between the quality of the samples and the inference speed of the LLM. In practice, we have chosen to work with a fast-inference model (rather than slower-inference, higher-quality)\"\n\nUnclear which PaLM2 model was used (of Gecko, Otter, Bison, and Unicorn); above quote indicates it was perhaps Otter or Bison, but not Unicorn. Exact parameter counts are not publicly disclosed for any of these models. In comparisons where FunSearch uses StarCoder-15B, Codey is an improvement but not obviously of an entirely different model class.\n\nI report the 15B parameters from StarCoder-15B, used as an open-source comparison"},{"id":"epoch:4404f100f1e3249dd19f:training-compute","entityId":"epoch:4404f100f1e3249dd19f","metric":"training-compute","value":3.87e+23,"unit":"FLOP","observationPeriod":"2023-12-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Appendix A.5: \"Finding the full-sized symmetric admissible set I(15, 10) required the generation and analysis of approximately two million programs... To reproduce admissible set experiments done above (generating 2 million samples) one would have to use 15 instances of StarCoder-15B running on A100 40 GB GPU each and 5 CPU servers (each running 32 evaluators in parallel) for two days. We estimate that when running on Google Cloud, the price of an experiment is around $800 \u2013 $1400, and the energy usage around 250 \u2013 500 kWh; i.e., 0.5% of the energy used for training StarCoder\"\n\n15 GPUs * 7.80E+13 FLOP/GPU-sec * 2 days * 24 hours/day * 3600 sec/hour = 2.02e20 FLOP for the GPU servers\n\nWe should also add the compute used to train the PaLM2 variant used as the base LLM. Since we don't have any details about this model, I use the compute from StarCoder-15B (used as the open source comparison point): 3.87e+23 FLOP\n\nUnclear how to evaluate the compute from the CPU servers implementing the evolutionary algorithm, but this is very likely dwarfed by the pre-training compute for the LLM."},{"id":"epoch:b3d1a0cf0f54f77ff4cb:parameters","entityId":"epoch:b3d1a0cf0f54f77ff4cb","metric":"parameters","value":29000000000.0,"unit":"parameters","observationPeriod":"2023-12-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"GigaChat Pro is a transformer containing 29 billion parameters\""},{"id":"epoch:359b5085fc52b5200b9b:parameters","entityId":"epoch:359b5085fc52b5200b9b","metric":"parameters","value":2700000000.0,"unit":"parameters","observationPeriod":"2023-12-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.7B"},{"id":"epoch:359b5085fc52b5200b9b:training-compute","entityId":"epoch:359b5085fc52b5200b9b","metric":"training-compute","value":2.27e+22,"unit":"FLOP","observationPeriod":"2023-12-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.7B params, trained on 1.4T tokens\n\n2.7 billion * 1.4 trillion * 6 = 2.27e22\n\n96*14 A100-days\n\n14 * 96 * 312 trillion * 24 * 3600 * 0.3 = 1.09e22"},{"id":"epoch:26c48a78261dca3983d2:parameters","entityId":"epoch:26c48a78261dca3983d2","metric":"parameters","value":13350839296.0,"unit":"parameters","observationPeriod":"2023-12-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/Efficient-Large-Model/VILA-13b/tree/main?show_file_info=model.safetensors.index.json\n"},{"id":"epoch:26c48a78261dca3983d2:training-compute","entityId":"epoch:26c48a78261dca3983d2","metric":"training-compute","value":2.3003136e+21,"unit":"FLOP","observationPeriod":"2023-12-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Appendix B: \"We perform training on 16 A100 GPU nodes, each node has 8 GPUs. The training hours for each stage of the 7B model are: projector initialization: 4 hours; visual language pre-training: 30 hours; visual instruction-tuning: 6 hours. The training corresponds to a total of 5.1k GPU hours. Most of the computation is spent on the pre-training stage.\"\n16 nodes * 8 GPU/node = 128 GPU\n128 GPU * (4 h + 30 h + 6 h) = 5120 GPU*h\n8 GPU/node -> A100 SMX4\n\nhttps://huggingface.co/Efficient-Large-Model/VILA-13b/tree/main?show_file_info=model.safetensors.index.json\n-> BF16\n\nAssume 80 GB variant (A100 SMX4 80 GB)\npk-BF16: 312000000000000 FLOP/s = 3.12e14 FLOP/s\n\nAssume 0.4 utilization (NVIDIA in-house)\n\n0.4 * 3.12e14 FLOP/s/GPU * (3600 s / 1 h) * 5120 GPU*h = 2.3003136e21 FLOP\n"},{"id":"epoch:312acfa943570979530d:parameters","entityId":"epoch:312acfa943570979530d","metric":"parameters","value":46700000000.0,"unit":"parameters","observationPeriod":"2023-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"46.7B *sparse* params. 12.9B params used on average:\n\n\"Concretely, Mixtral has 46.7B total parameters but only uses 12.9B parameters per token. It, therefore, processes input and generates output at the same speed and for the same cost as a 12.9B model.\""},{"id":"epoch:312acfa943570979530d:training-compute","entityId":"epoch:312acfa943570979530d","metric":"training-compute","value":7.74e+23,"unit":"FLOP","observationPeriod":"2023-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming the model was trained on ~1-10 trillions of tokens (same OOM as the models from the comparison in Figure 1. Llama 2 was trained on 2T tokens) + Mistral Small 3 was trained on 8T of tokens, we can estimate training compute with \"speculative\" confidence:\n\n6 FLOP / token / parameter * 12.9 * 10^9 active parameters * 10*10^12 tokens [speculatively] = 7.74e+23 FLOP"},{"id":"epoch:704feff601e8379cbb35:parameters","entityId":"epoch:704feff601e8379cbb35","metric":"parameters","value":11900000000.0,"unit":"parameters","observationPeriod":"2023-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" 11.9 billion parameters"},{"id":"epoch:704feff601e8379cbb35:training-compute","entityId":"epoch:704feff601e8379cbb35","metric":"training-compute","value":2.014908e+20,"unit":"FLOP","observationPeriod":"2023-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"11900000000 parameters *2822000000 tokens * 6 FLOP / token / parameter  = 2.014908 \u00d7 10^20 FLOP"},{"id":"epoch:182795d2d0be30e1fb97:training-compute","entityId":"epoch:182795d2d0be30e1fb97","metric":"training-compute","value":1.8e+18,"unit":"FLOP","observationPeriod":"2023-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We pre-trained our model on one NVIDIA A100-PCIE-40GB for about 16 h and fine-tuned for about 1 h.\" Assume 40% utilization and FP16 precision."},{"id":"epoch:86fe950af646df44a020:parameters","entityId":"epoch:86fe950af646df44a020","metric":"parameters","value":6700000000.0,"unit":"parameters","observationPeriod":"2023-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6.7B\nWe used the exact same model architecture as LLaMA 7B"},{"id":"epoch:86fe950af646df44a020:training-compute","entityId":"epoch:86fe950af646df44a020","metric":"training-compute","value":5.55564e+22,"unit":"FLOP","observationPeriod":"2023-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*6700000000.00*1382000000000 = 5.55564e+22"},{"id":"epoch:2b58239877343ea1ac4f:parameters","entityId":"epoch:2b58239877343ea1ac4f","metric":"parameters","value":6700000000.0,"unit":"parameters","observationPeriod":"2023-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6.7B\nWe used the exact same model architecture as LLaMA 7B"},{"id":"epoch:2b58239877343ea1ac4f:training-compute","entityId":"epoch:2b58239877343ea1ac4f","metric":"training-compute","value":4.7898069e+22,"unit":"FLOP","observationPeriod":"2023-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*6,7*10^9*1259130000000=5.0617026e+22\n\n312000000000000*600.5*3600*224*0.3 = 4.5325164e+22\n\nSqrt(4.5325164e+22*5.0617026e+22) = 4.7898069e+22"},{"id":"epoch:344e62be2952102a36d6:parameters","entityId":"epoch:344e62be2952102a36d6","metric":"parameters","value":4719000000.0,"unit":"parameters","observationPeriod":"2023-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train our base model at resolution 17 \u00d7 128 \u00d7 128 (3B parameters), and two 2\u00d7 cascaded super-resolution models for 17 \u00d7 128 \u00d7 224 \u2192 17 \u00d7 256 \u00d7 448 (L, 1.3B, p = 2) and 17 \u00d7 256 \u00d7 448 \u2192\n17 \u00d7 512 \u00d7 896 (L, 419M, p = 2) respectively\"\n\n3B + 1.3B + 419M = 4.719B"},{"id":"epoch:388d9ee302472e3a2e54:parameters","entityId":"epoch:388d9ee302472e3a2e54","metric":"parameters","value":65000000000.0,"unit":"parameters","observationPeriod":"2023-12-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Based on the name. Exact count unknown but may be listed on Hugging Face."},{"id":"epoch:388d9ee302472e3a2e54:training-compute","entityId":"epoch:388d9ee302472e3a2e54","metric":"training-compute","value":1.24800000000001e+24,"unit":"FLOP","observationPeriod":"2023-12-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"C = 6ND = 6 * 3.2T tokens * 65B params = 1.248e24 FLOP"},{"id":"epoch:2a0c49ae5e2ab605c2fc:parameters","entityId":"epoch:2a0c49ae5e2ab605c2fc","metric":"parameters","value":2300000000.0,"unit":"parameters","observationPeriod":"2023-12-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.3B"},{"id":"epoch:abb0d0a40fd744723c8b:parameters","entityId":"epoch:abb0d0a40fd744723c8b","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-12-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:abb0d0a40fd744723c8b:training-compute","entityId":"epoch:abb0d0a40fd744723c8b","metric":"training-compute","value":1.6e+23,"unit":"FLOP","observationPeriod":"2023-12-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.7e17 finetune compute, plus Llama 2-13B pretrain compute (1.6e+23)"},{"id":"epoch:38c6e92f48b51b819459:parameters","entityId":"epoch:38c6e92f48b51b819459","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2023-12-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3B"},{"id":"epoch:33f86dda008b131b8570:training-compute","entityId":"epoch:33f86dda008b131b8570","metric":"training-compute","value":5.0000000001e+25,"unit":"FLOP","observationPeriod":"2023-12-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"This number is an estimate based on limited evidence. In particular, we combine information about the performance of Gemini Ultra on various benchmarks compared to other models, and guesstimates about the hardware setup used for training to arrive at our estimate. Our reasoning and calculations are detailed in this Colab notebook.\nhttps://colab.research.google.com/drive/1sfG91UfiYpEYnj_xB5YRy07T5dv-9O_c"},{"id":"epoch:e56cf1d2793ac39fe3da:parameters","entityId":"epoch:e56cf1d2793ac39fe3da","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-12-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B from Table 1"},{"id":"epoch:6db0b9bdad06a87edd96:parameters","entityId":"epoch:6db0b9bdad06a87edd96","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B\n"},{"id":"epoch:3026d97159ca087ddb9d:parameters","entityId":"epoch:3026d97159ca087ddb9d","metric":"parameters","value":5000000000.0,"unit":"parameters","observationPeriod":"2023-12-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"5B\n\n\"We focus on the 5B PaLI-X variant, as more practical for the on-robot deployment than the 55B variant\""},{"id":"epoch:3d6f2ca8dcfb30712b22:parameters","entityId":"epoch:3d6f2ca8dcfb30712b22","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Total parameters: 13B\nTrainable parameters: 28M"},{"id":"epoch:72d44bcbc69ef44ebcea:parameters","entityId":"epoch:72d44bcbc69ef44ebcea","metric":"parameters","value":23400000.0,"unit":"parameters","observationPeriod":"2023-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 4"},{"id":"epoch:fd5a00be1c6989de2f6e:parameters","entityId":"epoch:fd5a00be1c6989de2f6e","metric":"parameters","value":2800000000.0,"unit":"parameters","observationPeriod":"2023-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.8B\n\nhttps://github.com/state-spaces/mamba"},{"id":"epoch:fd5a00be1c6989de2f6e:training-compute","entityId":"epoch:fd5a00be1c6989de2f6e","metric":"training-compute","value":5.400000000000001e+21,"unit":"FLOP","observationPeriod":"2023-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Table 3 shows the performance of Mamba on a range of popular downstream zero-shot evaluation tasks. We compare against the most well-known open source models at these sizes, most importantly Pythia (Biderman et al. 2023) and RWKV (B. Peng et al. 2023) which were trained with the same tokenizer, dataset, and training length (300B tokens) as our models.\"\n\n3b * 300b * 6 = 5.4e21\n\nNote: this is a new architecture so not sure how well 6*params*data works as a heuristic\n\nFigure 4 shows perplexity curves where Mamba is trained up to 2e20 FLOP, but those are for the 125M and 1.3B variants."},{"id":"epoch:fbdfd91207d693500bcc:parameters","entityId":"epoch:fbdfd91207d693500bcc","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:8f43ad7dcd07de997d2c:parameters","entityId":"epoch:8f43ad7dcd07de997d2c","metric":"parameters","value":125000000.0,"unit":"parameters","observationPeriod":"2023-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e902079582a60240feb1:parameters","entityId":"epoch:e902079582a60240feb1","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2023-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e902079582a60240feb1:training-compute","entityId":"epoch:e902079582a60240feb1","metric":"training-compute","value":2.1893426e+22,"unit":"FLOP","observationPeriod":"2023-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 3B parameters * 980B tokens = 1.764 \u00d7 10^22 FLOP\n\n312000000000000 FLOP / GPU / sec * 336 hours * 3600 sec / hour * 240 GPUs * 0.3 [assumed utilization] = 2.7172454e+22 FLOP\n\nsqrt(1.764 \u00d7 10^22 *  2.7172454e+22) = 2.1893426e+22 FLOP"},{"id":"epoch:ee0bb12fafe87be0eac9:parameters","entityId":"epoch:ee0bb12fafe87be0eac9","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ee0bb12fafe87be0eac9:training-compute","entityId":"epoch:ee0bb12fafe87be0eac9","metric":"training-compute","value":4.3297598e+22,"unit":"FLOP","observationPeriod":"2023-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 7B parameters * 980B tokens = 4.116e+22 FLOP\n\n312000000000000 FLOP / GPU / sec * 256 GPUs * 22 days * 24 hours / day * 3600 sec / hour * 0.3 [assumed utilization] = 4.5546209e+22 FLOP\n\nsqrt(4.116e+22*4.5546209e+22) = 4.3297598e+22 FLOP"},{"id":"epoch:c3b52214d09d2def1a32:parameters","entityId":"epoch:c3b52214d09d2def1a32","metric":"parameters","value":7380000000.0,"unit":"parameters","observationPeriod":"2023-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/SeaLLMs/SeaLLM-7B-v2"},{"id":"epoch:a26fd26a21e9c03bd080:parameters","entityId":"epoch:a26fd26a21e9c03bd080","metric":"parameters","value":8540000000.0,"unit":"parameters","observationPeriod":"2023-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/SeaLLMs/SeaLLM-7B-v2.5"},{"id":"epoch:1cf5e7dea6c62ada75d9:parameters","entityId":"epoch:1cf5e7dea6c62ada75d9","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2023-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"100B"},{"id":"epoch:1b443c5f537675182189:parameters","entityId":"epoch:1b443c5f537675182189","metric":"parameters","value":72000000000.0,"unit":"parameters","observationPeriod":"2023-11-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"72B"},{"id":"epoch:1b443c5f537675182189:training-compute","entityId":"epoch:1b443c5f537675182189","metric":"training-compute","value":1.3e+24,"unit":"FLOP","observationPeriod":"2023-11-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"72 billion params, 3 trillion tokens\n72b * 3T * 6 = 1.3e24"},{"id":"epoch:9bac24406ff6b3964f0d:parameters","entityId":"epoch:9bac24406ff6b3964f0d","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-11-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13 billion"},{"id":"epoch:9bac24406ff6b3964f0d:training-compute","entityId":"epoch:9bac24406ff6b3964f0d","metric":"training-compute","value":2.44e+23,"unit":"FLOP","observationPeriod":"2023-11-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Estimate using hardware:\n\n\"Granite.13b.v1 used 256 A100 GPUs for 1056 hours and 120 TFLOPs.\nGranite.13b.v2 was trained on the same infrastructure for an\nadditional 1152 hours with 120 TFLOPS, bringing the total to\n2208 hours\"\n\nSeems like 120 TFLOPS is the output per GPU after utilization, though they don't explicitly explain that part. That's 38% utilization.\n\n256 * 2208 * 3600 * 120 TFLOPS = 2.44e23\n\nUsing 6ND:\n\n\"The second version of the granite.13b models leverages an updated base model trained on 2.5T trillion tokens.\"\n\n\"The granite.13b.v1 base model is trained for 300K iterations,\nwith a batch size of 4M tokens, for a total of 1.25 trillion\n5 tokens. The granite.13b.v2 base model continued pre-training\non top of the granite.13b.v1 checkpoint for an additional 300K\niterations and a total of 2.5 trillion tokens.\"\n\n2.5T * 13B * 6 = 1.95e23"},{"id":"epoch:c6e738a4e46adbdd22b3:parameters","entityId":"epoch:c6e738a4e46adbdd22b3","metric":"parameters","value":6000000000.0,"unit":"parameters","observationPeriod":"2023-11-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6B\n\nhttps://aws.amazon.com/bedrock/cohere-command-embed/"},{"id":"epoch:c6e738a4e46adbdd22b3:training-compute","entityId":"epoch:c6e738a4e46adbdd22b3","metric":"training-compute","value":1.001e+22,"unit":"FLOP","observationPeriod":"2023-11-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://docs.cohere.com/docs/environmental-impact\n\n2700kg CO2 equivalent. Cohere used this calculator: https://mlco2.github.io/impact/ \n\nThis calculator claims that ~40000 TPUv3 hours causes ~3000 kg CO2 emissions in the \"us-west1\", \"us-west2\", and \"us-west3\" regions. Not clear what region the data center Cohere used was in. Google has data centers around the world; *most* regions are similarly carbon intensive as us-west but north-america-northeast is 10x less carbon intensive and south Asia is 5x more carbon intensive. So the calculation below could be quite off.\n\nCohere most likely used TPUv4s, which the calculator does not support, which seem to be much more efficient (2.7x more, according to this https://cloud.google.com/blog/topics/systems/tpu-v4-enables-performance-energy-and-co2e-efficiency-gains)\n\n40000 hours * 123 teraflops * 3600 * 0.3 utilization * 2.7 = 1.4e22\n\n"},{"id":"epoch:d3a8ac293c9b76fee0be:parameters","entityId":"epoch:d3a8ac293c9b76fee0be","metric":"parameters","value":16240000.0,"unit":"parameters","observationPeriod":"2023-11-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The pretrained potential has 16.24 million parameters.\"\nThis refers to the GNoME network, which is a \"Gaussian Network Model of Energy\" for predicting crystal potential energy of new crystals."},{"id":"epoch:605abdfd563356358729:parameters","entityId":"epoch:605abdfd563356358729","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2023-11-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:ea3afea64935e0de3e42:parameters","entityId":"epoch:ea3afea64935e0de3e42","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2023-11-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:cf4b9ba6f73e7323fa70:parameters","entityId":"epoch:cf4b9ba6f73e7323fa70","metric":"parameters","value":102600000000.0,"unit":"parameters","observationPeriod":"2023-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"102.6 billion"},{"id":"epoch:cf4b9ba6f73e7323fa70:training-compute","entityId":"epoch:cf4b9ba6f73e7323fa70","metric":"training-compute","value":1.78e+23,"unit":"FLOP","observationPeriod":"2023-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Trained on 288B tokens\n\n6*102.6b*288b = 1.78e23"},{"id":"epoch:9399bf9e93a6df9c64d5:parameters","entityId":"epoch:9399bf9e93a6df9c64d5","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:9399bf9e93a6df9c64d5:training-compute","entityId":"epoch:9399bf9e93a6df9c64d5","metric":"training-compute","value":8e+19,"unit":"FLOP","observationPeriod":"2023-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"see Figure at https://www.together.ai/blog/stripedhyena-7b"},{"id":"epoch:c1c660fd7803bc24a006:parameters","entityId":"epoch:c1c660fd7803bc24a006","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2023-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:78e126bcdf785863c4c1:parameters","entityId":"epoch:78e126bcdf785863c4c1","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-11-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:e1f967b5a7e8048f9777:training-compute","entityId":"epoch:e1f967b5a7e8048f9777","metric":"training-compute","value":6.7392e+22,"unit":"FLOP","observationPeriod":"2023-11-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000 FLOP / GPU / sec [bf16 assumed] * 200000 GPU-hours * 3600 sec / hour *  0.3 [assumed utilization] = 6.7392e+22 FLOP"},{"id":"epoch:32e1f22230bb9ce8f95d:parameters","entityId":"epoch:32e1f22230bb9ce8f95d","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2023-11-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"I assume the same as Qwen 14B"},{"id":"epoch:3fd98e2eeae24c30b540:parameters","entityId":"epoch:3fd98e2eeae24c30b540","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-11-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:687c553ba453260d287c:parameters","entityId":"epoch:687c553ba453260d287c","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-11-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a60d37abb44d6c436248:training-compute","entityId":"epoch:a60d37abb44d6c436248","metric":"training-compute","value":1.001e+25,"unit":"FLOP","observationPeriod":"2023-11-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Inflection-2 was trained on 5,000 NVIDIA H100 GPUs in fp8 mixed precision for ~10\u00b2\u2075 FLOPs\"\n\n(the second 1 is there because of airtable being wonky, it's not a real sig fig)"},{"id":"epoch:28f37efd0e4339b0a939:parameters","entityId":"epoch:28f37efd0e4339b0a939","metric":"parameters","value":12540000000.0,"unit":"parameters","observationPeriod":"2023-11-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The model uses InternViT-6B-448px V1-2 as the visual encoder, which undergoes pre-training, and GigaChat-7B as the LLM, which undergoes fine-tuning [1]. InternViT-6B-448px V1-2 has 5.54B parameters [2], and GigaChat-7B, being closed-source is assumed to have 7B parameters.\n1. https://arxiv.org/pdf/2404.06212\n2. https://huggingface.co/OpenGVLab/InternViT-6B-448px-V1-2"},{"id":"epoch:4662212122979fd77916:parameters","entityId":"epoch:4662212122979fd77916","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-11-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"based on Llama 13B"},{"id":"epoch:4662212122979fd77916:training-compute","entityId":"epoch:4662212122979fd77916","metric":"training-compute","value":4.6e+22,"unit":"FLOP","observationPeriod":"2023-11-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4.55e22 base compute from Llama-13 + 8.6e20 finetune compute ~= 4.6e22 "},{"id":"epoch:5867ebdec63959a17a34:parameters","entityId":"epoch:5867ebdec63959a17a34","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2023-11-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:f5d3ec563fd7ca8fd952:parameters","entityId":"epoch:f5d3ec563fd7ca8fd952","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-11-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B, based on Mistral 7B"},{"id":"epoch:1e46659c9818df0bc95c:parameters","entityId":"epoch:1e46659c9818df0bc95c","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2023-11-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1e46659c9818df0bc95c:training-compute","entityId":"epoch:1e46659c9818df0bc95c","metric":"training-compute","value":1.8e+23,"unit":"FLOP","observationPeriod":"2023-11-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/nvidia/nemotron-3-8b-base-4k\n\n\"This model was trained on a dataset containing 3.8 Trillion tokens of text\"\n\n8 billion * 3.8 trillion * 6 = 1.8e23\n\nAlso, using the hardware method: \"1,024 A100s were used for 19 days to train the model.\"\n\n19*1024 * 312 trillion * 24 * 3600 * 0.3 = 1.57e23"},{"id":"epoch:21c1f8dc49fc8aef4c6f:parameters","entityId":"epoch:21c1f8dc49fc8aef4c6f","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-11-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:c8b32194e1ed19dd261c:training-compute","entityId":"epoch:c8b32194e1ed19dd261c","metric":"training-compute","value":2.1000000000000002e+22,"unit":"FLOP","observationPeriod":"2023-11-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Training GraphCast took roughly four weeks on 32 Cloud TPU v4 devices using batch parallelism.\"\n\n4.6: \"we use bfloat16 floating point precision\"\n\n2.1e22 = 2.75E+14 FLOP/s * 32 * 60* 60 * 24 * 7 * 4"},{"id":"epoch:14384b1d89b7b1edefda:parameters","entityId":"epoch:14384b1d89b7b1edefda","metric":"parameters","value":8460000000.0,"unit":"parameters","observationPeriod":"2023-11-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"the model has two components - audio and language.\n670M + 7.7B = 8.46B\n\"The audio encoder is composed of 640M parameters\"\n\n\"Qwen-Audio incorporates a large language model as its foundational component.\nThe model is initialized using pre-trained weights derived from Qwen-7B (Bai et al., 2023a). Qwen-7B is a 32-layer Transformer decoder model with a hidden size of 4096, encompassing a total of 7.7B parameters.\""},{"id":"epoch:aa9345a7d8aab9333ddf:parameters","entityId":"epoch:aa9345a7d8aab9333ddf","metric":"parameters","value":19900000000.0,"unit":"parameters","observationPeriod":"2023-11-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"SPHINX + Llama 2 13B\nSPHINX component involves four vision encoders:\n- CLIP - ViT\n- CLIP - ConvNeXt V2 (89M to 659M params, depending on size)\n- DinoV2 - ViT (22M to 1.14B params, depending on size)\n- Q-former (188M params)\nAlso involves to projection networks\n\nHuggingface Hub model files appear to be 39.8GB. Assuming models are stored in fp16 there are 2 bytes per parameter, so 39.8 / 2 = 19.9B parameters."},{"id":"epoch:aa9345a7d8aab9333ddf:training-compute","entityId":"epoch:aa9345a7d8aab9333ddf","metric":"training-compute","value":3.04e+22,"unit":"FLOP","observationPeriod":"2023-11-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The pre-training time is around 125 hours on 32 A100 GPUs with a 7B\nlanguage model and about twice the time with a 13B language model... The fine-tuning takes about 38 hours with 16 A100 GPUs with a 13B\nlanguage model.\"\n\n((125*2 * 32) + (38 * 16)) * 3.12e14 * 3600 * 0.3 = 2.9e21\n\nComponent vision encoders were initialized from pre-trained:\n- CLIP ViT: 1.5e22 FLOPs for L/14@336\n- ConvNeXt V2: 6.8e21 FLOPs for largest\n- DinoV2: 7.42e+21 FLOPs for largest\n- Q-former: 1.2e21 FLOPs for largest\n\n(Based on full parameter count, SPHINX probably uses largest models)\n\nSum: 3.04e22 FLOPs"},{"id":"epoch:2f55728f4f181e6683ed:parameters","entityId":"epoch:2f55728f4f181e6683ed","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-11-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:2f55728f4f181e6683ed:training-compute","entityId":"epoch:2f55728f4f181e6683ed","metric":"training-compute","value":4.56e+22,"unit":"FLOP","observationPeriod":"2023-11-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Base model is LLaVa-1.5 13B, which used 4.55e22 FLOP (mostly coming from Llama base)\n\n\"For this research, we used an NVIDIA A100-SXM4-80GB GPU and an AMD EPYC 7513 32-Core Processor running at 2.0778 GHz. Training\nVOLCANO 7B required 8 GPUs and took a total of 15 hours, while training VOLCANO 13B took 30 hours.\"\n3.12e14 * 8 * 30 * 3600 * 0.3 = 8.1e19 finetune compute"},{"id":"epoch:c6508e3e65dacd262ea6:parameters","entityId":"epoch:c6508e3e65dacd262ea6","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-11-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0a0a3a85c6ac1f349319:parameters","entityId":"epoch:0a0a3a85c6ac1f349319","metric":"parameters","value":1000000000000.0,"unit":"parameters","observationPeriod":"2023-11-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1T"},{"id":"epoch:b9fc9d1447a4de6faa93:training-compute","entityId":"epoch:b9fc9d1447a4de6faa93","metric":"training-compute","value":2.6e+19,"unit":"FLOP","observationPeriod":"2023-11-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"It takes around 2 days on 4 Nvidia V100 with 16 GB to train one of the 4 models.\"\n\n125 tflop/s for V100 SXM (not clear which they used, could be PCI given small number - still same OOM thus confident)\n4 * 125 trillion * 2 * 24 * 3600 * 0.3 = 2.6e19"},{"id":"epoch:f6011b1db0687a9a3087:parameters","entityId":"epoch:f6011b1db0687a9a3087","metric":"parameters","value":100000000.0,"unit":"parameters","observationPeriod":"2023-11-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f6011b1db0687a9a3087:training-compute","entityId":"epoch:f6011b1db0687a9a3087","metric":"training-compute","value":2.299133952e+21,"unit":"FLOP","observationPeriod":"2023-11-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"compute hardware: \n0.3*311.84 (peak TFLOPS)*384 (s per epoch)*64 (GPUs)*1000 (epochs) *(10^12) = 2299133952000000000000 = 2.299133952\u00d710^21"},{"id":"epoch:1c6e4d69e197b1df517f:parameters","entityId":"epoch:1c6e4d69e197b1df517f","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2023-11-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1B (table 4 and hugging  face repo)"},{"id":"epoch:1c6e4d69e197b1df517f:training-compute","entityId":"epoch:1c6e4d69e197b1df517f","metric":"training-compute","value":1.96608e+19,"unit":"FLOP","observationPeriod":"2023-11-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6* 3 276 800 000 *10^9 = 1.96608e+19"},{"id":"epoch:de7f1342c15df1fbe560:parameters","entityId":"epoch:de7f1342c15df1fbe560","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2023-11-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"30B\n\n\"The backbone of Jais-30B is a causal decoder-only large language model. It is engineered with 48 transformer blocks, 56 attention heads, and an embedding dimension of 7168. \""},{"id":"epoch:de7f1342c15df1fbe560:training-compute","entityId":"epoch:de7f1342c15df1fbe560","metric":"training-compute","value":1.0372049e+23,"unit":"FLOP","observationPeriod":"2023-11-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 30 * 10^9 parameters * 427 * 10^9 tokens [ see dataset size notes] = 7.686e+22 FLOP\n\n7500000000000000 FLOP / chip / sec * 16 chips *  1080 hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.39968e+23 FLOP\n\nsqrt(7.686e+22*1.39968e+23) = 1.0372049e+23 FLOP"},{"id":"epoch:bceb9bd684413df52366:parameters","entityId":"epoch:bceb9bd684413df52366","metric":"parameters","value":110000000.0,"unit":"parameters","observationPeriod":"2023-11-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we use bert-base\" \nBERT base has 110M parameters from https://arxiv.org/pdf/1810.04805"},{"id":"epoch:bceb9bd684413df52366:training-compute","entityId":"epoch:bceb9bd684413df52366","metric":"training-compute","value":2.162688e+18,"unit":"FLOP","observationPeriod":"2023-11-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 110000000 parameters * 3276800000 tokens = 2162688000000000000 FLOP"},{"id":"epoch:0bf710f6084bb79e1fbf:parameters","entityId":"epoch:0bf710f6084bb79e1fbf","metric":"parameters","value":7120000000.0,"unit":"parameters","observationPeriod":"2023-11-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"As depicted in Figure 2, our model, referred to as mPLUGOwl2, is composed of three main components: a fundamental vision encoder, a visual abstractor, and a language decoder. Specifically, we utilize ViT-L/14 as the\nvision encoder and LLaMA-2-7B [58] as the language decoder\"\nViT-L/14 has 123M parameters and Llama 2 7B has 7B parameters."},{"id":"epoch:51b56cecd19ac9e601a7:parameters","entityId":"epoch:51b56cecd19ac9e601a7","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2023-11-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B"},{"id":"epoch:f931e0be7f28b9aee39f:parameters","entityId":"epoch:f931e0be7f28b9aee39f","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-11-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:317b21f9cff281456b13:parameters","entityId":"epoch:317b21f9cff281456b13","metric":"parameters","value":1550000000.0,"unit":"parameters","observationPeriod":"2023-11-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:317b21f9cff281456b13:training-compute","entityId":"epoch:317b21f9cff281456b13","metric":"training-compute","value":2.7e+23,"unit":"FLOP","observationPeriod":"2023-11-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Could derive this in terms of Whisper v1, which according to the paper was trained for 680k hours for between 2-3 epochs. Whisper v3 was trained on 5 million hours for 2 epochs, or ~5-7x as much data, and has the same architecture. We have an estimate of 4.65e22 for Whisper 1.\n\nAssume Whisper v1 was trained on 2.5 epochs, or 2.5*680k = 1.7M hours. Whisper v3 was trained on 10M hours. 10/1.7 * 4.65e22 ~= 2.7e23"},{"id":"epoch:b88d584a5665cf157729:parameters","entityId":"epoch:b88d584a5665cf157729","metric":"parameters","value":17000000000.0,"unit":"parameters","observationPeriod":"2023-11-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"CogVLM-17B has 10 billion vision parameters and 7 billion language parameters. However, \"the total number of trainable parameters is 6.5B\".\n\n\"CogVLM model comprises four fundamental components: a vision transformer (ViT) encoder, an MLP adapter, a pretrained large language model (GPT), and a visual expert module.\"\n\nViT: EVA2-CLIP-E, last layer removed (5B params with last layer, non-trainable)\nMLP adapter: 2 layers, parameter count unavailable\nGPT: Vicuna1.5-7B (7B params)\nVisual expert module: parameter count unclear"},{"id":"epoch:b88d584a5665cf157729:training-compute","entityId":"epoch:b88d584a5665cf157729","metric":"training-compute","value":6.331e+22,"unit":"FLOP","observationPeriod":"2023-11-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from table 8 on page 17\n\n230.1 FLOPS*days \nso \n10**15*24*3600*230.1= 1.988e22\n\nSince this training uses pretrained weights from EVA02-CLIP-E and Vicuna1.5-7B, we report the full number of FLOPs baked into the model.\n\nEVA02-CLIP-g/14 is stated to have taken 25 days to train 12B samples using 64 A100-40GB GPUs, implying: \n25 days * 24 hr/day * 3600 sec/hr * 64 GPU * 7.80E+13 FLOP/GPU-sec * 30% efficiency = 3.23e21\n\nEVA02-CLIP-E doesn't give a training time; it saw 1/4 as many samples as the g/14 model but has 4.27x more parameters; as a rough estimate, assume it took the same number of FLOPs to train.\n\nVicuna1.5-7B's training compute is ~entirely embedded in the base Llama-7b weights, which took 4.02e+22 FLOPs\n\nTotal: 1.988e22 + 3.23e21 + 4.02e22 = 6.331E22"},{"id":"epoch:0452cf8c28ad40311ebb:parameters","entityId":"epoch:0452cf8c28ad40311ebb","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-11-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from abstract \"Our final 13B checkpoint uses merely 1.2M publicly available data, and finishes full training in ~1 day on a single 8-A100 node. \""},{"id":"epoch:0452cf8c28ad40311ebb:training-compute","entityId":"epoch:0452cf8c28ad40311ebb","metric":"training-compute","value":7.807e+22,"unit":"FLOP","observationPeriod":"2023-11-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Due to the increased image input resolution to 336^2, the training of LLaVA-1.5 is \u223c2\u00d7 as long as LLaVA: \u223c6 hours of pretraining and \u223c20 hours of visual instruction tuning using 8\u00d7 A100s.\"\n26 * 3600 * 8 * 3.12e14 * 0.3 = 7.0e19\n\nFine-tuned from Vicuna-13B (itself finetuned from LLaMa-13B), which used 7.8e22 FLOPs\n\n7.0e19 + 7.8e22"},{"id":"epoch:c78f2668c8ec706ec290:parameters","entityId":"epoch:c78f2668c8ec706ec290","metric":"parameters","value":314000000000.0,"unit":"parameters","observationPeriod":"2023-11-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"314B parameter Mixture-of-Experts model with 25% of the weights active on a given token\". So effectively 78B parameters\n\nMixture of 8 experts: https://github.com/xai-org/grok-1"},{"id":"epoch:c78f2668c8ec706ec290:training-compute","entityId":"epoch:c78f2668c8ec706ec290","metric":"training-compute","value":2.90000000001e+24,"unit":"FLOP","observationPeriod":"2023-11-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"On these benchmarks, Grok-1 displayed strong results, surpassing all other models in its compute class, including ChatGPT-3.5 and Inflection-1. It is only surpassed by models that were trained with a significantly larger amount of training data and compute resources like GPT-4\"\n\nPer table, Grok-1 is surpassed by Palm 2, Claude 2, GPT-4, so it required less compute than these three models. Palm 2 was trained on 7e24 FLOP.\n\nGPT-3.5 is ~2.6e24. Inflection-1's compute is not public/known by us but Inflection says Inflection-1 compute was <= Palm-540B's (which was ~2.5e24). \n\nFor optimal training, our current working hypothesis is that you still need something like Chinchilla scaling on the total number of parameters in the model, even for MoE models, so optimal dataset size would be 20*310B tokens. With 25%*314B params active per forward pass, this would be around 3e24 FLOP.\nhttps://www.wolframalpha.com/input?i=20*310+billion+*+6+*+25%25+*+314+billion"},{"id":"epoch:f4e7152286924de00404:parameters","entityId":"epoch:f4e7152286924de00404","metric":"parameters","value":33000000000.0,"unit":"parameters","observationPeriod":"2023-11-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"33 billion"},{"id":"epoch:964f4f91658d4b1544b2:parameters","entityId":"epoch:964f4f91658d4b1544b2","metric":"parameters","value":176000000000.0,"unit":"parameters","observationPeriod":"2023-11-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"176 billion"},{"id":"epoch:ead7b1dc11f69f32f684:parameters","entityId":"epoch:ead7b1dc11f69f32f684","metric":"parameters","value":33000000000.0,"unit":"parameters","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"33B"},{"id":"epoch:ead7b1dc11f69f32f684:training-compute","entityId":"epoch:ead7b1dc11f69f32f684","metric":"training-compute","value":3.96e+23,"unit":"FLOP","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Step 1: Initially pre-trained with a dataset consisting of 87% code, 10% code-related language (Github Markdown and StackExchange), and 3% non-code-related Chinese language. Models are pre-trained using 1.8T tokens and a 4K window size in this step.\nStep 2: Further Pre-training using an extended 16K window size on an additional 200B tokens, resulting in foundational models (DeepSeek-Coder-Base).\nStep 3: Instruction Fine-tuning on 2B tokens of instruction data, resulting in instruction-tuned models (DeepSeek-Coder-Instruct).\"\n\nThis means it was trained on 2T tokens. 2T * 33B * 6 = 3.96e23"},{"id":"epoch:2a3898913468339977dc:parameters","entityId":"epoch:2a3898913468339977dc","metric":"parameters","value":34000000000.0,"unit":"parameters","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"34b"},{"id":"epoch:2a3898913468339977dc:training-compute","entityId":"epoch:2a3898913468339977dc","metric":"training-compute","value":6.1e+23,"unit":"FLOP","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The dataset we use contains Chinese & English only. We used approximately 3T tokens\" sounds like this means it was trained on 3T tokens, not necessarily that the dataset contains 3T tokens?\n\nIf so, 34b * 3T * 6 = 6.1e23"},{"id":"epoch:dc6931eded6cfcd6e4ab:training-compute","entityId":"epoch:dc6931eded6cfcd6e4ab","metric":"training-compute","value":1e+19,"unit":"FLOP","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 1x NVIDIA GeForce RTX 3090 (1.60 \u00d7 10\u00b9\u2074 FLOP/s)\n\n2. Training duration: Estimated based on dataset size and batch processing\n   - 71,861,060 total samples (3,593,053 samples \u00d7 20 epochs)\n   - 561,719 batches (batch size 128)\n   - ~15.6 hours (56,172 seconds)\n\n3. Utilization rate: 40%\n\n4. Final calculation:\n   1.60 \u00d7 10\u00b9\u2074 FLOP/s \u00d7 1 GPU \u00d7 5.6 \u00d7 10\u2074 s \u00d7 0.4 = 3.6 \u00d7 10\u00b9\u2078 FLOPs\n   (Rounded to 1.0 \u00d7 10\u00b9\u2079 FLOPs)"},{"id":"epoch:6e98ee6d9e96666867f0:parameters","entityId":"epoch:6e98ee6d9e96666867f0","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6e98ee6d9e96666867f0:training-compute","entityId":"epoch:6e98ee6d9e96666867f0","metric":"training-compute","value":4.2e+23,"unit":"FLOP","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*70B*1000B=4.2e+23"},{"id":"epoch:2bf7bb1d053bea8b4135:parameters","entityId":"epoch:2bf7bb1d053bea8b4135","metric":"parameters","value":130000000000.0,"unit":"parameters","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2bf7bb1d053bea8b4135:training-compute","entityId":"epoch:2bf7bb1d053bea8b4135","metric":"training-compute","value":7.8e+23,"unit":"FLOP","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*130B*1000B=7.8e+23"},{"id":"epoch:e80c738411461f0477d1:parameters","entityId":"epoch:e80c738411461f0477d1","metric":"parameters","value":175000000000.0,"unit":"parameters","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e80c738411461f0477d1:training-compute","entityId":"epoch:e80c738411461f0477d1","metric":"training-compute","value":1.05e+24,"unit":"FLOP","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*175B*1000B=1.05e+24"},{"id":"epoch:4ec6b48b16bd4ce43250:parameters","entityId":"epoch:4ec6b48b16bd4ce43250","metric":"parameters","value":6000000000.0,"unit":"parameters","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6B"},{"id":"epoch:4ec6b48b16bd4ce43250:training-compute","entityId":"epoch:4ec6b48b16bd4ce43250","metric":"training-compute","value":1.26e+23,"unit":"FLOP","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*7*10^9*3*10^12 = 1.26e+23"},{"id":"epoch:28a066afdf7234f4d3c9:parameters","entityId":"epoch:28a066afdf7234f4d3c9","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.3B"},{"id":"epoch:28a066afdf7234f4d3c9:training-compute","entityId":"epoch:28a066afdf7234f4d3c9","metric":"training-compute","value":1.56e+22,"unit":"FLOP","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2T tokens * 1.3B parameters * 6 FLOP / parameter / token = 15.6 * 10^21 = 1.56 * 10^22 FLOP"},{"id":"epoch:601d0dc212208efcad03:parameters","entityId":"epoch:601d0dc212208efcad03","metric":"parameters","value":6700000000.0,"unit":"parameters","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6.7B"},{"id":"epoch:601d0dc212208efcad03:training-compute","entityId":"epoch:601d0dc212208efcad03","metric":"training-compute","value":8.04e+22,"unit":"FLOP","observationPeriod":"2023-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2T tokens * 6.7B parameters * 6 FLOP / parameter / token = 8.04*10^22 FLOP"},{"id":"epoch:723c4fa0f274d8dfbdde:parameters","entityId":"epoch:723c4fa0f274d8dfbdde","metric":"parameters","value":16000000000.0,"unit":"parameters","observationPeriod":"2023-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"16 billion"},{"id":"epoch:723c4fa0f274d8dfbdde:training-compute","entityId":"epoch:723c4fa0f274d8dfbdde","metric":"training-compute","value":2.4e+23,"unit":"FLOP","observationPeriod":"2023-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"It uses 2.5T Tokens for pre-training\". I think that's the number of tokens the model was trained on, not the dataset size, but I'm not sure.\n\n16 billion * 2.5 trillion * 6 = 2.4e23"},{"id":"epoch:f00dde58c1b386a40dc3:parameters","entityId":"epoch:f00dde58c1b386a40dc3","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2023-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9fa545a8bf5a2e1fec41:parameters","entityId":"epoch:9fa545a8bf5a2e1fec41","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:9fa545a8bf5a2e1fec41:training-compute","entityId":"epoch:9fa545a8bf5a2e1fec41","metric":"training-compute","value":5.46e+22,"unit":"FLOP","observationPeriod":"2023-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*7B*1.3T = 54600000000000000000000\n6ND aproximation"},{"id":"epoch:5ef27d8b707791f7dae7:parameters","entityId":"epoch:5ef27d8b707791f7dae7","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:4a3d432694fa66c7c5a5:parameters","entityId":"epoch:4a3d432694fa66c7c5a5","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2023-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"model: Apache 2  https://huggingface.co/OFA-Sys/MuggleMath_7B\ncode: undefined license https://github.com/OFA-Sys/gsm8k-ScRel?tab=readme-ov-file#readme"},{"id":"epoch:e743dd6893c3b731bb0f:parameters","entityId":"epoch:e743dd6893c3b731bb0f","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"same as base model"},{"id":"epoch:e743dd6893c3b731bb0f:training-compute","entityId":"epoch:e743dd6893c3b731bb0f","metric":"training-compute","value":1.92000000001e+20,"unit":"FLOP","observationPeriod":"2023-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming 2457600000 tokens: C = 6 * 2.46B * 13B = 1.92e+20 FLOP"},{"id":"epoch:c9b49d00205e9e633b8c:parameters","entityId":"epoch:c9b49d00205e9e633b8c","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2023-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"same as base model"},{"id":"epoch:c9b49d00205e9e633b8c:training-compute","entityId":"epoch:c9b49d00205e9e633b8c","metric":"training-compute","value":1.03000000001e+21,"unit":"FLOP","observationPeriod":"2023-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming 2457600000 tokens: C = 6 * 2.46B * 70B = 1.03e+21 FLOP"},{"id":"epoch:72a3f5e53293aa4b13ce:parameters","entityId":"epoch:72a3f5e53293aa4b13ce","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-10-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"BlueLM is a large-scale pre-trained language model independently developed by vivo AI Global Research Institute. This release includes 7B base (base) model and 7B conversation (chat) model. At the same time, we have open sourced the long text base (base) model that supports 32K and conversation (chat) model.\" from GitHub https://github.com/vivo-ai-lab/BlueLM\n\n"},{"id":"epoch:72a3f5e53293aa4b13ce:training-compute","entityId":"epoch:72a3f5e53293aa4b13ce","metric":"training-compute","value":1.0920000000001e+23,"unit":"FLOP","observationPeriod":"2023-10-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"C = 6DN = 6 * 2.6T * 7B = 1.092*10^23 FLOP\nhttps://www.wolframalpha.com/input?i=6*7+billion+*+2.6+trillion\n(assuming 1 epoch)\n\nFigure 1 gives compute of 10^12 FLOPs  but this seems improbable\n\nTraining over 2.59T tokens took approximately 26 days using the vivolm system, with a throughput of 3150 tokens/sec/GPU."},{"id":"epoch:84fb2f64eea42598ab1b:parameters","entityId":"epoch:84fb2f64eea42598ab1b","metric":"parameters","value":200000000000.0,"unit":"parameters","observationPeriod":"2023-10-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"200B"},{"id":"epoch:84fb2f64eea42598ab1b:training-compute","entityId":"epoch:84fb2f64eea42598ab1b","metric":"training-compute","value":1.2e+24,"unit":"FLOP","observationPeriod":"2023-10-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND=1000000000000*200000000000.00*6=1.2 \u00d7 10^24"},{"id":"epoch:11f1bc8f6df4d7d55af8:parameters","entityId":"epoch:11f1bc8f6df4d7d55af8","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-10-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:11f1bc8f6df4d7d55af8:training-compute","entityId":"epoch:11f1bc8f6df4d7d55af8","metric":"training-compute","value":2.5e+23,"unit":"FLOP","observationPeriod":"2023-10-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our Skywork-13B is trained on a cluster of 64 NVIDIA-HGX-A800 nodes, a total of 512 A800-80G SXM GPUs... The training process of Skywork-13B spanned a total of 39 days.\"\n\nThey note that \"we achieved a token throughput of 1873 per GPU per second and a model flops utilization (MFU) of 56.5%... \". \n\n\"MFU\" was coined in the Palm paper (https://arxiv.org/pdf/2204.02311.pdf) and only counts operations used to train the model, not all operations observed on the hardware. MFU is lower than traditionally measured utilization.\n\nUsing the 56.5% number, and a peak tensor performance of 623.8 TFLOPS for the A800, this suggests 512 * 623.8 TFLOPS * 39 days * 86400 seconds/day * 0.565 = 6.08e23 FLOP.\n\nBased on C=6ND, with 13B parameters and 3.2T tokens, we have C=6*(13B)*(3.2T)=2.5e23 FLOP.\n\nSince the reported MFU is quite high, and would imply a higher compute usage than 6ND, it seems they may have trained on mixed precision and with the GPUs not always operating in the 623.8 TFLOPS mode."},{"id":"epoch:e061fe1bc0aa2e58bb63:parameters","entityId":"epoch:e061fe1bc0aa2e58bb63","metric":"parameters","value":500000000.0,"unit":"parameters","observationPeriod":"2023-10-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"0.5B\n12-layer encoder + 2-layer decoder, d=512/2048"},{"id":"epoch:e061fe1bc0aa2e58bb63:training-compute","entityId":"epoch:e061fe1bc0aa2e58bb63","metric":"training-compute","value":1.179648e+20,"unit":"FLOP","observationPeriod":"2023-10-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 500000000 parameters * 39321600000 tokens = 117964800000000000000 FLOP"},{"id":"epoch:a76c37eac9a67e7b3302:parameters","entityId":"epoch:a76c37eac9a67e7b3302","metric":"parameters","value":6000000000.0,"unit":"parameters","observationPeriod":"2023-10-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6B from https://arxiv.org/abs/2406.12793"},{"id":"epoch:a76c37eac9a67e7b3302:training-compute","entityId":"epoch:a76c37eac9a67e7b3302","metric":"training-compute","value":5.04e+22,"unit":"FLOP","observationPeriod":"2023-10-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Highly speculative.\nAssume 1 epoch on 1.4T tokens.\n6 FLOP/token/param * 1.4T tokens * 6B params=50.4 * 10 ^(12+9) = 5.04*10^(22)"},{"id":"epoch:4c403f687b44d5a26947:parameters","entityId":"epoch:4c403f687b44d5a26947","metric":"parameters","value":75000000.0,"unit":"parameters","observationPeriod":"2023-10-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:4c403f687b44d5a26947:training-compute","entityId":"epoch:4c403f687b44d5a26947","metric":"training-compute","value":7.92e+18,"unit":"FLOP","observationPeriod":"2023-10-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"V100 performance: 125 teraFLOPS according to https://www.nvidia.com/en-us/data-center/v100/\n\n11 hours * 4 GPUs * 125 teraFLOPS/GPU * 0.40 utilization = 7.92e18 FLOP"},{"id":"epoch:0eb2f94f287a0712dcf1:parameters","entityId":"epoch:0eb2f94f287a0712dcf1","metric":"parameters","value":675000000.0,"unit":"parameters","observationPeriod":"2023-10-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"original parameter count for DiT-XL/2"},{"id":"epoch:dbd123860bd8b0aafe88:parameters","entityId":"epoch:dbd123860bd8b0aafe88","metric":"parameters","value":1571000000000.0,"unit":"parameters","observationPeriod":"2023-10-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Concretely, QMoE can compress the 1.6 trillion parameter SwitchTransformer-c2048 model to less than 160GB\"\nSame parameter count as base model. This paper compresses the model; there is no learning from data."},{"id":"epoch:36887e0cf1311555c525:parameters","entityId":"epoch:36887e0cf1311555c525","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-10-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:012b757021524e936b7a:parameters","entityId":"epoch:012b757021524e936b7a","metric":"parameters","value":13200000000.0,"unit":"parameters","observationPeriod":"2023-10-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13.2B from https://huggingface.co/stockmark/stockmark-13b"},{"id":"epoch:012b757021524e936b7a:training-compute","entityId":"epoch:012b757021524e936b7a","metric":"training-compute","value":1.716e+22,"unit":"FLOP","observationPeriod":"2023-10-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*13B*220B = 17160000000000000000000\n\"Stockmark-13b is a 13 billion parameter LLM pretrained from scratch based on Japanese corpus of about 220B tokens. This model is developed by Stockmark Inc.\"\n"},{"id":"epoch:a4f6dfef84b8cb079f25:parameters","entityId":"epoch:a4f6dfef84b8cb079f25","metric":"parameters","value":14700000000.0,"unit":"parameters","observationPeriod":"2023-10-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:aaa0798b8da5b12d4b85:parameters","entityId":"epoch:aaa0798b8da5b12d4b85","metric":"parameters","value":86000000.0,"unit":"parameters","observationPeriod":"2023-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"SILC models set a new state-of-the-art for these tasks at ViT/B16 model size\" (https://arxiv.org/pdf/2310.13355, page 5). VIT/B16-224 has 86.6M parameters, and VIT/B16-384 has 86.9M parameters (https://huggingface.co/google/vit-base-patch16-224, https://huggingface.co/google/vit-base-patch16-384), so I will assume SILC-S has 86M parameters."},{"id":"epoch:aaa0798b8da5b12d4b85:training-compute","entityId":"epoch:aaa0798b8da5b12d4b85","metric":"training-compute","value":1.004e+22,"unit":"FLOP","observationPeriod":"2023-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"SILC* at ViT B/16 can be trained on 256 TPUv4 chips meanwhile the B/8 and L/16 models require 512 chips. The training takes around 5 days\" (https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/03093-supp.pdf, page 7). Assuming a 33% utilization rate,\nTraining compute\n= utilization rate * # of chips used * peak FLOPS / chip * training time\n= 0.33 * 256 TPUv4 chip * 2.75e14 FLOPS / TPUv4 chip * 5 days * 24 h / day * 3600 s / h\n~= 1.004e22 FLOPS"},{"id":"epoch:685fb820b92c952f81ac:parameters","entityId":"epoch:685fb820b92c952f81ac","metric":"parameters","value":86000000.0,"unit":"parameters","observationPeriod":"2023-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"SILC models set a new state-of-the-art for these tasks at ViT/B16 model size\" (https://arxiv.org/pdf/2310.13355, page 5). VIT/B16-224 has 86.6M parameters, and VIT/B16-384 has 86.9M parameters (https://huggingface.co/google/vit-base-patch16-224, https://huggingface.co/google/vit-base-patch16-384), so I will assume SILC-S has 86M parameters."},{"id":"epoch:23788fae003e76304140:parameters","entityId":"epoch:23788fae003e76304140","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-10-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B - from Table 2"},{"id":"epoch:12a3fd797493c537bd3e:parameters","entityId":"epoch:12a3fd797493c537bd3e","metric":"parameters","value":358000000.0,"unit":"parameters","observationPeriod":"2023-10-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Acoustic Model: 330 million parameters\nDuration Model:\n28M for English-only model (VB-En)"},{"id":"epoch:1b20d0cb44bb69652a08:parameters","entityId":"epoch:1b20d0cb44bb69652a08","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2023-10-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8B\n\nAlso a \"Fuyu-medium\" with unstated param count (<56B: \"Fuyu-Medium performs comparably to PALM-E-562B despite having fewer than a tenth as many parameters\")"},{"id":"epoch:46bbcb7773a1ff7bfba4:parameters","entityId":"epoch:46bbcb7773a1ff7bfba4","metric":"parameters","value":5000000000.0,"unit":"parameters","observationPeriod":"2023-10-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"5B"},{"id":"epoch:f6cd47c1bd8a0371357c:parameters","entityId":"epoch:f6cd47c1bd8a0371357c","metric":"parameters","value":34000000000.0,"unit":"parameters","observationPeriod":"2023-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://arxiv.org/abs/2310.10631"},{"id":"epoch:f6cd47c1bd8a0371357c:training-compute","entityId":"epoch:f6cd47c1bd8a0371357c","metric":"training-compute","value":5.4270979e+23,"unit":"FLOP","observationPeriod":"2023-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"5.3e+23 FLOP [base model compute] + 1.2709785e+22 FLOP [finetune compute] = 5.4270979e+23 FLOP"},{"id":"epoch:04e6c15df845073ee565:parameters","entityId":"epoch:04e6c15df845073ee565","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://arxiv.org/abs/2310.10631"},{"id":"epoch:04e6c15df845073ee565:training-compute","entityId":"epoch:04e6c15df845073ee565","metric":"training-compute","value":1.180685e+23,"unit":"FLOP","observationPeriod":"2023-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.1e+23 FLOP [base model compute] + 8.0684987e+21 FLOP [finetune compute] = 1.180685e+23 FLOP"},{"id":"epoch:61be6c5fcc8baa5c0b64:parameters","entityId":"epoch:61be6c5fcc8baa5c0b64","metric":"parameters","value":34000000000.0,"unit":"parameters","observationPeriod":"2023-10-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"34B\nsafetensors say it is 18.2B params\n\nThere's also a 70B \"experimental\" version: https://github.com/FlagAI-Open/Aquila2"},{"id":"epoch:ab543b4adaed4584d174:parameters","entityId":"epoch:ab543b4adaed4584d174","metric":"parameters","value":175000000000.0,"unit":"parameters","observationPeriod":"2023-10-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ed38601ae8a7ecd39fee:parameters","entityId":"epoch:ed38601ae8a7ecd39fee","metric":"parameters","value":55000000000.0,"unit":"parameters","observationPeriod":"2023-10-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"55B"},{"id":"epoch:9ba93249af34cbfbd654:parameters","entityId":"epoch:9ba93249af34cbfbd654","metric":"parameters","value":13900000000.0,"unit":"parameters","observationPeriod":"2023-10-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"A 13.9B parameter model is mentioned prominently at\nhttps://jiutian.10086.cn/portal/#/home\n2025-01-13."},{"id":"epoch:9ba93249af34cbfbd654:training-compute","entityId":"epoch:9ba93249af34cbfbd654","metric":"training-compute","value":1.668e+23,"unit":"FLOP","observationPeriod":"2023-10-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*13.9e9*2e12=1.668e23"},{"id":"epoch:2ab1fa12dd9310aa60bf:parameters","entityId":"epoch:2ab1fa12dd9310aa60bf","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-10-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:9967be9cdff3d866d4ae:parameters","entityId":"epoch:9967be9cdff3d866d4ae","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-10-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:073606db6a077f56510f:parameters","entityId":"epoch:073606db6a077f56510f","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-10-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:073606db6a077f56510f:training-compute","entityId":"epoch:073606db6a077f56510f","metric":"training-compute","value":3.09e+23,"unit":"FLOP","observationPeriod":"2023-10-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"CodeFuse-13B was trained using 512 Nvidia A100 GPU cards, with\na Hardware FLOPs Utilization (HFU) of approximately 60%. The\ntraining process took approximately 40 days to complete.\" Later they state utilization of 56%\n\n512 * 312 trillion * 40 * 24 * 3600 * 0.56 = 3.09e23\n\nUsing params*tokens, we have 13 billion * 1 trillion * 6 = 7.8e22. might be a sign of multiple epochs? 1T is the size of the dataset; they don't clearly state the number of training tokens"},{"id":"epoch:b4691b32fe11ac4c0a72:training-compute","entityId":"epoch:b4691b32fe11ac4c0a72","metric":"training-compute","value":2.14e+20,"unit":"FLOP","observationPeriod":"2023-10-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Supplementary material: \"This took 8 days on 8 NVIDIA A6000 GPUs.\"\n8 GPUS *3.87E+13 FLOPS/A6000 *60 * 60 * 24 * 8 days"},{"id":"epoch:2c8943a6654e621d47db:parameters","entityId":"epoch:2c8943a6654e621d47db","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-10-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Finetunes using LoRA, so only trains 3.67 million parameters"},{"id":"epoch:2c8943a6654e621d47db:training-compute","entityId":"epoch:2c8943a6654e621d47db","metric":"training-compute","value":1.6e+23,"unit":"FLOP","observationPeriod":"2023-10-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From Llama 2-13B"},{"id":"epoch:56e26914c771822b2dd3:training-compute","entityId":"epoch:56e26914c771822b2dd3","metric":"training-compute","value":1.1000000000000008e+20,"unit":"FLOP","observationPeriod":"2023-10-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware: 4x NVIDIA A100-80GB GPUs (3.12e14 FLOP/s per GPU)\n2. Training duration: 2.5 days = 216,000 seconds (directly provided)\n3. Utilization rate: 40%\n4. Calculation: 3.12e14 FLOP/s \u00d7 4 GPUs \u00d7 216,000s \u00d7 0.40 = 1.1e20 FLOP"},{"id":"epoch:0bfa3b250545c5b90ee7:parameters","entityId":"epoch:0bfa3b250545c5b90ee7","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-10-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"In the name"},{"id":"epoch:0bfa3b250545c5b90ee7:training-compute","entityId":"epoch:0bfa3b250545c5b90ee7","metric":"training-compute","value":2.78e+22,"unit":"FLOP","observationPeriod":"2023-10-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Estimate 1: 1T tokens * 6.7B parameters * 6 FLOP/token/parameter = 4e22 FLOP\n\nfrom paper, Llama-7B took 82,432 GPU hours using A100s\nEstimate 2: 312 trillion FLOP/s * (82,432 * 3600) s * 0.3 = 2.78e22 FLOP"},{"id":"epoch:29d068ded4dca43621fb:parameters","entityId":"epoch:29d068ded4dca43621fb","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2023-10-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.3B\nThe architecture for our 1.3B parameter phi-1 model consists of 24 layers, hidden dimension of 2048, MLP-inner dimension of 8192, and 32 attention heads of dimension 64 each."},{"id":"epoch:29d068ded4dca43621fb:training-compute","entityId":"epoch:29d068ded4dca43621fb","metric":"training-compute","value":3.3234195e+20,"unit":"FLOP","observationPeriod":"2023-10-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 *1.3*10^9 parameters * 51 * 10^9 tokens = 3.978e+20\n\n312000000000000 FLOP/s * 8 GPUs *103 hours *3600 sec / hour *0.3 [assumed utilization] = 2.7765504e+20\n\ngeometric mean sqrt(2.7765504e+20*3.978e+20) = 3.3234195e+20\n"},{"id":"epoch:d5f665b70ea8ac601690:parameters","entityId":"epoch:d5f665b70ea8ac601690","metric":"parameters","value":1100000000.0,"unit":"parameters","observationPeriod":"2023-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.1B"},{"id":"epoch:d5f665b70ea8ac601690:training-compute","entityId":"epoch:d5f665b70ea8ac601690","metric":"training-compute","value":7.24598784e+21,"unit":"FLOP","observationPeriod":"2023-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND approximation: 6*1.1B * 1T = 6600000000000000000000\n\nBased on reported GPU-time and utilization:\nflops = (num gpu) * (peak flops) * (time in seconds) * (reported utilization rate) = (16) * (312 * 10**12) * (30 * 24 * 3600) * (0.56) = 7245987840000001048576 = 7.24598784e+21\n\n\nsource: https://github.com/jzhang38/TinyLlama\n\"Thanks to those optimizations, we achieve a throughput of 24k tokens per second per A100-40G GPU, which translates to 56% model flops utilization\"\nand Releases Schedule from the same link"},{"id":"epoch:5c1ab0ef9517d8577544:parameters","entityId":"epoch:5c1ab0ef9517d8577544","metric":"parameters","value":1100000000.0,"unit":"parameters","observationPeriod":"2023-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.1B"},{"id":"epoch:5c1ab0ef9517d8577544:training-compute","entityId":"epoch:5c1ab0ef9517d8577544","metric":"training-compute","value":2.173796352e+22,"unit":"FLOP","observationPeriod":"2023-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND approximation: 6*1.1B * 3T = 19800000000000000000000\nExtrapolation from the 1T checkpoint:\nflops = (16) * (312 * 10**12) * (3 * 30 * 24 * 3600) * (0.56) = 7245987840000001048576\n(num gpu) * (peak flops) * (time in seconds) * (reported utilization rate)\n\nsource: https://github.com/jzhang38/TinyLlama\n\"Thanks to those optimizations, we achieve a throughput of 24k tokens per second per A100-40G GPU, which translates to 56% model flops utilization\"\nand Releases Schedule from the same link"},{"id":"epoch:2cac335c8795c8590665:parameters","entityId":"epoch:2cac335c8795c8590665","metric":"parameters","value":650000000.0,"unit":"parameters","observationPeriod":"2023-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" BITTERS has 650 million parameters in total."},{"id":"epoch:2cac335c8795c8590665:training-compute","entityId":"epoch:2cac335c8795c8590665","metric":"training-compute","value":7.8015e+17,"unit":"FLOP","observationPeriod":"2023-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"WaveVAE: \n\"We train the model for 10 epochs with batch size 3840.\"\n\"We reduced the number of parameters to 25 million in total\"\n\n25000*100000000*6*10=1.5e+14\n\nBiART:\n\"We train our model for 2 epochs in total with batch size 1280\"\n\"BITTERS has 650 million parameters in total\"\n\n650000000*100000000*6*2=7.8e+17\n\n7.8e+17+1.5e+14=7.8015e+17"},{"id":"epoch:b7c9d8da010380ea28a2:parameters","entityId":"epoch:b7c9d8da010380ea28a2","metric":"parameters","value":600000000.0,"unit":"parameters","observationPeriod":"2023-09-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"0.6B"},{"id":"epoch:b7c9d8da010380ea28a2:training-compute","entityId":"epoch:b7c9d8da010380ea28a2","metric":"training-compute","value":1.541475e+21,"unit":"FLOP","observationPeriod":"2023-09-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"PixArt-\u03b1 only takes 12% of Stable Diffusion v1.5's training time (753 vs. 6,250 A100 GPU days), saving nearly $300,000 ($28,000 vs. $320,000) and reducing 90% CO2 emissions. Moreover, compared with a larger SOTA model, RAPHAEL, our training cost is merely 1%.\nTo ensure fairness, we convert the V100 GPU days (1656) of our training to A100 GPU days (753)\n\nthey compare their compute with Imagen as 7132:753 (see Table 2)\nImagen compute was 1.4600000000000002e+22 FLOPS (from Epoch table) then PIXART-\u03b1 is\n(1.4600000000000002e+22/7132)*753=1.541475e+21 FLOPS (the most likely)\n\nAnother calculation: \n\n753 A100 GPU days (most likely A100 SXM4 80 GB)\n753 days * 24 hours * 3600 s * 77970000000000 FLOPS/s (assume FP16) * 40% utilization rate = 2.0290663296e+21 FLOPS\n753 days * 24 hours * 3600 s * 312000000000000 FLOPS/s (assume FP16 Tensor core) * 40% utilization rate = 8.11938816e+21 FLOPS (unlikely)\n\n1656 V100 GPU days \n1656*24*3600*125000000000000 (assume Tensor float 32)*40%=7.15392e+21 FLOPS (unlikely)\n1656*24*3600*31330000000000 (assume the most popular NVIDIA Tesla V100 DGXS 32 GB FP16)*40% = 1.7930585088e+21 FLOPS\n\nThe count of GPU days excludes the time for data labeling."},{"id":"epoch:349a3a5c4e1f9950c1b1:parameters","entityId":"epoch:349a3a5c4e1f9950c1b1","metric":"parameters","value":6000000000.0,"unit":"parameters","observationPeriod":"2023-09-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6B"},{"id":"epoch:55c5b1d4f24afdb6efc2:parameters","entityId":"epoch:55c5b1d4f24afdb6efc2","metric":"parameters","value":2795443200.0,"unit":"parameters","observationPeriod":"2023-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:55c5b1d4f24afdb6efc2:training-compute","entityId":"epoch:55c5b1d4f24afdb6efc2","metric":"training-compute","value":6.21e+22,"unit":"FLOP","observationPeriod":"2023-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"StableLM-3B-4E1T was trained on the Stability AI cluster across 256 NVIDIA A100 40GB GPUs (AWS P4d instances). Training began on August 23, 2023, and took approximately 30 days to complete.\"\n\n256 * 30 * 24* 3600 * 312 trillion * 0.3 utilization (assumption) = 6.21e22\n\n6ND = 6*2795443200*1000000000000*4 epochs = 6.7090637e+22"},{"id":"epoch:91191d61aea09cd4cfda:parameters","entityId":"epoch:91191d61aea09cd4cfda","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2023-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1B\n"},{"id":"epoch:91191d61aea09cd4cfda:training-compute","entityId":"epoch:91191d61aea09cd4cfda","metric":"training-compute","value":8.2898899e+21,"unit":"FLOP","observationPeriod":"2023-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2: 24,602 GPU Hours A100\n\n312*10^12 FLOP / sec / GPU * 24602 GPU - hours * 3600 sec / hour * 0.3 [assumed utilization] = 8.2898899e+21 FLOP"},{"id":"epoch:50c2a05f91e45eb95aa6:parameters","entityId":"epoch:50c2a05f91e45eb95aa6","metric":"parameters","value":9000000000.0,"unit":"parameters","observationPeriod":"2023-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9B\n\n\"GAIA-1\u2019s world model has 6.5 billion parameters\"\n\n\"GAIA-1\u2019s video decoder has 2.6 billion parameters\""},{"id":"epoch:50c2a05f91e45eb95aa6:training-compute","entityId":"epoch:50c2a05f91e45eb95aa6","metric":"training-compute","value":1.1645338e+22,"unit":"FLOP","observationPeriod":"2023-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000 FLOP / GPU / sec [A100 reported, bf16 assumed] * 34560 GPU-hours [see training time notes] * 3600 sec/ hour * 0.3 [assumed utilization] = 1.1645338e+22 FLOP\n\nthey also report compute of ~1*10^22 via a graph here: https://wayve.ai/thinking/scaling-gaia-1/"},{"id":"epoch:7c9c2da8f4fb08c7192b:parameters","entityId":"epoch:7c9c2da8f4fb08c7192b","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2023-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B"},{"id":"epoch:7c9c2da8f4fb08c7192b:training-compute","entityId":"epoch:7c9c2da8f4fb08c7192b","metric":"training-compute","value":2.5e+23,"unit":"FLOP","observationPeriod":"2023-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3T tokens per Table 1\n\n14B*3T*6 = 2.5e23"},{"id":"epoch:d8acc13e15272b47ac63:parameters","entityId":"epoch:d8acc13e15272b47ac63","metric":"parameters","value":200000000000.0,"unit":"parameters","observationPeriod":"2023-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"200B dense model\nhttps://importai.substack.com/p/import-ai-365-wmd-benchmark-amazon"},{"id":"epoch:d8acc13e15272b47ac63:training-compute","entityId":"epoch:d8acc13e15272b47ac63","metric":"training-compute","value":4.8e+24,"unit":"FLOP","observationPeriod":"2023-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"trained using NVIDIA NeMo: https://blogs.nvidia.com/blog/nemo-amazon-titan/\n\n13,760 NVIDIA A100 chips (using 1,720 P4d nodes). It took 48 days to train.\nfrom https://importai.substack.com/p/import-ai-365-wmd-benchmark-amazon\n\ncounting operations: 6*200000000000*4000000000000=4.8e+24\n\ngpu usage: 312000000000000(FLOP/s)*0.3*13760*1152*3600=5.3413281792e+24"},{"id":"epoch:694c8f1c794cba2b2688:parameters","entityId":"epoch:694c8f1c794cba2b2688","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:694c8f1c794cba2b2688:training-compute","entityId":"epoch:694c8f1c794cba2b2688","metric":"training-compute","value":1.01e+23,"unit":"FLOP","observationPeriod":"2023-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.4T tokens per Table 1\n\n7b*2.4T*6 = 1.01e23"},{"id":"epoch:d08da0efdb277a8f454b:parameters","entityId":"epoch:d08da0efdb277a8f454b","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d08da0efdb277a8f454b:training-compute","entityId":"epoch:d08da0efdb277a8f454b","metric":"training-compute","value":1.17e+23,"unit":"FLOP","observationPeriod":"2023-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*13e9*1.5e12=1.17e+23\nfrom https://huggingface.co/pfnet/plamo-13b#model-details\n\n480 GPUs * 30 days [assumed, likely less] * 24 hours * 3600 s * 77970000000000 FLOP/s * 41.0 [reported utilization] = 3.9772934e+24\n\n\n"},{"id":"epoch:53683b04e5ff5203cf41:parameters","entityId":"epoch:53683b04e5ff5203cf41","metric":"parameters","value":20000000000.0,"unit":"parameters","observationPeriod":"2023-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"20B parameters according to Table 1 in Microsoft's CODEFUSION paper: https://arxiv.org/pdf/2310.17680.pdf"},{"id":"epoch:13f21326e48c7600daec:parameters","entityId":"epoch:13f21326e48c7600daec","metric":"parameters","value":2800000000.0,"unit":"parameters","observationPeriod":"2023-09-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use a large U-Net with 2.8B trainable parameters.\""},{"id":"epoch:79b27c386797be4e5366:parameters","entityId":"epoch:79b27c386797be4e5366","metric":"parameters","value":2000000000.0,"unit":"parameters","observationPeriod":"2023-09-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2B"},{"id":"epoch:79b27c386797be4e5366:training-compute","entityId":"epoch:79b27c386797be4e5366","metric":"training-compute","value":1.2e+22,"unit":"FLOP","observationPeriod":"2023-09-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 2000000000 parameters * 1000000000000 tokens = 1.2e+22 FLOP\n\n\"Likely\" confidence because I amnot sure that \"signals\" correspond directly to gradient updates"},{"id":"epoch:d913cbea189689472df8:parameters","entityId":"epoch:d913cbea189689472df8","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B from https://huggingface.co/internlm/internlm-xcomposer-vl-7b"},{"id":"epoch:d913cbea189689472df8:training-compute","entityId":"epoch:d913cbea189689472df8","metric":"training-compute","value":5.28e+21,"unit":"FLOP","observationPeriod":"2023-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The training procedure employs a batch size of approximately 15.7 million tokens and spans 8,000 iterations\"\n6ND = 6 * 7B params * 15.7M tokens/iteration * 8000 iterations = 5.28e21 FLOP"},{"id":"epoch:f791336f65463cab8666:parameters","entityId":"epoch:f791336f65463cab8666","metric":"parameters","value":93000000.0,"unit":"parameters","observationPeriod":"2023-09-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The model architecture is similar to that of AlphaFold (21), with minor modifications\"\nReference is to the AlphaFold 2 paper; that model had 93 million parameters"},{"id":"epoch:7a69129853ff7902a04d:parameters","entityId":"epoch:7a69129853ff7902a04d","metric":"parameters","value":200000000000.0,"unit":"parameters","observationPeriod":"2023-09-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"PengCheng.Mind \u53c8\u79f0\u9e4f\u57ce\u00b7\u8111\u6d77\u662f\u9e4f\u57ce\u5b9e\u9a8c\u5ba4\u5f00\u53d1\u3001\u5f00\u6e90\u3001\u5f00\u653e\u7684\u57fa\u4e8eTransformer\u67b6\u6784\u7684\u81ea\u56de\u5f52\u5f0f\u8bed\u8a00\u6a21\u578b\uff0c\u76ee\u524d\u5305\u542b7B\u3001200B\u4e24\u4e2a\u7248\u672c\""},{"id":"epoch:ef2ea70ce13bb03047e4:parameters","entityId":"epoch:ef2ea70ce13bb03047e4","metric":"parameters","value":2600000000.0,"unit":"parameters","observationPeriod":"2023-09-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.6B, per paper"},{"id":"epoch:ef2ea70ce13bb03047e4:training-compute","entityId":"epoch:ef2ea70ce13bb03047e4","metric":"training-compute","value":9.8e+21,"unit":"FLOP","observationPeriod":"2023-09-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.6b params * 627b tokens * 6 = 9.8e21"},{"id":"epoch:743ea2cd909e4de01ade:parameters","entityId":"epoch:743ea2cd909e4de01ade","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-09-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:743ea2cd909e4de01ade:training-compute","entityId":"epoch:743ea2cd909e4de01ade","metric":"training-compute","value":7.547904e+20,"unit":"FLOP","observationPeriod":"2023-09-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 13: 128xA800 GPU for (6+10+1.5) hours\n\nflops = (128) * ( 312 * 10**12) * (17.5 * 3600) * (0.3) = 7.5e20\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)\n"},{"id":"epoch:c66ba29fd6a724b0ce1a:parameters","entityId":"epoch:c66ba29fd6a724b0ce1a","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-09-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c66ba29fd6a724b0ce1a:training-compute","entityId":"epoch:c66ba29fd6a724b0ce1a","metric":"training-compute","value":1.092e+23,"unit":"FLOP","observationPeriod":"2023-09-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7b * 2.6t * 6 = 1.092e23\nAlso mentions 1,024 NVIDIA A800 GPUs at 180 TFLOPS per GPU"},{"id":"epoch:d7d79861f37b7fef9fb4:parameters","entityId":"epoch:d7d79861f37b7fef9fb4","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-09-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The openchat-13b is based on the llama-2-13b,\" according to arxiv.org/pdf/2309.11235."},{"id":"epoch:d7d79861f37b7fef9fb4:training-compute","entityId":"epoch:d7d79861f37b7fef9fb4","metric":"training-compute","value":7.805217030000001e+22,"unit":"FLOP","observationPeriod":"2023-09-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The developers only fine-tuned Llama-2-13b, according to arxiv.org/pdf/2309.11235.\n\nFinetune FLOP: 5.23e19\nBase model FLOP: 7.8e+22\nTotal: 78052170300000010000000"},{"id":"epoch:03992adc9a80bff9e945:parameters","entityId":"epoch:03992adc9a80bff9e945","metric":"parameters","value":560000000.0,"unit":"parameters","observationPeriod":"2023-09-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"560M from https://huggingface.co/BAAI/bge-reranker-large\n\"This reranker is initialized from xlm-roberta-base\" from https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/reranker\n"},{"id":"epoch:d863804f279e3b15f75b:parameters","entityId":"epoch:d863804f279e3b15f75b","metric":"parameters","value":5700000000.0,"unit":"parameters","observationPeriod":"2023-09-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"DeciLM 6B is a 5.7 billion parameter decoder-only text generation model. \""},{"id":"epoch:d863804f279e3b15f75b:training-compute","entityId":"epoch:d863804f279e3b15f75b","metric":"training-compute","value":1.026e+22,"unit":"FLOP","observationPeriod":"2023-09-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"assume they used 50% of SlimPajama dataset (300B tokens) - then we are  within 'Likely' confidence interval\n\n6*300B*5700000000=1.026e+22"},{"id":"epoch:435ccab7f220bf06ab02:parameters","entityId":"epoch:435ccab7f220bf06ab02","metric":"parameters","value":500000.0,"unit":"parameters","observationPeriod":"2023-09-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Parkour policy details on page 8, table 11."},{"id":"epoch:70fcb3d0e457ec07c895:parameters","entityId":"epoch:70fcb3d0e457ec07c895","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2023-09-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:70fcb3d0e457ec07c895:training-compute","entityId":"epoch:70fcb3d0e457ec07c895","metric":"training-compute","value":1.17e+21,"unit":"FLOP","observationPeriod":"2023-09-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"150B training tokens\n\n150B*1.3B*6 = 1.17e21\n\nalso, took 1.5k GPU-hours with A100s, per Table 1\n\n1500 * 312 trillion * 3600 * 0.3 (utilization assumption) = 5.05e20\n\nso utilization was likely ~60%. may be high due to the relatively small cluster.\n"},{"id":"epoch:277838ba0ad07aa1e2d2:parameters","entityId":"epoch:277838ba0ad07aa1e2d2","metric":"parameters","value":10700000000.0,"unit":"parameters","observationPeriod":"2023-09-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"10.7B  from appendix A.8"},{"id":"epoch:277838ba0ad07aa1e2d2:training-compute","entityId":"epoch:277838ba0ad07aa1e2d2","metric":"training-compute","value":1.605e+22,"unit":"FLOP","observationPeriod":"2023-09-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 10.7B * 250B = 1.6e22\n'MADLAD-400-10B-MT is a multilingual machine translation model based on the T5 architecture that was trained on 250 billion tokens covering over 450 languages using publicly available data. '\n10.7B  params from appendix A.8"},{"id":"epoch:bd9e4274bcf1c7871112:parameters","entityId":"epoch:bd9e4274bcf1c7871112","metric":"parameters","value":9300000000.0,"unit":"parameters","observationPeriod":"2023-09-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The checkpoint we are releasing has approximately 9.3B parameters. In order to make pipelining during training more efficient, we chose to decouple the input and output embeddings. Doing this does not increase the capacity of the model\u2013it is purely a systems optimization to avoid all-reducing the gradients for the (very large) embeddings across potentially slow communication links. In terms of inference cost, the model is equivalent to an 8B parameter model with coupled input/output embeddings.\""},{"id":"epoch:bd9e4274bcf1c7871112:training-compute","entityId":"epoch:bd9e4274bcf1c7871112","metric":"training-compute","value":4.11246e+22,"unit":"FLOP","observationPeriod":"2023-09-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*9300000000*737000000000=4.11246e+22"},{"id":"epoch:950b6110fbddb076550d:parameters","entityId":"epoch:950b6110fbddb076550d","metric":"parameters","value":101000000000.0,"unit":"parameters","observationPeriod":"2023-09-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:950b6110fbddb076550d:training-compute","entityId":"epoch:950b6110fbddb076550d","metric":"training-compute","value":5.72e+22,"unit":"FLOP","observationPeriod":"2023-09-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"192 GPUs * 160 TFLOP/s per GPU (reported, adjusted for utilization) * 21.54 days * 24 * 3600 = 5.72e22 (confident)\n\n6*101000000000*311540000000=1.8879324e+23 (less confident)"},{"id":"epoch:02ae8023f63cd2d9eaef:parameters","entityId":"epoch:02ae8023f63cd2d9eaef","metric":"parameters","value":6700000000.0,"unit":"parameters","observationPeriod":"2023-09-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"I think this suggests the same number of params as Llama-7b. In any case ~7B. \n\"The model architecture follows LLaMA with exact numerical compatibility to ease adoption in third-party frameworks. The hyperparameters closely follow LLaMA-7B [34] with the following alterations\""},{"id":"epoch:02ae8023f63cd2d9eaef:training-compute","entityId":"epoch:02ae8023f63cd2d9eaef","metric":"training-compute","value":8.02e+22,"unit":"FLOP","observationPeriod":"2023-09-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"270,336 TPUv4-hours per the carbon emissions section. They must mean chip-hours, not core-hours, because they multiply by 192W, which is the max power consumption of a TPU-v4 chip.\n\nhttps://cloud.google.com/tpu/docs/system-architecture-tpu-vm#tpu_v4\n\nThis is probably the total for the whole paper? They released two versions of the base model, one of which is just trained further than the other, and they have two fine-tuned versions. If fine-tuning compute is minor, then this total is close to the largest training compute for any single model in the paper. \n\n270336 * 275 teraflop/s * 3600 * 0.3 = 8.02e22 FLOP\n\nalso, using 6ND:\n\n1484 billion tokens * 6.7 billion * 6 = 5.97e22"},{"id":"epoch:a34ae52f194578cc25c4:parameters","entityId":"epoch:a34ae52f194578cc25c4","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2023-09-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Presently, the Hunyuan model has over 100 billion parameters, with more than two trillion tokens in pre-training data.\""},{"id":"epoch:a34ae52f194578cc25c4:training-compute","entityId":"epoch:a34ae52f194578cc25c4","metric":"training-compute","value":1.2e+24,"unit":"FLOP","observationPeriod":"2023-09-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*100*10^9*2*10^12 = 1.2*10^24"},{"id":"epoch:96b18990adf626435abc:parameters","entityId":"epoch:96b18990adf626435abc","metric":"parameters","value":180000000000.0,"unit":"parameters","observationPeriod":"2023-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Falcon 180B is a super-powerful language model with 180 billion parameters\""},{"id":"epoch:96b18990adf626435abc:training-compute","entityId":"epoch:96b18990adf626435abc","metric":"training-compute","value":3.76e+24,"unit":"FLOP","observationPeriod":"2023-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"43,500 petaflop-days per Table 1 of the paper\n\n43500 * 1e15 * 24 * 3600 = 3.76e24\n\n\nC = 6ND = 6 FLOP/token/parameter * 3.5 trillion tokens * 180 billion parameters = 3.78*10^24 FLOP"},{"id":"epoch:86c1645c25dc27d2cae0:parameters","entityId":"epoch:86c1645c25dc27d2cae0","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:86c1645c25dc27d2cae0:training-compute","entityId":"epoch:86c1645c25dc27d2cae0","metric":"training-compute","value":2.03e+23,"unit":"FLOP","observationPeriod":"2023-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"They describe the dataset as having 2.6T tokens, but the checkpoint graph makes it clear that's also the number of tokens the model was trained on.\n\n13b * 2.6t * 6 = 2.03e23"},{"id":"epoch:c0634123475f64cdb7d3:parameters","entityId":"epoch:c0634123475f64cdb7d3","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2023-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:c0634123475f64cdb7d3:training-compute","entityId":"epoch:c0634123475f64cdb7d3","metric":"training-compute","value":1.02e+24,"unit":"FLOP","observationPeriod":"2023-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"~1.02e24\n\nTigerobo did ~2.1e23 additional pre-training. We estimated Llama 2 was trained on 8.1e23 FLOP."},{"id":"epoch:4a646953d0570a4e9ae5:parameters","entityId":"epoch:4a646953d0570a4e9ae5","metric":"parameters","value":56804.0,"unit":"parameters","observationPeriod":"2023-08-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The control network is an MLP with input dimension 31, two hidden layers of size 128, and an output of dimension 4.\n(31+1)*128+(128+1)*128+(128+1)*4 = 21124\n\nGate detector is a 6 layer U-net with \n8*(3^3*3+1) + 16*(3^2*8+1) + 16*(3^2*16+1) + 16*(5^2*16+1) + 16*(7^2*16+1) + 16*(7^2*16+1) = 35680\n\n35680 + 21124 = 56804"},{"id":"epoch:4a646953d0570a4e9ae5:training-compute","entityId":"epoch:4a646953d0570a4e9ae5","metric":"training-compute","value":5.337e+16,"unit":"FLOP","observationPeriod":"2023-08-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Policies are trained for a total of 1\u2009\u00d7\u2009108 environment interactions, which takes 50\u2009min on a workstation (i9 12900K, RTX 3090, 32\u2009GB RAM DDR5). Fine-tuning is performed for 2\u2009\u00d7\u2009107 environment interactions.\n\n35.58 TFLOPS * 50 min * 60 s/min * 0.50 utilization = 5.337*10^16 FLOP"},{"id":"epoch:258d780bac1b4bb7d32d:parameters","entityId":"epoch:258d780bac1b4bb7d32d","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-08-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"With 13 billion parameters, they demonstrate better knowledge and reasoning capabilities in Arabic\""},{"id":"epoch:258d780bac1b4bb7d32d:training-compute","entityId":"epoch:258d780bac1b4bb7d32d","metric":"training-compute","value":4.8946763e+22,"unit":"FLOP","observationPeriod":"2023-08-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"C = 6ND = 6 * 13 billion params * 395 billion tokens = 3.081e+22 FLOP\n\n7500000000000000 FLOP / chip / sec * 16 chips *  600 hours [see training time notes] * 3600 sec / hour * 0.3 [assumed utilization] = 7.776e+22 FLOP\n\nsqrt(3.081e+22*7.776e+22) = 4.8946763e+22 FLOP"},{"id":"epoch:31a90462839c1744ce4f:parameters","entityId":"epoch:31a90462839c1744ce4f","metric":"parameters","value":1600000000.0,"unit":"parameters","observationPeriod":"2023-08-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.6B"},{"id":"epoch:31a90462839c1744ce4f:training-compute","entityId":"epoch:31a90462839c1744ce4f","metric":"training-compute","value":1.152e+22,"unit":"FLOP","observationPeriod":"2023-08-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 * 1.6B * 1.2T = 11520000000000000000000 = 1.152e22\ncitation \"model trained for 1.2T tokens. \"\nalternative\nflops = (64) * (27770 * 10**9) * (28 * 24 * 3600) * (0.3) = 1.2898787328e+21\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)\nPrecision: bfloat16\nGPUs 64 NVidia A5000\nTraining time 28 days\n27.77 TFLOPS - peak flop from https://www.techpowerup.com/gpu-specs/rtx-a5000.c3748"},{"id":"epoch:d0abd1a677773c8bb5e2:parameters","entityId":"epoch:d0abd1a677773c8bb5e2","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2023-08-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.leiphone.com/category/ai/23kbzQXj60xZgUgO.html \nEnglish translation:\n\"Li Dahai: From a technical point of view, the CPM2 (Chinese Pretrained Model) 100 billion model we launched at that time was a sparse model of MoE, which is different from the 100 billion model we are promoting now.\"\n\nThis suggests it is a dense model."},{"id":"epoch:d0abd1a677773c8bb5e2:training-compute","entityId":"epoch:d0abd1a677773c8bb5e2","metric":"training-compute","value":1.2e+24,"unit":"FLOP","observationPeriod":"2023-08-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assume Chinchilla-optimal dataset size:\n20 * 100B * 100B * 6 = 1.2e24 FLOP"},{"id":"epoch:047e2b73eb86e57ccd15:training-compute","entityId":"epoch:047e2b73eb86e57ccd15","metric":"training-compute","value":4.9e+16,"unit":"FLOP","observationPeriod":"2023-08-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Compute for fine-tuning ProtBERT: 1 NVidia GeForce GTX 1080Ti, 30 epochs, batch size 32, model trained for individual tasks with training time ranging from 58-116 minutes, assuming \nfrom Table 1 we have 244 minutes\n11.34e12 FLOPs and 0.3 utilization rate FLOP = 244 min * 60 sec/min * 11.34e12 FLOP/sec *0.3 = 4.9e16 FLOP,"},{"id":"epoch:d1b125f20d10ee184192:parameters","entityId":"epoch:d1b125f20d10ee184192","metric":"parameters","value":40000000000.0,"unit":"parameters","observationPeriod":"2023-08-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f650b455ba1ebb1b4198:parameters","entityId":"epoch:f650b455ba1ebb1b4198","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-08-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5c1c0b0e1b193678dfd9:parameters","entityId":"epoch:5c1c0b0e1b193678dfd9","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-08-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:466557b55a209c4c6850:parameters","entityId":"epoch:466557b55a209c4c6850","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2023-08-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"This model is an Open-Assistant fine-tuning of Meta's Llama2 70B LLM\" (https://huggingface.co/OpenAssistant/llama2-70b-oasst-sft-v10)."},{"id":"epoch:100431ba1bfe82570826:parameters","entityId":"epoch:100431ba1bfe82570826","metric":"parameters","value":9600000000.0,"unit":"parameters","observationPeriod":"2023-08-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9.6B total - Table 1"},{"id":"epoch:913133193de50b34af2d:parameters","entityId":"epoch:913133193de50b34af2d","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2023-08-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"same as llama 2 70b"},{"id":"epoch:1e9e87500dfa03b54ad4:parameters","entityId":"epoch:1e9e87500dfa03b54ad4","metric":"parameters","value":6700000000.0,"unit":"parameters","observationPeriod":"2023-08-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6.7B"},{"id":"epoch:1e9e87500dfa03b54ad4:training-compute","entityId":"epoch:1e9e87500dfa03b54ad4","metric":"training-compute","value":5.8e+21,"unit":"FLOP","observationPeriod":"2023-08-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8 A100s for three months\n\n8 * 312 trillion * 24 * 3600 * 90 * 0.3 (utilization assumption) = 5.8e21"},{"id":"epoch:a6951021675ad5ba6920:parameters","entityId":"epoch:a6951021675ad5ba6920","metric":"parameters","value":2700000.0,"unit":"parameters","observationPeriod":"2023-08-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a6951021675ad5ba6920:training-compute","entityId":"epoch:a6951021675ad5ba6920","metric":"training-compute","value":2.59999999999998e+19,"unit":"FLOP","observationPeriod":"2023-08-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware: 1x Tesla V100 PCIe 32GB (1.30\u00d710\u00b9\u2074 FLOP/s)\n2. Training duration: directly provided - 140 hours total (80h Shape Encoder + 60h Diffusion Model) = 504,000 seconds\n3. Utilization: 40%\n4. Calculation: 1.30\u00d710\u00b9\u2074 FLOP/s \u00d7 1 GPU \u00d7 504,000s \u00d7 0.4 = 2.60\u00d710\u00b9\u2079 FLOPs"},{"id":"epoch:65665a329a370a122215:parameters","entityId":"epoch:65665a329a370a122215","metric":"parameters","value":80000000000.0,"unit":"parameters","observationPeriod":"2023-08-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"IDEFICS... comes in two variants\u2014the base version and the instructed version. Each variant is available at the 9 billion and 80 billion parameter sizes."},{"id":"epoch:65665a329a370a122215:training-compute","entityId":"epoch:65665a329a370a122215","metric":"training-compute","value":1.1593580544e+23,"unit":"FLOP","observationPeriod":"2023-08-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"flops = 512 * 312e12 * 28*24*3600 * 0.3 = 1.159e23\n(num gpus) * (peak perforemence) * (time in seconds) * (assumed utilization rate)\n\n\"The IDEFICS models were trained on an AWS SageMaker cluster with 8x80GB A100 GPUs nodes and EFA network.\n    IDEFICS-80B took ~28 days of training on 64 nodes (512 GPUs).\"\n\nhttps://huggingface.co/HuggingFaceM4/idefics-80b-instruct \n\ntrained on 150B text tokens + images\n\n6ND = 6*734000000000*80*10^9 = 3.5232e+23\n\n"},{"id":"epoch:e2b42742e3bb6c0715c6:parameters","entityId":"epoch:e2b42742e3bb6c0715c6","metric":"parameters","value":9000000000.0,"unit":"parameters","observationPeriod":"2023-08-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9B"},{"id":"epoch:853b710a254ab3752450:parameters","entityId":"epoch:853b710a254ab3752450","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-08-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:346fc074cd657ba310ee:parameters","entityId":"epoch:346fc074cd657ba310ee","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-08-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:346fc074cd657ba310ee:training-compute","entityId":"epoch:346fc074cd657ba310ee","metric":"training-compute","value":1.248e+23,"unit":"FLOP","observationPeriod":"2023-08-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"=1600000000000 tokens * 6 FLOP / token / parameter * 13000000000 parameters = 1.248\u00d710^23 FLOP"},{"id":"epoch:93bb0b4d926aa77b9927:parameters","entityId":"epoch:93bb0b4d926aa77b9927","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-08-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:93bb0b4d926aa77b9927:training-compute","entityId":"epoch:93bb0b4d926aa77b9927","metric":"training-compute","value":6.72e+22,"unit":"FLOP","observationPeriod":"2023-08-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"=1600000000000*6*7000000000=6.72 \u00d7 10^22"},{"id":"epoch:8db7bd47a6201f216164:parameters","entityId":"epoch:8db7bd47a6201f216164","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-08-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:8db7bd47a6201f216164:training-compute","entityId":"epoch:8db7bd47a6201f216164","metric":"training-compute","value":2.73e+22,"unit":"FLOP","observationPeriod":"2023-08-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"=350000000000*6*13000000000=2.73 \u00d7 10^22"},{"id":"epoch:29d9555c5ea3cefab6d6:parameters","entityId":"epoch:29d9555c5ea3cefab6d6","metric":"parameters","value":1100000000.0,"unit":"parameters","observationPeriod":"2023-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.1B, per model card https://huggingface.co/Deci/DeciCoder-1b"},{"id":"epoch:29d9555c5ea3cefab6d6:training-compute","entityId":"epoch:29d9555c5ea3cefab6d6","metric":"training-compute","value":2.9e+21,"unit":"FLOP","observationPeriod":"2023-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"446b * 1.1b * 6 = 2.9e21"},{"id":"epoch:ec6cf8c451d36585a02b:parameters","entityId":"epoch:ec6cf8c451d36585a02b","metric":"parameters","value":13100000000.0,"unit":"parameters","observationPeriod":"2023-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"'Konan LLM' has 13.1 billion parameters."},{"id":"epoch:ec6cf8c451d36585a02b:training-compute","entityId":"epoch:ec6cf8c451d36585a02b","metric":"training-compute","value":3.86712e+22,"unit":"FLOP","observationPeriod":"2023-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"=13100000000 parameters * 492000000000 tokens [see dataset size notes] * 6 FLOP / token / parameter =3.86712 \u00d7 10^22 FLOP"},{"id":"epoch:60bab75fcff5f1611794:parameters","entityId":"epoch:60bab75fcff5f1611794","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:cc5dd7ed4e59a611728c:parameters","entityId":"epoch:cc5dd7ed4e59a611728c","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2023-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"fine-tuned from LLaMa 2-70B"},{"id":"epoch:3ab2fe721c7e0e53e5a0:parameters","entityId":"epoch:3ab2fe721c7e0e53e5a0","metric":"parameters","value":34000000000.0,"unit":"parameters","observationPeriod":"2023-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"34B"},{"id":"epoch:3ab2fe721c7e0e53e5a0:training-compute","entityId":"epoch:3ab2fe721c7e0e53e5a0","metric":"training-compute","value":5.3e+23,"unit":"FLOP","observationPeriod":"2023-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.22e23 finetune compute, or ~5.3e23 including Llama-2 34B base compute. See finetune compute notes for calculation.\n"},{"id":"epoch:9b6143d7f0b92bc912a6:parameters","entityId":"epoch:9b6143d7f0b92bc912a6","metric":"parameters","value":3600000000.0,"unit":"parameters","observationPeriod":"2023-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.6B"},{"id":"epoch:1fc9c4401e4c35b0decc:parameters","entityId":"epoch:1fc9c4401e4c35b0decc","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:1fc9c4401e4c35b0decc:training-compute","entityId":"epoch:1fc9c4401e4c35b0decc","metric":"training-compute","value":1.1e+23,"unit":"FLOP","observationPeriod":"2023-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.5e22 finetune compute + 8.4e22 base compute for Llama 2-7B, for ~1.1e23 compute overall\n\nTable 26: \"In aggregate, training all 12 Code Llama models required 1400K GPU hours of computation on hardware of type A100-80GB\"\nSuggests all versions took a combined 4.7e23 FLOPs:\n3.12e14 * 1400000 * 3600 * 0.3 = 4.7e23\n\nAssuming this refers to the finetune compute only, agrees with our finetune estimate if compute is proportional to parameter count:\n7 / (7+13+34+70) = 0.056\n0.056 * 4.7e23 = 2.65e22\n"},{"id":"epoch:1c04442d32e715c1e093:parameters","entityId":"epoch:1c04442d32e715c1e093","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:4a627b01f80852b17612:parameters","entityId":"epoch:4a627b01f80852b17612","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-08-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:4a627b01f80852b17612:training-compute","entityId":"epoch:4a627b01f80852b17612","metric":"training-compute","value":3.15e+22,"unit":"FLOP","observationPeriod":"2023-08-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7b params, 750b tokens\n7b * 750b * 6 = 3.15e22"},{"id":"epoch:7fc4f002242d8da1c3a5:parameters","entityId":"epoch:7fc4f002242d8da1c3a5","metric":"parameters","value":53000000000.0,"unit":"parameters","observationPeriod":"2023-08-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:7fc4f002242d8da1c3a5:training-compute","entityId":"epoch:7fc4f002242d8da1c3a5","metric":"training-compute","value":8.268e+23,"unit":"FLOP","observationPeriod":"2023-08-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Given that it was announced at a similar time to the other Baichuan2 models, this assumes that the dataset size is the same at 2.6T tokens while the parameter count was scaled up. This would be consistent with many other model releases, such as Meta's Llama models.\n53b * 2.6t * 6 = 8.268e23\n"},{"id":"epoch:c3e595fe744b1ea0ce1d:parameters","entityId":"epoch:c3e595fe744b1ea0ce1d","metric":"parameters","value":14800000.0,"unit":"parameters","observationPeriod":"2023-08-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From Table 3"},{"id":"epoch:c3e595fe744b1ea0ce1d:training-compute","entityId":"epoch:c3e595fe744b1ea0ce1d","metric":"training-compute","value":2.28096e+19,"unit":"FLOP","observationPeriod":"2023-08-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 4x NVIDIA A30 GPUs (1.50\u00d710\u00b9\u2074 FLOP/s per GPU in FP16)\n2. Training duration: 1 day (86,400 seconds) - directly provided\n3. Utilization rate: 40%\n4. Calculation: 4 GPUs \u00d7 1.50\u00d710\u00b9\u2074 FLOP/s \u00d7 86,400s \u00d7 0.40 = 2.07\u00d710\u00b9\u2079 FLOPs\n\n4*1.7e+14*0.4*86400s=2.3e+19"},{"id":"epoch:68aeb5179a1d43d1a287:parameters","entityId":"epoch:68aeb5179a1d43d1a287","metric":"parameters","value":6890209280.0,"unit":"parameters","observationPeriod":"2023-08-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:68aeb5179a1d43d1a287:training-compute","entityId":"epoch:68aeb5179a1d43d1a287","metric":"training-compute","value":4.5e+22,"unit":"FLOP","observationPeriod":"2023-08-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"StableLM-Base-Alpha-7B-v2 is pre-trained using a multi-stage context length extension schedule following similar work (Nijkamp et al. 2023); first pre-training at a context length of 2048 for 1 trillion tokens, then fine-tuning at a context length of 4096 for another 100B tokens\"\n\n6890209280 params * 1.1 trillion tokens * 6 = 4.5e22\n\nalternatively: \"StableLM-Base-Alpha-7B-v2 was trained on the Stability AI cluster - occupying 384 NVIDIA A100 40GB GPUs across AWS P4d instances. Training took approximately 16.33 days to complete across both stages.\"\n\n312 teraflops * 384 * 16.33 * 24 * 3600 * 0.3 = 5.07e22"},{"id":"epoch:a70e820ec0c6bade0769:training-compute","entityId":"epoch:a70e820ec0c6bade0769","metric":"training-compute","value":7.56e+21,"unit":"FLOP","observationPeriod":"2023-08-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"ESM-2 650M is very likely the majority of FLOPs, since they only used 2 A100s (ESM-2 650M used 512 V100s for 8 days). As such I'm reporting the compute from ESM-2 650M here only."},{"id":"epoch:552f6a3409fef61fa668:parameters","entityId":"epoch:552f6a3409fef61fa668","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2023-08-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:552f6a3409fef61fa668:training-compute","entityId":"epoch:552f6a3409fef61fa668","metric":"training-compute","value":3.6e+22,"unit":"FLOP","observationPeriod":"2023-08-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 10B*600B * 6 = 3.6e22\" The model was trained on around 600B tokens from a mixture of the following corpora.\"\n\nSee also: https://weblab.t.u-tokyo.ac.jp/en/100%E5%84%84%E3%83%91%E3%83%A9%E3%83%A1%E3%83%BC%E3%82%BF%E3%82%B5%E3%82%A4%E3%82%BA%E3%83%BB%E6%97%A5%E8%8B%B12%E3%83%B6%E5%9B%BD%E8%AA%9E%E5%AF%BE%E5%BF%9C%E3%81%AE%E5%A4%A7%E8%A6%8F%E6%A8%A1/"},{"id":"epoch:553530f209d09ecbbb38:parameters","entityId":"epoch:553530f209d09ecbbb38","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-08-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:be372a234a5d6370ee95:parameters","entityId":"epoch:be372a234a5d6370ee95","metric":"parameters","value":9000000000.0,"unit":"parameters","observationPeriod":"2023-08-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We introduce OpenFlamingo, a family of autoregressive vision-language models ranging from 3B to 9B parameters.\""},{"id":"epoch:0bad6361f04ce4b12ac1:training-compute","entityId":"epoch:0bad6361f04ce4b12ac1","metric":"training-compute","value":4.03e+22,"unit":"FLOP","observationPeriod":"2023-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The training was conducted using 96 A100 80G GPUs, and the entire process took approximately 52 days.\"\n\n312 teraflop/s * 96 * 52 * 24 * 3600 * 0.3 = 4e22"},{"id":"epoch:eaedc6fad40c691ee51d:parameters","entityId":"epoch:eaedc6fad40c691ee51d","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:949d40bb57c906a2ac88:parameters","entityId":"epoch:949d40bb57c906a2ac88","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a7786c347ad75ec55bce:parameters","entityId":"epoch:a7786c347ad75ec55bce","metric":"parameters","value":3800000000.0,"unit":"parameters","observationPeriod":"2023-07-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.8 billion"},{"id":"epoch:a7786c347ad75ec55bce:training-compute","entityId":"epoch:a7786c347ad75ec55bce","metric":"training-compute","value":1.2e+22,"unit":"FLOP","observationPeriod":"2023-07-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.8 billion params * 524b tokens * 6 = 1.2e22"},{"id":"epoch:28e5493ff916858af6d3:parameters","entityId":"epoch:28e5493ff916858af6d3","metric":"parameters","value":55000000000.0,"unit":"parameters","observationPeriod":"2023-07-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train two specific instantiations of RT-2 that leverage pre-trained VLMs: (1) RT-2-PaLI-X is built from 5B and 55B PaLI-X (Chen et al., 2023a), and (2) RT-2-PaLM-E is built from 12B PaLM-E (Driess et al., 2023).\"\n\n55B and 12B have similar overall performance"},{"id":"epoch:ca07e4f45837479d7eb9:parameters","entityId":"epoch:ca07e4f45837479d7eb9","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-07-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d7eb15d3ae147aaa1f75:parameters","entityId":"epoch:d7eb15d3ae147aaa1f75","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2023-07-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use identical decoder-only Transformers in\nall stages, with 12 layers, 16 attention heads, embedding\ndimension of 1024, feed-forward layer dimension of 4096\nand dropout of 0.1, together with T5-style relative positional\nembeddings [38], resulting in a model parameter size of\n0.3B per stage.\"\n\nThree stages (figure 2), and 300M per stage. Plus 600M parameters for w2v-BERT XL, so 1.5B total"},{"id":"epoch:d7eb15d3ae147aaa1f75:training-compute","entityId":"epoch:d7eb15d3ae147aaa1f75","metric":"training-compute","value":3.9e+18,"unit":"FLOP","observationPeriod":"2023-07-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train each stage on 16 TPUv4s with batch size of 256 for 1M steps.\"\n\nThat's for the 900M-param transformers\n\nIf there's 256 passes in each batch, then using 6ND that's 900m * 256m * 6 = 1.3e18. sanity check: 16 tpu4s is 4.4e15 FLOP/s. 1.3e18 FLOP / 4.4e15 FLOP/s is 295 seconds. adjusting for utilization it would be ~1000 seconds or 15 minutes? probably too short, so 1.3e18 seems too low.\n\nupd there are 3 stages -> 1.3e18*3 = 3.9e+18 (Speculative due to reasoning above)"},{"id":"epoch:e96e61c645040e5febcf:parameters","entityId":"epoch:e96e61c645040e5febcf","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-07-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:1284e6c2093ac971d0c8:parameters","entityId":"epoch:1284e6c2093ac971d0c8","metric":"parameters","value":6900000.0,"unit":"parameters","observationPeriod":"2023-07-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"That is because LMDESIGN has 6.9M parameters while ProteinMPNN+CMLM only has 1.6M parameters.\""},{"id":"epoch:1284e6c2093ac971d0c8:training-compute","entityId":"epoch:1284e6c2093ac971d0c8","metric":"training-compute","value":1.400000000000012e+20,"unit":"FLOP","observationPeriod":"2023-07-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 8x NVIDIA V100 GPUs (1.25 x 10^14 FLOP/s per GPU)\n\n2. Training duration: 4 days (345,600 seconds) - estimated based on 10 epochs over 50M sequences with 6000 residues per batch\n\n3. Utilization rate: 40%\n\n4. Final calculation:\n(1.25 x 10^14 FLOP/s/GPU \u00d7 8 GPUs) \u00d7 345,600 seconds \u00d7 0.4 = 1.4 x 10^20 FLOPs"},{"id":"epoch:73e7ded0832e5ab225ad:parameters","entityId":"epoch:73e7ded0832e5ab225ad","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-07-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:bada4d040bc225f318a7:parameters","entityId":"epoch:bada4d040bc225f318a7","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-07-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:193ebaf53ee82aa3c81c:parameters","entityId":"epoch:193ebaf53ee82aa3c81c","metric":"parameters","value":65200000000.0,"unit":"parameters","observationPeriod":"2023-07-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"65.2B"},{"id":"epoch:6d39d314e168f7167129:parameters","entityId":"epoch:6d39d314e168f7167129","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2023-07-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"fine-tuned from Llama 2-70B"},{"id":"epoch:3d7d9fd0451e3ec24eea:parameters","entityId":"epoch:3d7d9fd0451e3ec24eea","metric":"parameters","value":300000000000.0,"unit":"parameters","observationPeriod":"2023-07-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"300 billion"},{"id":"epoch:b8a094554c7fc3dc265c:parameters","entityId":"epoch:b8a094554c7fc3dc265c","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2023-07-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Llama has been released in 7B, 13B, 34B, and 70B variants."},{"id":"epoch:b8a094554c7fc3dc265c:training-compute","entityId":"epoch:b8a094554c7fc3dc265c","metric":"training-compute","value":8.1e+23,"unit":"FLOP","observationPeriod":"2023-07-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Pretraining utilized a cumulative 3.3M GPU hours of computation on hardware of type A100-80GB\" of which 1720320 GPU hours were used to train the 70B model.\n\n311.84 BF16 TFLOP/s * 1720320 hours * 0.40 utilization = 7.725e+23 FLOP.\n\nAlternatively: the model was trained for 1 epoch on 2 trillion tokens and has 70B parameters. C = 6ND = 6*70B*2T = 8.4e+23 FLOP."},{"id":"epoch:cda7c117f2bf6d383e9f:parameters","entityId":"epoch:cda7c117f2bf6d383e9f","metric":"parameters","value":34000000000.0,"unit":"parameters","observationPeriod":"2023-07-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Llama has been released in 7B, 13B, 34B, and 70B variants."},{"id":"epoch:cda7c117f2bf6d383e9f:training-compute","entityId":"epoch:cda7c117f2bf6d383e9f","metric":"training-compute","value":4.08e+23,"unit":"FLOP","observationPeriod":"2023-07-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"All models sizes trained on 2.0T tokens, per table 1\n2T * 34b * 6 = 4.08e23\n\nAlso trained on 1038336 A100-hours, which is 3.5e23 at 30% utilization. So the utilization was probably around 35%."},{"id":"epoch:c12c3628066cb3ae4c37:parameters","entityId":"epoch:c12c3628066cb3ae4c37","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-07-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Llama has been released in 7B, 13B, and 70B variants."},{"id":"epoch:c12c3628066cb3ae4c37:training-compute","entityId":"epoch:c12c3628066cb3ae4c37","metric":"training-compute","value":8.4e+22,"unit":"FLOP","observationPeriod":"2023-07-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Trained on 2 trillion tokens per Table 1. \n\nC = 6ND = 6 FLOP / token / parameter * 7B parameters * 2T tokens = 8.4e+22 FLOP.\n\nAlso, 7B model was trained on 184320 GPU-hours\n\n312 trillion * 184320 GPU-hours * 3600 sec/hour * 0.3 [utilization] = 6.21e22 FLOP"},{"id":"epoch:1eb59cec44d5d07eeb79:parameters","entityId":"epoch:1eb59cec44d5d07eeb79","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-07-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Llama has been released in 7B, 13B, and 70B variants."},{"id":"epoch:1eb59cec44d5d07eeb79:training-compute","entityId":"epoch:1eb59cec44d5d07eeb79","metric":"training-compute","value":1.6e+23,"unit":"FLOP","observationPeriod":"2023-07-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13 billion parameters * 2 trillion tokens * 6 FLOP / token / parameter = 1.6e23 FLOP"},{"id":"epoch:e54c99ab9808a4739d3b:parameters","entityId":"epoch:e54c99ab9808a4739d3b","metric":"parameters","value":2700000000.0,"unit":"parameters","observationPeriod":"2023-07-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c3e605adf00c3f56d2c0:parameters","entityId":"epoch:c3e605adf00c3f56d2c0","metric":"parameters","value":6700000000.0,"unit":"parameters","observationPeriod":"2023-07-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2\n\nThey later mention testing the memory and throughput of a 13B-parameter model, but it doesn't sound like they trained it long enough to test its perplexity."},{"id":"epoch:c3e605adf00c3f56d2c0:training-compute","entityId":"epoch:c3e605adf00c3f56d2c0","metric":"training-compute","value":4.02e+21,"unit":"FLOP","observationPeriod":"2023-07-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"C = 6ND = 6 * 6.7 billion * 100 billion\n\"We train the models with 512 AMD MI200 GPUs.\"\n\nAnother rough method: Table 4 shows throughput is 8642.2 words per second during training on an 8xA100 setup, with batch size of 8192. This suggests 2543.1 hours to go through the full 100B tokens ~=  75B words.\n(8 * 3.12e14) * (2543.1 * 3600) * (0.3) = 6.9e21 FLOPs\n\n(low confidence in this as actual training setup used different hardware and much larger batch size, but same OOM is encouraging)"},{"id":"epoch:363e5e9b252aa3ceec13:parameters","entityId":"epoch:363e5e9b252aa3ceec13","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2023-07-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"ChatRhino sets a new benchmark as a 100-billion-parameter model\", could be substantially rounded"},{"id":"epoch:0ce782e18ab7018dcf66:parameters","entityId":"epoch:0ce782e18ab7018dcf66","metric":"parameters","value":25000000.0,"unit":"parameters","observationPeriod":"2023-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 8: Model architecture parameters of different Uni-RNA models"},{"id":"epoch:3ca472f6e7cdda3d4918:parameters","entityId":"epoch:3ca472f6e7cdda3d4918","metric":"parameters","value":540000000000.0,"unit":"parameters","observationPeriod":"2023-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We performed instruction prompt tuning on Flan-PaLM 540B with a soft prompt length of 100 to produce Med-PaLM. We froze the rest of the model, and used an embedding dimension of 18432 as in PaLM [1], which resulted in 1.84M trainable parameters.\"\n(from supplementary materials)"},{"id":"epoch:19492f14ba2176865c63:parameters","entityId":"epoch:19492f14ba2176865c63","metric":"parameters","value":400000000.0,"unit":"parameters","observationPeriod":"2023-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 8: Model architecture parameters of different Uni-RNA models"},{"id":"epoch:2164948363852c0083cd:parameters","entityId":"epoch:2164948363852c0083cd","metric":"parameters","value":85000000.0,"unit":"parameters","observationPeriod":"2023-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 8: Model architecture parameters of different Uni-RNA models"},{"id":"epoch:54f3267f0a6182123c57:parameters","entityId":"epoch:54f3267f0a6182123c57","metric":"parameters","value":169000000.0,"unit":"parameters","observationPeriod":"2023-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 8: Model architecture parameters of different Uni-RNA models\n\nEstimating sequence length at 500 tokens.\n\n500000000*500=250000000000"},{"id":"epoch:45d2e51b513a645a94b8:training-compute","entityId":"epoch:45d2e51b513a645a94b8","metric":"training-compute","value":3.866e+24,"unit":"FLOP","observationPeriod":"2023-07-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://colab.research.google.com/drive/1MdPuhS4Emaf23VXYZ-ooExDW-5GXZkw0#scrollTo=Ds0Q5X8aMnOY"},{"id":"epoch:d064f985df845d615ae8:parameters","entityId":"epoch:d064f985df845d615ae8","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2023-07-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The total number of parameters of Emu is 14B and is trained end-to-end.\"\n\n\" We leverage pretrained EVA-CLIP (Sun et al., 2023), LLaMA (Touvron et al., 2023) and Stable Diffusion (Rombach et al., 2022) to initialize the Visual Encoder, the Multimodal Modeling LLM and the Visual Decoder, respectively.\""},{"id":"epoch:d064f985df845d615ae8:training-compute","entityId":"epoch:d064f985df845d615ae8","metric":"training-compute","value":2.70000000001e+21,"unit":"FLOP","observationPeriod":"2023-07-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train the model on 128 NVIDIA 80G-A100 GPUs for 10k steps with around 82M samples (150B tokens in total), and the pretraining takes approximately 2 days.\"\nhttps://www.wolframalpha.com/input?i=128*312+TFLOPS+*+2+days+*+0.4"},{"id":"epoch:4cc903dc1abdcf54ec4b:parameters","entityId":"epoch:4cc903dc1abdcf54ec4b","metric":"parameters","value":13264901120.0,"unit":"parameters","observationPeriod":"2023-07-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:4cc903dc1abdcf54ec4b:training-compute","entityId":"epoch:4cc903dc1abdcf54ec4b","metric":"training-compute","value":9.36e+22,"unit":"FLOP","observationPeriod":"2023-07-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13b parameters * 1.2t tokens * 6 FLOP / parameter / token = 9.36e22 FLOP"},{"id":"epoch:00a0f94bdb1462c586cc:parameters","entityId":"epoch:00a0f94bdb1462c586cc","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2023-07-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"100B? I think the five foundation models are all included in the same system, instead of being five different variants of Pangu, but that's not very clear. I think that's implied by \"All of these capability sets will be made available to customers and partners, and will be consistent regardless of the model size\". \n\nhttps://www.huaweicloud.com/intl/en-us/news/20230707180809498.html \n\"Huawei Cloud Pangu models were designed to focus on the practical needs of specific industry scenarios. The newly-launched Pangu Models 3.0 use a \"5+N+X\" three-layer architecture.\n\nThe L0 layer consists of five foundation models: NLP, CV, multimodal, prediction, and scientific computing, which provide general skills to power an endless possibility of industry-specific applications. Pangu Models 3.0 are available in different sizes: 10 billion parameters, 38 billion parameters, 71 billion parameters, and 100 billion parameters, meeting diverse customer needs and different standards on latency and response times. Brand-new capability sets are also provided, such as knowledge-based Q&A, copywriting, and code generation for the Pangu NLP model; and image generation and understanding for the Pangu multimodal model. All of these capability sets will be made available to customers and partners, and will be consistent regardless of the model size (number of parameters).\n\nThe L1 layer consists of N industry-tailored models. Huawei Cloud can provide customers with industry models it has trained on open industry datasets, including Pangu models for government, finance, manufacturing, mining, and meteorology. Alternatively, customers can train their own models using their own datasets based on Huawei's L0 or L1 Pangu models.\n\nThe L2 layer provides pre-trained models for specific industry scenarios and tasks, such as intelligent government hotline, intelligent branch assistant, lead compound screening, conveyor belt foreign object detection, and typhoon trajectory prediction. These models can be quickly deployed off-the-shelf.\"\n\n"},{"id":"epoch:46855fb2d44d42558d17:parameters","entityId":"epoch:46855fb2d44d42558d17","metric":"parameters","value":104000000000.0,"unit":"parameters","observationPeriod":"2023-07-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We present InternLM, a multilingual foundational language model with 104B parameters\""},{"id":"epoch:46855fb2d44d42558d17:training-compute","entityId":"epoch:46855fb2d44d42558d17","metric":"training-compute","value":9.984e+23,"unit":"FLOP","observationPeriod":"2023-07-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 * 104b * 1.6T = 9.984e23"},{"id":"epoch:65e20aef10567b73c88c:parameters","entityId":"epoch:65e20aef10567b73c88c","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-07-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:65e20aef10567b73c88c:training-compute","entityId":"epoch:65e20aef10567b73c88c","metric":"training-compute","value":5.9e+22,"unit":"FLOP","observationPeriod":"2023-07-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B parameters, trained on 1.4T tokens\n\n7 billion * 1.4 trillion * 6 = 5.9e22"},{"id":"epoch:457e7cd3c78ae743ae37:parameters","entityId":"epoch:457e7cd3c78ae743ae37","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2023-07-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Abstract: \"training xTrimoPGLM at an unprecedented scale of 100 billion\nparameters and 1 trillion training tokens\""},{"id":"epoch:457e7cd3c78ae743ae37:training-compute","entityId":"epoch:457e7cd3c78ae743ae37","metric":"training-compute","value":6.2e+23,"unit":"FLOP","observationPeriod":"2023-07-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"xTrimoPGLM-100B is trained on a cluster of 96 DGX-A100 GPU (8\u00d740G) servers in FP16 precision from January 18 to June 30, 2023. During this time, xTrimoPGLM-100B has consumed 1 trillion tokens from the dataset consisting of Uniref90 and ColAbFoldDB. As of the current date, xTrimoPGLM-100B continues its pre-training process to pass through as many tokens as possible\"\n\n6 * 100 billion params * 1T tokens = 6e23\n\n8*96 * 312 trillion * 163 days * 24 * 3600 * 0.3 ~= 1e24\n\ndirectly given in the paper (Table 9, or Table 4 in new version): 6.2E+23 "},{"id":"epoch:31b0926f7f16aa633e2f:parameters","entityId":"epoch:31b0926f7f16aa633e2f","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-07-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:9d17c89a0906407b7bcd:parameters","entityId":"epoch:9d17c89a0906407b7bcd","metric":"parameters","value":256000000.0,"unit":"parameters","observationPeriod":"2023-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4*64 million = 256M params\n\n\"We trained four deep networks with lead times (the time difference\nbetween input and output) at 1\u2009h, 3\u2009h, 6\u2009h and 24\u2009h, respectively... \n\nThis modification increases the number of bias parameters by a factor of 527, with each 3D deep network containing approximately 64\u2009million parameters.\""},{"id":"epoch:9d17c89a0906407b7bcd:training-compute","entityId":"epoch:9d17c89a0906407b7bcd","metric":"training-compute","value":3.98e+22,"unit":"FLOP","observationPeriod":"2023-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Each of the four deep networks was trained for 100 epochs, and\neach of them takes approximately 16\u2009days on a cluster of 192 NVIDIA\nTesla-V100 GPUs.\"\n\n192 * 4 * 16 * 24 * 3600 * 125 teraflops * 0.3 utilization = 3.98e22"},{"id":"epoch:2613c3a24212862bee9d:parameters","entityId":"epoch:2613c3a24212862bee9d","metric":"parameters","value":2700000000.0,"unit":"parameters","observationPeriod":"2023-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.7B"},{"id":"epoch:2613c3a24212862bee9d:training-compute","entityId":"epoch:2613c3a24212862bee9d","metric":"training-compute","value":4.86e+21,"unit":"FLOP","observationPeriod":"2023-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.7B params * 300B tokens * 6 = 4.86e21\n\nNote: not sure if there are very long sequences in the training data that would affect this calculation. Per paper, complexity of their attention mechanism scales linearly with sequence length."},{"id":"epoch:9af3bdb731ac02dbabcf:parameters","entityId":"epoch:9af3bdb731ac02dbabcf","metric":"parameters","value":3400000000.0,"unit":"parameters","observationPeriod":"2023-07-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"...result in a model size of 2.6B parameters in the UNet, see Tab. 1. The text encoders have a total size of 817M parameters.\""},{"id":"epoch:7044d0c6e2612b7b8e88:parameters","entityId":"epoch:7044d0c6e2612b7b8e88","metric":"parameters","value":560000000.0,"unit":"parameters","observationPeriod":"2023-06-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"560M from https://huggingface.co/intfloat/multilingual-e5-large"},{"id":"epoch:7044d0c6e2612b7b8e88:training-compute","entityId":"epoch:7044d0c6e2612b7b8e88","metric":"training-compute","value":3.370752e+18,"unit":"FLOP","observationPeriod":"2023-06-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*560000000*(1000000000+1600000*2 epochs) = 3.370752e+18\n\nconfidence 'likely\" because pre-training epochs are unknown"},{"id":"epoch:11902e71d684ff55236c:parameters","entityId":"epoch:11902e71d684ff55236c","metric":"parameters","value":2000000000.0,"unit":"parameters","observationPeriod":"2023-06-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The model now has 800 million training parameters and 2 billion training parameters in two versions, which can fulfill functions such as text-to-picture and picture-to-picture, said staff members of the company.\" https://www.chinadaily.com.cn/a/202309/02/WS64f2715aa310d2dce4bb3865.html"},{"id":"epoch:7cda389f135fd31ecff1:parameters","entityId":"epoch:7cda389f135fd31ecff1","metric":"parameters","value":6600000.0,"unit":"parameters","observationPeriod":"2023-06-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table A.1 shows details, largest experiment is on far right."},{"id":"epoch:7cda389f135fd31ecff1:training-compute","entityId":"epoch:7cda389f135fd31ecff1","metric":"training-compute","value":1.811e+21,"unit":"FLOP","observationPeriod":"2023-06-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8 Nvidia A100 (80GB) GPUs, ~4 weeks\n(4 * 7 * 24 * 3600) seconds * (8 * 3.12e14) FLOP/sec * 0.3 (utilization) = 1.811e21"},{"id":"epoch:cde599c6d7b1fede0516:parameters","entityId":"epoch:cde599c6d7b1fede0516","metric":"parameters","value":2000000000.0,"unit":"parameters","observationPeriod":"2023-06-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"On Google Vertex AI, you can access Chirp, a 2B-parameter speech model which achieved incredible accuracy improvements in speech recognition over several languages\" (https://medium.com/google-cloud/getting-started-with-chirp-the-googles-universal-speech-model-usm-on-vertex-ai-f54edaf4da93)."},{"id":"epoch:c36e316c60361a75aa22:parameters","entityId":"epoch:c36e316c60361a75aa22","metric":"parameters","value":7393000000.0,"unit":"parameters","observationPeriod":"2023-06-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B\nTable 2 https://arxiv.org/pdf/2305.13048"},{"id":"epoch:309a766222d14a949770:parameters","entityId":"epoch:309a766222d14a949770","metric":"parameters","value":1600000000.0,"unit":"parameters","observationPeriod":"2023-06-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.6B"},{"id":"epoch:309a766222d14a949770:training-compute","entityId":"epoch:309a766222d14a949770","metric":"training-compute","value":4.5500354284e+20,"unit":"FLOP","observationPeriod":"2023-06-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train the model on 256 V100 GPUs and the training takes approximately one day to complete\"\n\"We train KOSMOS-2 for 60k steps, equivalent to around 25 billion tokens\"\n\nGPU-time method\n(256) * (1.3e14) * (24 * 3600) * (0.3) = 8.626176e20\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)\n\nParameter-data method\n6ND = 6*25B*1.6B = 2.4e20\n\nUsed geometric mean of two estimates."},{"id":"epoch:b1dc46f914520275792b:parameters","entityId":"epoch:b1dc46f914520275792b","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-06-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:eb2f99da05ea4965be20:training-compute","entityId":"epoch:eb2f99da05ea4965be20","metric":"training-compute","value":1.0001e+24,"unit":"FLOP","observationPeriod":"2023-06-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"<= 2.5e24\n\nThey define two \"compute classes\", one for models with more compute than PaLM 540B, i.e. GPT-4 and PaLM 2, and one for models with as much compute or less, i.e. GPT-3.5, Chinchilla, LLaMA, and Inflection-1.\n\nPaLM 540B required 2.5e24 FLOP to train (confirmed by Google)"},{"id":"epoch:080d7f2a20668a7bfd01:parameters","entityId":"epoch:080d7f2a20668a7bfd01","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2023-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"30B"},{"id":"epoch:080d7f2a20668a7bfd01:training-compute","entityId":"epoch:080d7f2a20668a7bfd01","metric":"training-compute","value":1.8900000000001e+23,"unit":"FLOP","observationPeriod":"2023-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"According to their blog post,\n\"MPT-30B FLOPs ~= 6 * 30e9 [params] * 1.05e12 [tokens] = 1.89e23 FLOPs\""},{"id":"epoch:845d312088523b191566:parameters","entityId":"epoch:845d312088523b191566","metric":"parameters","value":33000000000.0,"unit":"parameters","observationPeriod":"2023-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5a035ef8bfd2aa6e106b:parameters","entityId":"epoch:5a035ef8bfd2aa6e106b","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2b425f82835908d545ab:parameters","entityId":"epoch:2b425f82835908d545ab","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:280a4df48f53601c93d8:parameters","entityId":"epoch:280a4df48f53601c93d8","metric":"parameters","value":1180000000.0,"unit":"parameters","observationPeriod":"2023-06-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Most of the experimental results are based on models with a 1.18B-parameter decoder-only transformer (Vaswani et al., 2017) with 24 layers, an embedding size of 2048, and a post-attention feedforward hidden size of 8196.\" page 8"},{"id":"epoch:34694a28bff27e845601:parameters","entityId":"epoch:34694a28bff27e845601","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2023-06-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1B (Table 2)"},{"id":"epoch:34694a28bff27e845601:training-compute","entityId":"epoch:34694a28bff27e845601","metric":"training-compute","value":3.8680312e+22,"unit":"FLOP","observationPeriod":"2023-06-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"GigaGAN and SD-v1.5 require 4,783 and 6,250 A100 GPU days\"\n\n312*10^12 FLOP / sec / GPU * 4783 GPU-days * 24 hours / day * 3600 sec / hour * 0.3 [assumed utilization] = 3.8680312e+22 FLOP"},{"id":"epoch:6ec4e0e6abdf968bed39:parameters","entityId":"epoch:6ec4e0e6abdf968bed39","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2023-06-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.3B"},{"id":"epoch:6ec4e0e6abdf968bed39:training-compute","entityId":"epoch:6ec4e0e6abdf968bed39","metric":"training-compute","value":1.7380147e+20,"unit":"FLOP","observationPeriod":"2023-06-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The large model is pretrained for 170K steps with a batch size of 1024\"\n\"The decoder sequence length is 128 tokens\"\n\n6 FLOP / token / parameter * 1.3 * 10^9 parameters * 128 tokens / sample* 1024 samples / step * 170000 steps = 1.7380147e+20 FLOP"},{"id":"epoch:795da8095d041a593a32:parameters","entityId":"epoch:795da8095d041a593a32","metric":"parameters","value":15500000000.0,"unit":"parameters","observationPeriod":"2023-06-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"15.5B"},{"id":"epoch:795da8095d041a593a32:training-compute","entityId":"epoch:795da8095d041a593a32","metric":"training-compute","value":1.12e+23,"unit":"FLOP","observationPeriod":"2023-06-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.12e23 base compute (StarCoder estimate) + 1.95e19 finetune compute (see below) ~= 1.12e23"},{"id":"epoch:ac802aeb5ad7a4796023:parameters","entityId":"epoch:ac802aeb5ad7a4796023","metric":"parameters","value":1800000000000.0,"unit":"parameters","observationPeriod":"2023-06-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Rumored to be 1.8T parameter MoE with 280B activated on the forward pass, per https://www.semianalysis.com/p/gpt-4-architecture-infrastructure. Other sources estimate 1.76T with 220B per forward pass https://web.archive.org/web/20230712123915/https://the-decoder.com/gpt-4-architecture-datasets-costs-and-more-leaked/"},{"id":"epoch:ac802aeb5ad7a4796023:training-compute","entityId":"epoch:ac802aeb5ad7a4796023","metric":"training-compute","value":2.1e+25,"unit":"FLOP","observationPeriod":"2023-06-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"90% CI: 8.2E+24 to 4.4E+25\n\nNOTE: this is a rough estimate based on public information, much less information than most other systems in the database.\n\nCalculation and confidence intervals here: https://colab.research.google.com/drive/1O99z9b1I5O66bT78r9ScslE_nOj5irN9?usp=sharing"},{"id":"epoch:6405c236a41ff0e40197:parameters","entityId":"epoch:6405c236a41ff0e40197","metric":"parameters","value":20000000000.0,"unit":"parameters","observationPeriod":"2023-06-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"20B parameters according to Table 1 in Microsoft's CODEFUSION paper: https://arxiv.org/pdf/2310.17680.pdf"},{"id":"epoch:993059ba7e26ab6bbaed:parameters","entityId":"epoch:993059ba7e26ab6bbaed","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-06-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:18c9419f1fa6286b599c:parameters","entityId":"epoch:18c9419f1fa6286b599c","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-06-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:514ce9fa401a26ef2f11:parameters","entityId":"epoch:514ce9fa401a26ef2f11","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-06-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:33b0326c45822a251851:parameters","entityId":"epoch:33b0326c45822a251851","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-06-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:4b1cb45bd170772eaea3:parameters","entityId":"epoch:4b1cb45bd170772eaea3","metric":"parameters","value":33000000000.0,"unit":"parameters","observationPeriod":"2023-06-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"33B for largest model: https://huggingface.co/BAAI/Aquila-7B"},{"id":"epoch:1f6319039fe4a6efa599:parameters","entityId":"epoch:1f6319039fe4a6efa599","metric":"parameters","value":57000000.0,"unit":"parameters","observationPeriod":"2023-06-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1f6319039fe4a6efa599:training-compute","entityId":"epoch:1f6319039fe4a6efa599","metric":"training-compute","value":2.3000000000000197e+20,"unit":"FLOP","observationPeriod":"2023-06-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 7x NVIDIA A100 GPUs, 3.12e14 FLOP/s per GPU\n2. Training duration: 3 days = 259,200 seconds (directly provided)\n3. Utilization rate: 40% (default assumption)\n4. Calculation: 3.12e14 FLOP/s \u00d7 7 GPUs \u00d7 259,200s \u00d7 0.4 = 2.3e20 FLOPs"},{"id":"epoch:b002942b9cc355ac8e2e:parameters","entityId":"epoch:b002942b9cc355ac8e2e","metric":"parameters","value":420000000.0,"unit":"parameters","observationPeriod":"2023-06-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:93befb226a72412c68ff:parameters","entityId":"epoch:93befb226a72412c68ff","metric":"parameters","value":3359000000.0,"unit":"parameters","observationPeriod":"2023-06-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train autoregressive transformer models at different sizes: 300M, 1.5B, 3.3B parameters\"\n\nUses EnCodec 32kHz (HF version has 59M params) for audio tokenization."},{"id":"epoch:0546fc6fcb57a143bf00:parameters","entityId":"epoch:0546fc6fcb57a143bf00","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2023-06-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14B from https://inside.ai/en/news/2023/06/08/aiinside-xresearch/\n"},{"id":"epoch:d172c52893dc37f3cc9b:parameters","entityId":"epoch:d172c52893dc37f3cc9b","metric":"parameters","value":6900000000.0,"unit":"parameters","observationPeriod":"2023-06-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6.9b"},{"id":"epoch:d172c52893dc37f3cc9b:training-compute","entityId":"epoch:d172c52893dc37f3cc9b","metric":"training-compute","value":4.1e+22,"unit":"FLOP","observationPeriod":"2023-06-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Trained over 1.001 trillion tokens.\n6.9b * 1 trillion * 6 = 4.1e22\n\n"},{"id":"epoch:39749a9459edc38a1cf2:parameters","entityId":"epoch:39749a9459edc38a1cf2","metric":"parameters","value":9888.0,"unit":"parameters","observationPeriod":"2023-06-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://docs.google.com/spreadsheets/d/1rnsk2ysbAra1UfQWD9TFDki-0T9OmKqBdd7fw6O7QWE/edit?usp=sharing"},{"id":"epoch:39749a9459edc38a1cf2:training-compute","entityId":"epoch:39749a9459edc38a1cf2","metric":"training-compute","value":741600000000.0,"unit":"FLOP","observationPeriod":"2023-06-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*9888*50000*250=741600000000"},{"id":"epoch:75aaea891fcbcce402bb:parameters","entityId":"epoch:75aaea891fcbcce402bb","metric":"parameters","value":8400000.0,"unit":"parameters","observationPeriod":"2023-06-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Appendix B: 8.4m"},{"id":"epoch:75aaea891fcbcce402bb:training-compute","entityId":"epoch:75aaea891fcbcce402bb","metric":"training-compute","value":163905134400000.0,"unit":"FLOP","observationPeriod":"2023-06-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"=6*3252086*8400000=1.639051344 \u00d7 10^14\n\nI am not sure about datasize thus Speculative confidence level"},{"id":"epoch:7ea5090576290727df23:parameters","entityId":"epoch:7ea5090576290727df23","metric":"parameters","value":12898631680.0,"unit":"parameters","observationPeriod":"2023-06-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:7ea5090576290727df23:training-compute","entityId":"epoch:7ea5090576290727df23","metric":"training-compute","value":1.28e+22,"unit":"FLOP","observationPeriod":"2023-06-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"trained for 167 billion tokens\n\n167b * 12.8b * 6 = 1.28e22"},{"id":"epoch:533ec29c9705731f6d53:parameters","entityId":"epoch:533ec29c9705731f6d53","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:4e5d614e9491db423c59:parameters","entityId":"epoch:4e5d614e9491db423c59","metric":"parameters","value":7000559616.0,"unit":"parameters","observationPeriod":"2023-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:4e5d614e9491db423c59:training-compute","entityId":"epoch:4e5d614e9491db423c59","metric":"training-compute","value":5.04e+22,"unit":"FLOP","observationPeriod":"2023-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7b parameters * 1.2t tokens * 6 FLOP / parameter / token = 5.04e22 FLOP"},{"id":"epoch:79ce2106371288d21cbc:parameters","entityId":"epoch:79ce2106371288d21cbc","metric":"parameters","value":92000000.0,"unit":"parameters","observationPeriod":"2023-05-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"92M (Table 9)"},{"id":"epoch:79ce2106371288d21cbc:training-compute","entityId":"epoch:79ce2106371288d21cbc","metric":"training-compute","value":1.656e+20,"unit":"FLOP","observationPeriod":"2023-05-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 92 * 10^6 parameters * 3 * 10^11 tokens [see dataset size notes] = 1.656e+20 FLOP"},{"id":"epoch:73fc904d18fa8911d25e:training-compute","entityId":"epoch:73fc904d18fa8911d25e","metric":"training-compute","value":1.992646656e+22,"unit":"FLOP","observationPeriod":"2023-05-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000 FLOP/GPU/sec * 672 hours * 3600 sec / hour * 88 GPUs * 0.3 [assumed utilization] = 1.992646656e+22 FLOP"},{"id":"epoch:5d55270f2d6dd8727827:parameters","entityId":"epoch:5d55270f2d6dd8727827","metric":"parameters","value":55000000000.0,"unit":"parameters","observationPeriod":"2023-05-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"55B (table 1)"},{"id":"epoch:5d55270f2d6dd8727827:training-compute","entityId":"epoch:5d55270f2d6dd8727827","metric":"training-compute","value":5.6e+23,"unit":"FLOP","observationPeriod":"2023-05-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we present PaLI-X [...] consisting of a pretrained large-capacity visual encoder (using [6] as the starting point) and a pretrained language-only encoder-decoder (using [7] as the starting point), further trained at-scale on a vision-and-language data mixture using a combination of self-supervision and full-supervision signals.\"\n\n\"Visual component Our visual backbone is scaled to 22B parameters, as introduced by [6], the largest dense ViT model to date.\" (elsewhere they specify this is ViT-22B)\nOriginal ViT-22B pre-training compute: 1.93248e+23 FLOP\n\n\"Overall model The encoder-decoder backbone is initialized from a variant of the UL2 [7] encoder-decoder model that uses 32B parameters. The architecture of this variant has 50 layers in both encoder and decoder (up from 32 layers in [7]), and is pretrained on a mixture of text data similar to [7].\" ([7] = UL2 paper)\nOriginal UL2 pre-training compute: 1.2e+23 FLOP\nBut that was for a 20B model. Assuming the same amount of text data, compute is roughly 32B/20B * 1.2e23 = 1.92e+23\n\nFurther training:]\n\n\"WebLI [5], consisting of roughly one billion images with alt-texts from the web and OCR annotations.\"'\n\n\"In addition to WebLI \u27e8image, text\u27e9 pairs, we introduce here Episodic WebLI data, where each episode corresponds to a set of such pairs. We aim to have each episode contain loosely related images (i.e., they are clustered according to their URL field), so as to encourage attention among examples in an \u201cepisode\u201d. We find this new dataset (with 75M episodes and around 400M images in total) important for developing the few-shot capabilities of the model.\"\n\n\"The pretraining mixture consists of the following data and objectives [...]\" - this seems to be the \"mixed-objective training\" referred to later\n\n\"Our model is trained in two stages. In stage 1, the visual encoder (after mixed-objective training) is kept frozen, while the rest of the parameters are trained on a total of 2.2B examples at the base resolution 224\u00d7224 (native to ViT-22B), using the entire mixture. In stage 2, it continues training using only the OCR-related objectives (pix2struct and split-ocr) plus the object detection objective; this is done in several substages, during which image resolution is gradually increased to 448\u00d7448, 672\u00d7672 and finally 756\u00d7756.\"\n\n\"256 tokens per image\"\n\n2.2B images * 256 tokens per image = 5.632e+11 tokens\n32B parameters for backward pass because the 22B visual encoder is frozen.\nSo stage 1 compute estimate is: 5.632e+11 * (55e9 * 2 forward + 32e9 * 6 backward) = 1.7e+23 FLOP\n\nUnsure how to estimate compute for stage 2 - will neglect that here.\n\nFine-tuning compute seems negligible - Table 20 implies millions of examples.\n\nSo total compute estimate:\n1.93248e+23 FLOP ViT-22B original\n1.92e+23 FLOP UL2 32B original\n1.7e+23 FLOP stage 1 training\n5.6e23 FLOP total"},{"id":"epoch:08fa7a89259db8be0059:parameters","entityId":"epoch:08fa7a89259db8be0059","metric":"parameters","value":2800000000.0,"unit":"parameters","observationPeriod":"2023-05-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"same as base model"},{"id":"epoch:fab36804822bbdd56158:parameters","entityId":"epoch:fab36804822bbdd56158","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2023-05-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d0786cf4a0cf521a3842:parameters","entityId":"epoch:d0786cf4a0cf521a3842","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-05-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"HuatuoGPT-13B is trained on Ziya-LLaMA-13B-Pretrain-v1"},{"id":"epoch:c8eed4e4308e29e0f56d:parameters","entityId":"epoch:c8eed4e4308e29e0f56d","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-05-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:39003c512f069bb10383:parameters","entityId":"epoch:39003c512f069bb10383","metric":"parameters","value":65200000000.0,"unit":"parameters","observationPeriod":"2023-05-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c841d61653c80ebeb477:parameters","entityId":"epoch:c841d61653c80ebeb477","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-05-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b0a6537047dbf7667e25:parameters","entityId":"epoch:b0a6537047dbf7667e25","metric":"parameters","value":33000000000.0,"unit":"parameters","observationPeriod":"2023-05-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:557797e0d5571f8aba2f:parameters","entityId":"epoch:557797e0d5571f8aba2f","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-05-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:94d39a20fd3bf2169be2:parameters","entityId":"epoch:94d39a20fd3bf2169be2","metric":"parameters","value":65000000000.0,"unit":"parameters","observationPeriod":"2023-05-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from Llama-65B \n(also 33B, 13B, 7B variants)"},{"id":"epoch:94d39a20fd3bf2169be2:training-compute","entityId":"epoch:94d39a20fd3bf2169be2","metric":"training-compute","value":5.5e+23,"unit":"FLOP","observationPeriod":"2023-05-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Fine-tune of Llama-65B, which appears to have been trained on a \"professional grade GPU\" with 48GB VRAM (likely A6000) for 24 hours.\n\nFine-tune compute is negligible compared to pretraining (5.5e23 for Llama-65b)"},{"id":"epoch:aebf1dcb69e1b69764d7:parameters","entityId":"epoch:aebf1dcb69e1b69764d7","metric":"parameters","value":182000000.0,"unit":"parameters","observationPeriod":"2023-05-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:4844ba259a8307691c56:parameters","entityId":"epoch:4844ba259a8307691c56","metric":"parameters","value":14000000000.0,"unit":"parameters","observationPeriod":"2023-05-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14b"},{"id":"epoch:4844ba259a8307691c56:training-compute","entityId":"epoch:4844ba259a8307691c56","metric":"training-compute","value":2.78e+22,"unit":"FLOP","observationPeriod":"2023-05-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from HuggingFace page: https://huggingface.co/BlinkDL/rwkv-4-pile-14b\n\ntrained for 330B tokens\n14 billion * 330 billion * 6 = 2.78e22\n\npaper notes that a forward pass is almost exactly 2x parameters (within 2%): \"Alternative approximations for FLOPs include doubling the parameters which yields similar results within 2% for 14B and a 30% discrepancy for 169M variant.\" and that 6*params*tokens is a good approximation because it's not a transformer: \"FLOPs is for a forward pass for one token. It was calculated as 6(V D + 13D2L), which is the\ntwice (add and multiply) the number of parameters\nin linear layers. The backwards pass FLOPs can be\napproximated as twice that of the forward pass. So\nthe total is 6(V D + 13D2L) per token for training\n(3x fw FLOPs). It is noteworthy that FLOPs are\nindependent of the context length, unlike regular\ntransformers\""},{"id":"epoch:8c2cdf2bef6cf6215907:parameters","entityId":"epoch:8c2cdf2bef6cf6215907","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2023-05-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/facebook/mms-1b-all/tree/main"},{"id":"epoch:eb0f0f0e1aa56a87cbd6:parameters","entityId":"epoch:eb0f0f0e1aa56a87cbd6","metric":"parameters","value":16000000000.0,"unit":"parameters","observationPeriod":"2023-05-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We implemented a family of CodeT5+ models, with model sizes ranging from 220M to 16B\""},{"id":"epoch:b84f4eedf512ef0fd3fe:parameters","entityId":"epoch:b84f4eedf512ef0fd3fe","metric":"parameters","value":176200000000.0,"unit":"parameters","observationPeriod":"2023-05-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"176.2B"},{"id":"epoch:ae247fffb59dd13b769d:parameters","entityId":"epoch:ae247fffb59dd13b769d","metric":"parameters","value":4000000000.0,"unit":"parameters","observationPeriod":"2023-05-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we propose ONE-PEACE, a model with 4B parameters\""},{"id":"epoch:ae247fffb59dd13b769d:training-compute","entityId":"epoch:ae247fffb59dd13b769d","metric":"training-compute","value":1.8e+20,"unit":"FLOP","observationPeriod":"2023-05-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4 billion params * 7.5 billion data * 6 = 1.8e20.\n\nsee training dataset size notes. this estimate required some more assumptions than usual."},{"id":"epoch:471ecebbfed4bedf6d43:parameters","entityId":"epoch:471ecebbfed4bedf6d43","metric":"parameters","value":65000000000.0,"unit":"parameters","observationPeriod":"2023-05-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train LIMA (Less Is More for Alignment) using the following protocol. Starting from LLaMa 65B [Touvron et al., 2023], we fine-tune on our 1,000-example alignment training set,\" according to page 4 of https://arxiv.org/pdf/2305.11206."},{"id":"epoch:471ecebbfed4bedf6d43:training-compute","entityId":"epoch:471ecebbfed4bedf6d43","metric":"training-compute","value":5.5000439e+23,"unit":"FLOP","observationPeriod":"2023-05-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Finetune:  4.39e18 FLOP\nBase model: 5.5e+23 FLOP \nTotal: 4.39e18+5.5e+23=5.5e+23 FLOP"},{"id":"epoch:7055fca8962cd95cb995:parameters","entityId":"epoch:7055fca8962cd95cb995","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2023-05-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"11B"},{"id":"epoch:daf72083bf1eb465e31a:parameters","entityId":"epoch:daf72083bf1eb465e31a","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2023-05-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"WeLM is trained with 10B parameters\u201d"},{"id":"epoch:daf72083bf1eb465e31a:training-compute","entityId":"epoch:daf72083bf1eb465e31a","metric":"training-compute","value":2.484338688e+22,"unit":"FLOP","observationPeriod":"2023-05-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The largest model is trained on 128 A100-SXM4-40GB GPUs in about 24 days\u201d, \"All models are trained with FP16 mixed precision.\"\n3.12e14 * 128 * 24 * 24 * 3600 * 0.3 = 2.48e22\n(FLOP/GPU-s) * (GPU) * (days) * (h/day) * (sec/h) * (utilization assumption)"},{"id":"epoch:12c974ccc3386361a9af:parameters","entityId":"epoch:12c974ccc3386361a9af","metric":"parameters","value":340000000000.0,"unit":"parameters","observationPeriod":"2023-05-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from PaLM 2"},{"id":"epoch:4363583f8c6556b094ff:parameters","entityId":"epoch:4363583f8c6556b094ff","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-05-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:4363583f8c6556b094ff:training-compute","entityId":"epoch:4363583f8c6556b094ff","metric":"training-compute","value":8.95104e+19,"unit":"FLOP","observationPeriod":"2023-05-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*2131200000*7000000000=8.95104e+19"},{"id":"epoch:4b9c9638495196a979ba:parameters","entityId":"epoch:4b9c9638495196a979ba","metric":"parameters","value":39000000000.0,"unit":"parameters","observationPeriod":"2023-05-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:181e5dcb10386e6586a6:parameters","entityId":"epoch:181e5dcb10386e6586a6","metric":"parameters","value":210000000.0,"unit":"parameters","observationPeriod":"2023-05-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"210 million (Figure 4)"},{"id":"epoch:181e5dcb10386e6586a6:training-compute","entityId":"epoch:181e5dcb10386e6586a6","metric":"training-compute","value":1.9782e+18,"unit":"FLOP","observationPeriod":"2023-05-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"210000000*1570000000*6=1.9782e+18"},{"id":"epoch:297046b9de3649effcde:parameters","entityId":"epoch:297046b9de3649effcde","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-05-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B form 2.6"},{"id":"epoch:297046b9de3649effcde:training-compute","entityId":"epoch:297046b9de3649effcde","metric":"training-compute","value":1.94e+20,"unit":"FLOP","observationPeriod":"2023-05-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"All models are trained utilizing 16 Nvidia A100 (40G) GPUs and are completed within 1.5 days.\"\n16 * 3.12e14 * 1.5 * 24 * 3600 * 0.3 = 1.94e20"},{"id":"epoch:c6f07dd3590cff713608:parameters","entityId":"epoch:c6f07dd3590cff713608","metric":"parameters","value":650000000.0,"unit":"parameters","observationPeriod":"2023-05-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c6f07dd3590cff713608:training-compute","entityId":"epoch:c6f07dd3590cff713608","metric":"training-compute","value":2.145e+19,"unit":"FLOP","observationPeriod":"2023-05-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Using 6*N*D with 50 epochs:\n6*650000000*110000001*50=2.145e+19"},{"id":"epoch:251c8eff4b8d66ce958d:parameters","entityId":"epoch:251c8eff4b8d66ce958d","metric":"parameters","value":340000000000.0,"unit":"parameters","observationPeriod":"2023-05-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Model Architecture: \"PaLM-2 is a new state-of-the-art language model. We have small, medium, and large variants that use stacked layers based on the Transformer architecture, with varying parameters depending on model size. Further details of model size and architecture are withheld from external publication.\"\nHowever, the parameter count was leaked to CNBC: https://www.cnbc.com/2023/05/16/googles-palm-2-uses-nearly-five-times-more-text-data-than-predecessor.html"},{"id":"epoch:251c8eff4b8d66ce958d:training-compute","entityId":"epoch:251c8eff4b8d66ce958d","metric":"training-compute","value":7.34e+24,"unit":"FLOP","observationPeriod":"2023-05-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Compute Requirements \"Not reported.\"\nPaper suggests heuristic of  C=6ND. Based on 340B parameters and 3.6T tokens, training compute would be around 7.3*10^24 FLOP."},{"id":"epoch:6e9ecebd99573903eda9:parameters","entityId":"epoch:6e9ecebd99573903eda9","metric":"parameters","value":15500000000.0,"unit":"parameters","observationPeriod":"2023-05-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We trained a 15.5B parameter model\""},{"id":"epoch:6e9ecebd99573903eda9:training-compute","entityId":"epoch:6e9ecebd99573903eda9","metric":"training-compute","value":8.46e+22,"unit":"FLOP","observationPeriod":"2023-05-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"FLOP reported here, 8.46e22\nhttps://huggingface.co/bigcode/starcoder\n\n\n\"We trained our model on a GPU cluster with 512 A100 80 GB GPUs... Based on the total number of GPU hours that training took (320,256) and an average power usage of 280W per GPU... The fine-tuned model adds 3.5% of training time\"\n\n320256 * 312 tFLOP/s * 3600 * 1.035 * 0.3 (utilization assumption) = 1.12e23"},{"id":"epoch:eb0f036fa5eb3eb3dba7:parameters","entityId":"epoch:eb0f036fa5eb3eb3dba7","metric":"parameters","value":932000000.0,"unit":"parameters","observationPeriod":"2023-05-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"used ViT-Huge 630M as an image/video encoder and OpenCLIP-302m as text encoder"},{"id":"epoch:6f233752fb056315e941:parameters","entityId":"epoch:6f233752fb056315e941","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-05-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6f233752fb056315e941:training-compute","entityId":"epoch:6f233752fb056315e941","metric":"training-compute","value":4.2000000000000004e+22,"unit":"FLOP","observationPeriod":"2023-05-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"As shown in Table 3, nearly all of the training budget was spent on the base MPT-7B model, which took ~9.5 days to train on 440xA100-40GB GPUs, and cost ~$200k.\""},{"id":"epoch:a46c8b6b9b76cff4f403:parameters","entityId":"epoch:a46c8b6b9b76cff4f403","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2023-05-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.3B from section 3.2:  'This results in approximately 1.3 billion trainable parameters for the Otter model.'"},{"id":"epoch:3e63d1bc0d239c646c10:parameters","entityId":"epoch:3e63d1bc0d239c646c10","metric":"parameters","value":16000000000.0,"unit":"parameters","observationPeriod":"2023-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"16B for largest CodeGen2 model"},{"id":"epoch:159380d0a4b4c19ff51c:parameters","entityId":"epoch:159380d0a4b4c19ff51c","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-05-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:159380d0a4b4c19ff51c:training-compute","entityId":"epoch:159380d0a4b4c19ff51c","metric":"training-compute","value":7.8e+22,"unit":"FLOP","observationPeriod":"2023-05-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13b * 1T * 6 = 7.8e22"},{"id":"epoch:8bd63a00dfab3c2fc30f:parameters","entityId":"epoch:8bd63a00dfab3c2fc30f","metric":"parameters","value":1289952427.0,"unit":"parameters","observationPeriod":"2023-04-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Manually loaded model with following code snippet:\n\nfrom diffusers import DiffusionPipeline, DPMSolverMultistepScheduler\nimport torch\n\nrepo_id = \"stabilityai/stable-diffusion-2\"\npipe = DiffusionPipeline.from_pretrained(repo_id, torch_dtype=torch.float16, revision=\"fp16\")\n\nn_params_vae = sum(p.numel() for p in pipe.components['vae'].parameters())\nn_params_text_encoder = sum(p.numel() for p in pipe.components['text_encoder'].parameters())\nn_params_unet = sum(p.numel() for p in pipe.components['unet'].parameters())\nn_params = n_params_vae + n_params_text_encoder + n_params_unet\n\nprint(f\"Total number of parameters: {n_params}\")"},{"id":"epoch:8bd63a00dfab3c2fc30f:training-compute","entityId":"epoch:8bd63a00dfab3c2fc30f","metric":"training-compute","value":1.07085888e+22,"unit":"FLOP","observationPeriod":"2023-04-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"23,835 A100-hours * 3.12e14 FLOP/GPU-sec * 3600 sec/hour * 0.4 (assumed utilization) = 1.07e22\n\n\"Both phases of training ran on 128 NVIDIA A100 GPUs. The first training phase was run for 550k iterations in 1.6 days while the second phase was run for 850k iterations in 4.9 days, for a total of 20,051 A100 hours for training. In addition to the training time, we pre-computed the latents for the VAE and CLIP model to reduce training time and cost when making multiple passes over the dataset. Pre-computing the latents required an additional 3,784 A100 hours, resulting in 23,835 A100 hours in total\"\n\nUses automatic mixed precision, which uses half precision (fp16) in most layers, but fp32 in a few numerically unstable layers like normalization and softmax. Unlike original stable diffusion 2, used half precision LayerNorm and GroupNorm layers."},{"id":"epoch:909594c5a89e6c35b0b9:parameters","entityId":"epoch:909594c5a89e6c35b0b9","metric":"parameters","value":6700000000.0,"unit":"parameters","observationPeriod":"2023-04-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"This is Llama-7b's parameter count"},{"id":"epoch:909594c5a89e6c35b0b9:training-compute","entityId":"epoch:909594c5a89e6c35b0b9","metric":"training-compute","value":4.02e+22,"unit":"FLOP","observationPeriod":"2023-04-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use pre-trained LLaMA 7B [4] to initialize our model. We adopt Adam optimizer as an initial learning rate of 2 \u00d710\u22125, a maximum number of tokens 2048, and the batch size is 8 for each GPU. We train our model on 8 V100 GPUs with Deepspeed Zero-3 for 70 hours on 3 epochs\"\n\nLlama-7b was ~4e22. 8*70 V100-hours is ~2e20, so fine-tuning was <1% of base training."},{"id":"epoch:f032db05c85e9754ed64:parameters","entityId":"epoch:f032db05c85e9754ed64","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-04-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7B"},{"id":"epoch:f032db05c85e9754ed64:training-compute","entityId":"epoch:f032db05c85e9754ed64","metric":"training-compute","value":6.3e+22,"unit":"FLOP","observationPeriod":"2023-04-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 * 7B * 1.5T = 6.3e22\n\"Falcon-7B is a 7B parameters causal decoder-only model built by TII and trained on 1,500B tokens of RefinedWeb enhanced with curated corpora. It is made available under the Apache 2.0 license.\"\n\nTable 1 from https://arxiv.org/pdf/2311.16867 Falcon paper\n\n730 petaflop-days * 1e15 * 24 * 3600 = 6.3072e+22 FLOPs"},{"id":"epoch:fc386fa0b1080eb8d9b8:parameters","entityId":"epoch:fc386fa0b1080eb8d9b8","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2023-04-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:dc7fe349a757649ad00c:parameters","entityId":"epoch:dc7fe349a757649ad00c","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-04-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:dc7fe349a757649ad00c:training-compute","entityId":"epoch:dc7fe349a757649ad00c","metric":"training-compute","value":1.0699776e+23,"unit":"FLOP","observationPeriod":"2023-04-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Model was trained using Deepspeed and Megatron libraries, on 300B tokens dataset for 3 epochs, around 45 days on 512 V100. After that model was finetuned 1 epoch with sequence length 2048 around 20 days on 200 GPU A100 on additional data\"\n\n512 GPUs * 125000000000000 FLOPs/s [peak] * 45 days * 24 hours * 3600 s * 0.3 + 200 GPUs * 312000000000000 FLOPs/s [peak for fp16] * 20 days * 24 hours * 3600 s * 0.3 = 1.0699776e+23\n\nthey probably used fp16 as in their similar project: https://habr.com/ru/companies/sberdevices/articles/780334/\n\n6ND = 6*13B*300B*3 = 70200*10^18 = 7*10^24"},{"id":"epoch:54e8bf4eab44dc56bb6d:parameters","entityId":"epoch:54e8bf4eab44dc56bb6d","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-04-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B as Vicuna"},{"id":"epoch:ba2de316913873bb4aca:parameters","entityId":"epoch:ba2de316913873bb4aca","metric":"parameters","value":93000000.0,"unit":"parameters","observationPeriod":"2023-04-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"To avoid training OpenFold from scratch, we start with the AlphaFold2 2.0 (https://github.com/deepmind/alphafold/releases/tag/v2.0.0) weights provided by Deepmind\""},{"id":"epoch:ef68eb44b14b4c53d2b1:parameters","entityId":"epoch:ef68eb44b14b4c53d2b1","metric":"parameters","value":16000000000.0,"unit":"parameters","observationPeriod":"2023-04-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"16B"},{"id":"epoch:ef68eb44b14b4c53d2b1:training-compute","entityId":"epoch:ef68eb44b14b4c53d2b1","metric":"training-compute","value":6.67e+22,"unit":"FLOP","observationPeriod":"2023-04-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6.67e22 including pre-training for CodeGen:\n\n\"The base language model of MOSS-003, which was initialized with CodeGen and further pre-trained on 100B Chinese tokens and 20B English tokens. The model has seen 700B tokens during pre-training and consumed ~6.67x10^22 FLOPs in total.\"\n"},{"id":"epoch:0d21f533ee9f25f00753:parameters","entityId":"epoch:0d21f533ee9f25f00753","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2023-04-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f41df359fc5926e7d0d1:parameters","entityId":"epoch:f41df359fc5926e7d0d1","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-04-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:f41df359fc5926e7d0d1:training-compute","entityId":"epoch:f41df359fc5926e7d0d1","metric":"training-compute","value":7.8049e+22,"unit":"FLOP","observationPeriod":"2023-04-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8 * 3.12e14 * (18 * 3600) * 0.3 = 4.9e19\nnum gpus * peak flops * time *assumed utilization rate \n\"We train all models with 8\u00d7 A100s. Pretraining on CC-595K completes within 4 hours. Finetuning on Instruct-158K completes within 10 hours. Finetuning on ScienceQA completes within 4 hours.\" so 18 hours of time in total.\n\nHowever, they use Vicuna as their LLM backbone, which used 7.8e22 FLOPs in training. Total FLOPs are then 4.9e19 + 7.8e22 = 7.8049e22"},{"id":"epoch:76f58e84095f9f49df61:parameters","entityId":"epoch:76f58e84095f9f49df61","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-04-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:3b0f266295cfe5883146:parameters","entityId":"epoch:3b0f266295cfe5883146","metric":"parameters","value":986710000.0,"unit":"parameters","observationPeriod":"2023-04-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See row 2 of the csv here: https://github.com/mlfoundations/open_clip/blob/main/docs/openclip_results.csv"},{"id":"epoch:481763bcdc18fb17afbc:parameters","entityId":"epoch:481763bcdc18fb17afbc","metric":"parameters","value":300000000.0,"unit":"parameters","observationPeriod":"2023-04-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b10d1d2a2f5c7e731ed1:parameters","entityId":"epoch:b10d1d2a2f5c7e731ed1","metric":"parameters","value":1140000000.0,"unit":"parameters","observationPeriod":"2023-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.14B from https://huggingface.co/facebook/dinov2-giant"},{"id":"epoch:b10d1d2a2f5c7e731ed1:training-compute","entityId":"epoch:b10d1d2a2f5c7e731ed1","metric":"training-compute","value":7.41851136e+21,"unit":"FLOP","observationPeriod":"2023-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"table 14\n\n22016 * 3600 * 312 * 10 ** 12 * 3/10 = 7.41851136e+21\ngpu hours in seconds * flops of A100 * assumed utilization  rate"},{"id":"epoch:9e227fb55b0800d654ba:parameters","entityId":"epoch:9e227fb55b0800d654ba","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9ac3b61b50d48c4bf1ad:parameters","entityId":"epoch:9ac3b61b50d48c4bf1ad","metric":"parameters","value":12000000000.0,"unit":"parameters","observationPeriod":"2023-04-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f885124a631412a7b3e7:parameters","entityId":"epoch:f885124a631412a7b3e7","metric":"parameters","value":180000000000.0,"unit":"parameters","observationPeriod":"2023-04-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.thepaper.cn/newsDetail_forward_22639611\nTranslation:\n\"SenseTime launched the \"SenseNova\" large model system, which includes natural language generation, image generation services, pre-labeling for perception models, and model development. The \"SenseChat\" application platform, powered by a 180-billion parameter Chinese language model, supports ultra-long text comprehension and offers capabilities such as question answering, understanding, and generation in Chinese.\"\n\nLink says \"hundreds of billions\" but the more precise number above seems more credible."},{"id":"epoch:f885124a631412a7b3e7:training-compute","entityId":"epoch:f885124a631412a7b3e7","metric":"training-compute","value":3.89e+24,"unit":"FLOP","observationPeriod":"2023-04-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\u201cOver the course of five years, SenseTime has built SenseCore, a leading AI infrastructure with 27,000 GPUs, capable of delivering a total computational power of 5,000 petaflops\u201d\n\nAssuming they used this entire cluster with 30 days of training (rough average of frontier model training times since 2016), 30% utilization rate: 5000e15 * 0.3 * 30 * 24 * 60 * 60 = 3.89e24 FLOP.\n\nAssuming the model is dense and trained Chinchilla-optimal: 20 tokens/parameter * (180e9 parameters)**2 * 6 = 3.89e24 FLOP. (The two estimates match by coincidence.)\n\nThe model seems more likely than not to be dense, given that news of SenseChat 5.0 makes a point of stating its MoE architecture, whereas SenseChat 1.0 does not mention architecture.\n\nGiven uncertainties (e.g. the model is possibly MoE, could have been overtrained or undertrained, could have trained longer or shorter), likely between 1e23 and 3e25 FLOP."},{"id":"epoch:b7a1195c0d0c53cc6faf:parameters","entityId":"epoch:b7a1195c0d0c53cc6faf","metric":"parameters","value":6700000000.0,"unit":"parameters","observationPeriod":"2023-04-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6.7B"},{"id":"epoch:b7a1195c0d0c53cc6faf:training-compute","entityId":"epoch:b7a1195c0d0c53cc6faf","metric":"training-compute","value":3.00001e+21,"unit":"FLOP","observationPeriod":"2023-04-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"per table 5, required 3 zettaflop (3e21) to train.\n\nalso, \"INCODER-6.7B was trained on 248 V100 GPUs for 24 days\"\n\nhardware method: 125 trillion * 248 * 24 * 24 * 3600 * 0.3 = 2e22. suggests their utilization was quite low, or 24 days was just calendar time.\n"},{"id":"epoch:965029a6b0095cc7591d:parameters","entityId":"epoch:965029a6b0095cc7591d","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2023-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our model consists of ~1B parameters which is at least a magnitude smaller  compared to state-of-the-art pLMs.\""},{"id":"epoch:965029a6b0095cc7591d:training-compute","entityId":"epoch:965029a6b0095cc7591d","metric":"training-compute","value":2.26437e+20,"unit":"FLOP","observationPeriod":"2023-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The training stage takes several weeks on four NVIDIA A100 GPUs.\"\n\nAssumption: 3 weeks, 40% utilization rate, 78 TFLOP peak rate\n\nEstimate: (3*7*24*3600) s * 78e12 FLOP/s *4 GPU * 0.4 = 2.3e20"},{"id":"epoch:108ba06a9ec8067baa6e:parameters","entityId":"epoch:108ba06a9ec8067baa6e","metric":"parameters","value":1620000.0,"unit":"parameters","observationPeriod":"2023-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:108ba06a9ec8067baa6e:training-compute","entityId":"epoch:108ba06a9ec8067baa6e","metric":"training-compute","value":3.5382841e+16,"unit":"FLOP","observationPeriod":"2023-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Using 6ND formula with 170 epochs\n6*21413000*1620000*170=3.5382841e+16"},{"id":"epoch:6ef1bbeda267a3e4d4aa:parameters","entityId":"epoch:6ef1bbeda267a3e4d4aa","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-04-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13 billion"},{"id":"epoch:6ef1bbeda267a3e4d4aa:training-compute","entityId":"epoch:6ef1bbeda267a3e4d4aa","metric":"training-compute","value":2.3e+22,"unit":"FLOP","observationPeriod":"2023-04-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.3e22, per table 2"},{"id":"epoch:64463ad2cf03fe3bf125:parameters","entityId":"epoch:64463ad2cf03fe3bf125","metric":"parameters","value":636000000.0,"unit":"parameters","observationPeriod":"2023-04-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From Facebook website: https://segment-anything.com/\n\"How big is the model? The image encoder has 632M parameters.\nThe prompt encoder and mask decoder have 4M parameters.\""},{"id":"epoch:64463ad2cf03fe3bf125:training-compute","entityId":"epoch:64463ad2cf03fe3bf125","metric":"training-compute","value":7.8e+21,"unit":"FLOP","observationPeriod":"2023-04-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"SAM was trained on 256 A100 GPUS for 68 hours. We acknowledge the environmental impact and cost of training\nlarge scale models. The environmental impact of training the released SAM model is approximately 6963 kWh\"\n\n68*256 A100-hours = \n17408 hours * 3600 * 312 trillion * 0.4 (utilization assumption for image models)\n= 7.82e21\n\nmax A100 power is 400W. 6,963,000 watt-hours / 400 watts = 17407.5 hours (so they probably just calculated backwards from power rating, and this doesn't give any info on utilization)"},{"id":"epoch:2970c8b576752dabbb71:parameters","entityId":"epoch:2970c8b576752dabbb71","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-04-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:23966325d4d5906bc75d:parameters","entityId":"epoch:23966325d4d5906bc75d","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-04-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:34f27e37b67ed3580b75:parameters","entityId":"epoch:34f27e37b67ed3580b75","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-04-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:dc92e28ea4f016b4563e:parameters","entityId":"epoch:dc92e28ea4f016b4563e","metric":"parameters","value":12000000000.0,"unit":"parameters","observationPeriod":"2023-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 1 for non-embedding parameters"},{"id":"epoch:dc92e28ea4f016b4563e:training-compute","entityId":"epoch:dc92e28ea4f016b4563e","metric":"training-compute","value":2.1590000000001e+22,"unit":"FLOP","observationPeriod":"2023-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.wolframalpha.com/input?i=6+FLOP+*+12+billion+*+299892736000"},{"id":"epoch:e8a54c69920007826808:parameters","entityId":"epoch:e8a54c69920007826808","metric":"parameters","value":2800000000.0,"unit":"parameters","observationPeriod":"2023-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 1 for non-embedding parameters"},{"id":"epoch:e8a54c69920007826808:training-compute","entityId":"epoch:e8a54c69920007826808","metric":"training-compute","value":5.038000000001e+21,"unit":"FLOP","observationPeriod":"2023-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.wolframalpha.com/input?i=6+FLOP+*+2.8+billion+*+299892736000"},{"id":"epoch:e3da5b6c3151e4f865e9:parameters","entityId":"epoch:e3da5b6c3151e4f865e9","metric":"parameters","value":6900000000.0,"unit":"parameters","observationPeriod":"2023-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 1 for non-embedding parameters"},{"id":"epoch:e3da5b6c3151e4f865e9:training-compute","entityId":"epoch:e3da5b6c3151e4f865e9","metric":"training-compute","value":1.2420000000001e+22,"unit":"FLOP","observationPeriod":"2023-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.wolframalpha.com/input?i=6+FLOP+*+6.9+billion+*+299892736000"},{"id":"epoch:98674d464b207339d7ad:parameters","entityId":"epoch:98674d464b207339d7ad","metric":"parameters","value":160000000.0,"unit":"parameters","observationPeriod":"2023-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 1 for non-embedding parameters"},{"id":"epoch:98674d464b207339d7ad:training-compute","entityId":"epoch:98674d464b207339d7ad","metric":"training-compute","value":2.87900000001e+20,"unit":"FLOP","observationPeriod":"2023-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.wolframalpha.com/input?i=6+FLOP+*+160+million+*+299892736000\n\nTable 5 gives confirmation:\n1030 GPU-hours * 3600 sec/hr * 3.12e14 FLOP/sec * 0.3 (utilization assumption) = 3.5e20"},{"id":"epoch:a697305a198bb67918e2:parameters","entityId":"epoch:a697305a198bb67918e2","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2023-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 1 for non-embedding parameters"},{"id":"epoch:a697305a198bb67918e2:training-compute","entityId":"epoch:a697305a198bb67918e2","metric":"training-compute","value":1.799000000001e+21,"unit":"FLOP","observationPeriod":"2023-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.wolframalpha.com/input?i=6+FLOP+*+1+billion+*+299892736000"},{"id":"epoch:e72d9a787831945df2aa:parameters","entityId":"epoch:e72d9a787831945df2aa","metric":"parameters","value":1400000000.0,"unit":"parameters","observationPeriod":"2023-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 1 for non-embedding parameters"},{"id":"epoch:e72d9a787831945df2aa:training-compute","entityId":"epoch:e72d9a787831945df2aa","metric":"training-compute","value":2.5190000000001e+21,"unit":"FLOP","observationPeriod":"2023-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.wolframalpha.com/input?i=6+FLOP+*+1.4+billion+*+299892736000"},{"id":"epoch:73b0c8819712f0a267cf:parameters","entityId":"epoch:73b0c8819712f0a267cf","metric":"parameters","value":70000000.0,"unit":"parameters","observationPeriod":"2023-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 1 for non-embedding parameters"},{"id":"epoch:73b0c8819712f0a267cf:training-compute","entityId":"epoch:73b0c8819712f0a267cf","metric":"training-compute","value":1.26000000001e+20,"unit":"FLOP","observationPeriod":"2023-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.wolframalpha.com/input?i=6+FLOP+*+70+million+*+299892736000"},{"id":"epoch:61395f76ad68777c5694:parameters","entityId":"epoch:61395f76ad68777c5694","metric":"parameters","value":410000000.0,"unit":"parameters","observationPeriod":"2023-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 1 for non-embedding parameters"},{"id":"epoch:61395f76ad68777c5694:training-compute","entityId":"epoch:61395f76ad68777c5694","metric":"training-compute","value":7.37700000001e+20,"unit":"FLOP","observationPeriod":"2023-04-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.wolframalpha.com/input?i=6+FLOP+*+410+million+*+299892736000"},{"id":"epoch:dd9a0cb6d57eb767710e:parameters","entityId":"epoch:dd9a0cb6d57eb767710e","metric":"parameters","value":540000000000.0,"unit":"parameters","observationPeriod":"2023-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"540B"},{"id":"epoch:5b7f27f94b7d4f03bf07:parameters","entityId":"epoch:5b7f27f94b7d4f03bf07","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-03-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6f4bb0809b911b9aa2d9:parameters","entityId":"epoch:6f4bb0809b911b9aa2d9","metric":"parameters","value":50558868480.0,"unit":"parameters","observationPeriod":"2023-03-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6f4bb0809b911b9aa2d9:training-compute","entityId":"epoch:6f4bb0809b911b9aa2d9","metric":"training-compute","value":2.36e+23,"unit":"FLOP","observationPeriod":"2023-03-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.36e23 per Table 4\n\n(using our usual hardware method, 512 A100s over 53 days would be 512 * 312 teraFLOP/s * 53 * 24 * 3600 * 0.3 = 2.19e23)"},{"id":"epoch:b9f75c435165582123bd:parameters","entityId":"epoch:b9f75c435165582123bd","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-03-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a66bf542a216ebfa807c:parameters","entityId":"epoch:a66bf542a216ebfa807c","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2023-03-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1B"},{"id":"epoch:a66bf542a216ebfa807c:training-compute","entityId":"epoch:a66bf542a216ebfa807c","metric":"training-compute","value":9.7e+21,"unit":"FLOP","observationPeriod":"2023-03-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"finetuned on ViT-g (smaller than ViT-G with 1B params)\n\n\"It takes more than two weeks to pre-train a ViT-g model with VideoMAE\non 64 A100 GPUs\"\n\n64 * 312 trillion * 2 * 7 * 24 * 3600 * 0.4 (utilization assumption) = 9.7e21"},{"id":"epoch:9630873618c12054abda:parameters","entityId":"epoch:9630873618c12054abda","metric":"parameters","value":24000000000.0,"unit":"parameters","observationPeriod":"2023-03-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we train ERNIEViLG 2.0 and scale up the model size to 24B parameters.\""},{"id":"epoch:9630873618c12054abda:training-compute","entityId":"epoch:9630873618c12054abda","metric":"training-compute","value":4.658135e+22,"unit":"FLOP","observationPeriod":"2023-03-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000 FLOP / GPU / sec [A100 reported, bf16 assumed]* 432 hours * 3600 sec / hour * 320 GPUs * 0.3 [assumed utilization] = 4.658135e+22 FLOP"},{"id":"epoch:8161ddbfe6e91e42d715:parameters","entityId":"epoch:8161ddbfe6e91e42d715","metric":"parameters","value":400000000.0,"unit":"parameters","observationPeriod":"2023-03-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 3"},{"id":"epoch:8161ddbfe6e91e42d715:training-compute","entityId":"epoch:8161ddbfe6e91e42d715","metric":"training-compute","value":4.9467301e+21,"unit":"FLOP","observationPeriod":"2023-03-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Operation Counting: \n6ND = 6 FLOP / token / parameter*400*10^6 parameters * 6705000000000 tokens [see Dataset size notes] = 1.6092e+22 FLOP\n\nHardware:\n275000000000000 FLOP/s/GPU * 32 GPUs * 120 hours * 3600 sec / hour * 0.4 = 1.52064e+21 FLOPs\n\ngeometric mean (1.6092e+22, 1.52064e+21) = 4.9467301e+21"},{"id":"epoch:7d4d54aaf991ed3c474f:training-compute","entityId":"epoch:7d4d54aaf991ed3c474f","metric":"training-compute","value":7.6032e+19,"unit":"FLOP","observationPeriod":"2023-03-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"275000000000000 FLOP/s *48 hours *4 GPUs *3600 sec / hour *0.4 = 7.6032e+19 FLOP"},{"id":"epoch:746231bb9faf79ba0dd9:parameters","entityId":"epoch:746231bb9faf79ba0dd9","metric":"parameters","value":5000000000.0,"unit":"parameters","observationPeriod":"2023-03-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"5b (table 1(a))\n\nimage parameters: 4.4B \ntext parameters: 695M"},{"id":"epoch:746231bb9faf79ba0dd9:training-compute","entityId":"epoch:746231bb9faf79ba0dd9","metric":"training-compute","value":3.456e+22,"unit":"FLOP","observationPeriod":"2023-03-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 5*10^9 parameters * 2304000000000/2 tokens [see dataset size notes] = 3.456e+22 FLOP"},{"id":"epoch:35383cdc9f8a9f547258:parameters","entityId":"epoch:35383cdc9f8a9f547258","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-03-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:8f8c9e248fdbf4cc2b85:parameters","entityId":"epoch:8f8c9e248fdbf4cc2b85","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2023-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"CPM-Bee 10B large model training will launch on October 13, 2022, with monthly model releases.\""},{"id":"epoch:8f8c9e248fdbf4cc2b85:training-compute","entityId":"epoch:8f8c9e248fdbf4cc2b85","metric":"training-compute","value":6.012e+22,"unit":"FLOP","observationPeriod":"2023-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 10*10e9 parameters * 1.002e11 tokens [see training dataset size notes] = 6.012e22 FLOP"},{"id":"epoch:411b2232061f7d67abd1:parameters","entityId":"epoch:411b2232061f7d67abd1","metric":"parameters","value":1639000000.0,"unit":"parameters","observationPeriod":"2023-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\u201cWe used two different models, GPT-2 Medium and GPT-2 XL for our supervised fine-tuning to verify our trainer is implemented correctly\u2026 We fine-tuned the model for about 100K step with a batch size 2 and learning rate of 1e-4 with an Adam optimizer. It\u2019s equivalent of about 3 epochs of our training dataset\u201d (https://github.com/ethanyanjiali/minChatGPT/blob/main/report.pdf, page 4). GPT-2 XL has 1.639B parameters (https://github.com/ethanyanjiali/minChatGPT/blob/main/report.pdf, page 5)."},{"id":"epoch:a560066e1f42f44d6e1d:training-compute","entityId":"epoch:a560066e1f42f44d6e1d","metric":"training-compute","value":1.8710548688e+22,"unit":"FLOP","observationPeriod":"2023-03-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 1x NVIDIA GeForce RTX 3090 Ti (1.60 x 10^14 FP16 FLOP/s)\n\n2. Training duration: Directly provided - 2,200 seconds/epoch \u00d7 43 epochs = 94,600 seconds\n\n3. Utilization rate: 40%\n\n4. Final calculation: 1.60 \u00d7 10^14 FLOP/s \u00d7 94,600 seconds \u00d7 0.4 = 6.05 \u00d7 10^18 FLOPs\n\nBase model: 18704498688000000000000\nTotal: 18704498688000000000000+6.05 \u00d7 10^18=18710548688000000000000"},{"id":"epoch:5924b44d16e65e2f319c:parameters","entityId":"epoch:5924b44d16e65e2f319c","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2023-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"GPT-3 Small: 125M\nAt 90% sparsity \u2192 ~1.25B total parameters\nAt 75% sparsity \u2192 125M * ksw^2 = 125M *(1/(1-0.75)) = 500M total pramters (I assume only 125M are active) "},{"id":"epoch:5924b44d16e65e2f319c:training-compute","entityId":"epoch:5924b44d16e65e2f319c","metric":"training-compute","value":1.875e+18,"unit":"FLOP","observationPeriod":"2023-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 125* 10^6 active parameters * 2.5 * 10^9 tokens = 1.875e+18 FLOP\n\n__________________\nin the Algorithmic progress paper the estimation was 8.84 \u00d7 10^19 FLOP, they assumed WT-103 dataset (not mentioned in the paper) and different number of parameters (1.3* 10^9 - I am unsure where it comes from)"},{"id":"epoch:8d928e0d434a0341ad57:parameters","entityId":"epoch:8d928e0d434a0341ad57","metric":"parameters","value":40000000000.0,"unit":"parameters","observationPeriod":"2023-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Boasting an impressive 40 billion parameters, Mini is a formidable addition to the growing array of language models available in the market today.\""},{"id":"epoch:8d928e0d434a0341ad57:training-compute","entityId":"epoch:8d928e0d434a0341ad57","metric":"training-compute","value":2.4e+23,"unit":"FLOP","observationPeriod":"2023-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND aproximation: 6*40B*1T = 2.4e23"},{"id":"epoch:82cc3c6fb3336885302a:parameters","entityId":"epoch:82cc3c6fb3336885302a","metric":"parameters","value":1085000000000.0,"unit":"parameters","observationPeriod":"2023-03-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In this work, we present PanGu-\u03a3 , a large language model with sparse architecture containing 1.085 trillion parameters.\""},{"id":"epoch:82cc3c6fb3336885302a:training-compute","entityId":"epoch:82cc3c6fb3336885302a","metric":"training-compute","value":4.67e+23,"unit":"FLOP","observationPeriod":"2023-03-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"It has sparse architecture, so we can't use C=6ND.\n\"We develop PanGu-\u03a3 model under the framework of MindSpore and train it on a cluster with only 512 Ascend 910 AI Accelerators with 329 billion tokens over 100 days.\"\n100 days * 512 processors * 320 teraFLOPS/processor * 33% utilization = 4.67e+23 FLOP\nhttps://www.wolframalpha.com/input?i=100+days+*+512+*+320+terahertz+*+0.33"},{"id":"epoch:2958d3f9a0a1ddf77b22:parameters","entityId":"epoch:2958d3f9a0a1ddf77b22","metric":"parameters","value":1800000000000.0,"unit":"parameters","observationPeriod":"2023-03-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Rumored to be 1.8T parameter MoE with 280B activated on the forward pass, per https://www.semianalysis.com/p/gpt-4-architecture-infrastructure. Other sources estimate 1.76T with 220B per forward pass https://web.archive.org/web/20230712123915/https://the-decoder.com/gpt-4-architecture-datasets-costs-and-more-leaked/"},{"id":"epoch:2958d3f9a0a1ddf77b22:training-compute","entityId":"epoch:2958d3f9a0a1ddf77b22","metric":"training-compute","value":2.1e+25,"unit":"FLOP","observationPeriod":"2023-03-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"90% CI: 8.2E+24 to 4.4E+25\n\nNOTE: this is a rough estimate based on public information, much less information than most other systems in the database.\n\nCalculation and confidence intervals here: https://colab.research.google.com/drive/1O99z9b1I5O66bT78r9ScslE_nOj5irN9?usp=sharing"},{"id":"epoch:3440ed3662690e00b7f9:parameters","entityId":"epoch:3440ed3662690e00b7f9","metric":"parameters","value":40000000000.0,"unit":"parameters","observationPeriod":"2023-03-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Model comes in 7B and 40B variants."},{"id":"epoch:3440ed3662690e00b7f9:training-compute","entityId":"epoch:3440ed3662690e00b7f9","metric":"training-compute","value":2.4e+23,"unit":"FLOP","observationPeriod":"2023-03-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"C = 6ND = 6 * 40B * 1000B = 2.4e+23 FLOP (assuming one epoch)\n\nTable 1 from https://arxiv.org/pdf/2311.16867 Falcon paper\n\n2,800 petaflop-days * 1e15 * 24 * 3600 = 2.4192e+23 FLOPs"},{"id":"epoch:89914dc9452e4d858d5c:parameters","entityId":"epoch:89914dc9452e4d858d5c","metric":"parameters","value":3007381000.0,"unit":"parameters","observationPeriod":"2023-03-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Uses ESM-2 3B. Table 2 gives details on the non-ESM layers. The GCN appears to have about 3.31M parameters and the linear layers should have 771k and 3.3M, respectively. So total is ~3.007B"},{"id":"epoch:b39e5c28f4d57584fbe1:parameters","entityId":"epoch:b39e5c28f4d57584fbe1","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2023-03-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"hundreds of billions of parameters\""},{"id":"epoch:041b3cc333b13bed8982:parameters","entityId":"epoch:041b3cc333b13bed8982","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-03-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7b"},{"id":"epoch:219a12f95aad057ce31a:parameters","entityId":"epoch:219a12f95aad057ce31a","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-03-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Pythia-Chat-Base-7B-v0.16 is based on ElutherAI\u2019s Pythia-7B model\" (https://huggingface.co/togethercomputer/Pythia-Chat-Base-7B)."},{"id":"epoch:1389bdfa981fca92afb6:parameters","entityId":"epoch:1389bdfa981fca92afb6","metric":"parameters","value":700000000.0,"unit":"parameters","observationPeriod":"2023-03-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For the cross-lingual codec language models, \u03c6MAR and \u03c6MNAR are both 12-layer Transformer decoders with an attention dimension of 1024 and the FFN dimension of 4096.\"\n\nThese are two parts of the model.\n\nAccording to Ben's script, that's 353M parameters, or ~700M for both\n(https://github.com/bencottier/ml-parameter-count/blob/main/parameter_count.py)"},{"id":"epoch:1389bdfa981fca92afb6:training-compute","entityId":"epoch:1389bdfa981fca92afb6","metric":"training-compute","value":1.2e+21,"unit":"FLOP","observationPeriod":"2023-03-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\" The batch sizes of speech and phonemes on each GPU are 1,400,000 (87.5 seconds) and 3,000, respectively. The maximum learning\nrate is 5e-4 with warm-up steps of 32,000. The model is pre-trained on 32 V100 GPUs for 400K step\"\n\n6CN:\n6 * (1.4M + 3000) * 353M * 400k = 1.2e21\n\n(2 OOM larger than our estimate for VALL-E, and would take ~46 days on the V100 cluster at 30% utilization. Seems perhaps a bit high?)"},{"id":"epoch:f7dcf4a9e6c8bddd3ca7:parameters","entityId":"epoch:f7dcf4a9e6c8bddd3ca7","metric":"parameters","value":562000000000.0,"unit":"parameters","observationPeriod":"2023-03-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"562B"},{"id":"epoch:48b1e7eed75731e1c140:training-compute","entityId":"epoch:48b1e7eed75731e1c140","metric":"training-compute","value":5.413824e+18,"unit":"FLOP","observationPeriod":"2023-03-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"31330000000000 FLOP / sec / GPU [fp16 asssumed] * 8 GPUs * 20 hours * 3600 sec / hour * 0.3 [assumed utilization] = 5.413824e+18 FLOP"},{"id":"epoch:e5bf9f623029c9078dd4:parameters","entityId":"epoch:e5bf9f623029c9078dd4","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2023-03-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We trained two sets of ALMs, one with 285M parameters (base) and the other with 1B parameters (large).\""},{"id":"epoch:e5bf9f623029c9078dd4:training-compute","entityId":"epoch:e5bf9f623029c9078dd4","metric":"training-compute","value":9.5e+21,"unit":"FLOP","observationPeriod":"2023-03-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"the large model was trained on 128 A100 GPUs for 200k steps (\u223c1 week)\"\nA100s are 312 teraflop/s\n128 * 312 trillion * 7 * 24 * 3600 * 0.3 (utilization assumption) = 7.2e21\n\nText encoding uses T5-Large, which used 2.3e21 FLOP in pre-training per Flan paper: https://arxiv.org/abs/2210.11416 "},{"id":"epoch:bfb7978289133e6994d5:parameters","entityId":"epoch:bfb7978289133e6994d5","metric":"parameters","value":19500000000.0,"unit":"parameters","observationPeriod":"2023-03-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"19.5B per https://www.yitay.net/blog/flan-ul2-20b"},{"id":"epoch:e9599513dc001a6539f9:parameters","entityId":"epoch:e9599513dc001a6539f9","metric":"parameters","value":675000000.0,"unit":"parameters","observationPeriod":"2023-03-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"675M\n\nTable 4: \"Parameter and flop counts exclude the VAE model which contains 84M parameters across the encoder and decoder.\"\n675e6 not including VAE (likely frozen), 759e6 including VAE"},{"id":"epoch:e9599513dc001a6539f9:training-compute","entityId":"epoch:e9599513dc001a6539f9","metric":"training-compute","value":6e+20,"unit":"FLOP","observationPeriod":"2023-03-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"~6e20, based on eyeballing Figure 9. It's between 1e11 and 1e12 gigaflop (1 gigaflop = 1e9 flop), and about 80% of the way towards 1e12 on a log scale. 10^0.8 is about 6. \n\n3M iterations with a batch size of 256.\n\n\"Compute. We implement all models in JAX [1] and train\nthem using TPU-v3 pods. DiT-XL/2, our most computeintensive model, trains at roughly 5.7 iterations/second on a\nTPU v3-256 pod with a global batch size of 256\"\n256*123000000000000 FLOPs/s * 800000 training steps / 5.7 iterations/second * 0.3 = 1.3258105e+21"},{"id":"epoch:13a39470ef3db60b2765:parameters","entityId":"epoch:13a39470ef3db60b2765","metric":"parameters","value":52000000000.0,"unit":"parameters","observationPeriod":"2023-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"52B for larger version\n\nhttps://aws.amazon.com/bedrock/cohere-command-embed/\n\nCohere Command has had a few different sizes over time and is continuously updated, but there's been a 52B version since at least March 2023: https://twitter.com/percyliang/status/1638236921754443776"},{"id":"epoch:03a19848367ea80b3f37:parameters","entityId":"epoch:03a19848367ea80b3f37","metric":"parameters","value":20000000000.0,"unit":"parameters","observationPeriod":"2023-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"20B parameters for Palmyra Large. There is also a 43B version called Palmyra X according to HELM."},{"id":"epoch:03a19848367ea80b3f37:training-compute","entityId":"epoch:03a19848367ea80b3f37","metric":"training-compute","value":9.6e+22,"unit":"FLOP","observationPeriod":"2023-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Palmyra-Large is a 20B parameters causal decoder-only model built by Writer and trained on +800B tokens of Palmyra-Index-Data enhanced with curated corpora.\"\n\nI'm not sure if the 800B is how many tokens the model was trained on, or the size of the dataset. But the dataset linked on HuggingFace has 1T tokens, so 800B as tokens trained is more likely.\n\n20B*800B*6 = 9.6e22"},{"id":"epoch:12b3d2dd8afbd0c82967:parameters","entityId":"epoch:12b3d2dd8afbd0c82967","metric":"parameters","value":178000000000.0,"unit":"parameters","observationPeriod":"2023-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Source is https://crfm.stanford.edu/helm/latest/#/leaderboard as viewed on 2023-12-06"},{"id":"epoch:2578ab7061a08c1dcb72:parameters","entityId":"epoch:2578ab7061a08c1dcb72","metric":"parameters","value":40000000000.0,"unit":"parameters","observationPeriod":"2023-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"40B"},{"id":"epoch:2578ab7061a08c1dcb72:training-compute","entityId":"epoch:2578ab7061a08c1dcb72","metric":"training-compute","value":7.68e+22,"unit":"FLOP","observationPeriod":"2023-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"aproximation 6ND = 6*320E9*40e9 = 7.68e22\n\"GPT-SW3 has been trained on a dataset containing 320B tokens in Swedish, Norwegian, Danish, Icelandic, English, and programming code.\""},{"id":"epoch:feaab2441ebeef4aa5cd:parameters","entityId":"epoch:feaab2441ebeef4aa5cd","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2023-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ba48a2bbe0d23d6e6f32:parameters","entityId":"epoch:ba48a2bbe0d23d6e6f32","metric":"parameters","value":1600000000.0,"unit":"parameters","observationPeriod":"2023-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The total number of parameters of KOSMOS-1 is about 1.6B\"\n\"The MLLM component has 24 layers with 2,048 hidden dimensions, 8,192 FFN intermediate size, and 32 attention heads, resulting in about 1.3B parameters.\""},{"id":"epoch:ba48a2bbe0d23d6e6f32:training-compute","entityId":"epoch:ba48a2bbe0d23d6e6f32","metric":"training-compute","value":3.456e+21,"unit":"FLOP","observationPeriod":"2023-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 1.6 * 10^9 parameters [assuming all parameters were updated each step including vision encoder] * 360 * 10^9 tokens = 3.456e+21 FLOP\n\n\"Likely\" confidence because the number of trained parameters could be slightly less"},{"id":"epoch:60cf6ec2813fedc7c2ee:parameters","entityId":"epoch:60cf6ec2813fedc7c2ee","metric":"parameters","value":16100000000.0,"unit":"parameters","observationPeriod":"2023-02-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"16.1B parameters"},{"id":"epoch:ebaf735a206d666db461:parameters","entityId":"epoch:ebaf735a206d666db461","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-02-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13.0B"},{"id":"epoch:ebaf735a206d666db461:training-compute","entityId":"epoch:ebaf735a206d666db461","metric":"training-compute","value":7.8e+22,"unit":"FLOP","observationPeriod":"2023-02-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1T tokens * 13B parameters * 6 FLOP/token/parameter = 7.8e22\n\nfrom paper, Llama-13B took 135,168 GPU hours using A100s\n\n312 trillion * 135,168 * 3600 = 1.518e23 FLOPs at full utilization\nThis implies that the actual utilization was:\nMFU = 7.8e22/1.518e23 = 0.514"},{"id":"epoch:d48b540948ec8568c817:parameters","entityId":"epoch:d48b540948ec8568c817","metric":"parameters","value":32500000000.0,"unit":"parameters","observationPeriod":"2023-02-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2 in the paper"},{"id":"epoch:d48b540948ec8568c817:training-compute","entityId":"epoch:d48b540948ec8568c817","metric":"training-compute","value":2.7300000000001e+23,"unit":"FLOP","observationPeriod":"2023-02-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.4T tokens * 32.5B params * 6 FLOP/token/param = 2.73e+23 FLOP"},{"id":"epoch:7e9f8b7b78339843ee41:parameters","entityId":"epoch:7e9f8b7b78339843ee41","metric":"parameters","value":65200000000.0,"unit":"parameters","observationPeriod":"2023-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Model card, table 1: https://github.com/facebookresearch/llama/blob/53011c3d7946dadb8274a4c5c7586ab54edf792d/MODEL_CARD.md"},{"id":"epoch:7e9f8b7b78339843ee41:training-compute","entityId":"epoch:7e9f8b7b78339843ee41","metric":"training-compute","value":5.5e+23,"unit":"FLOP","observationPeriod":"2023-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.4e12 tokens * 6.52e10 parameters * 6 FLOP/token/parameter = 5.5e23 FLOP"},{"id":"epoch:27caf645f7347afb9a6e:parameters","entityId":"epoch:27caf645f7347afb9a6e","metric":"parameters","value":6700000000.0,"unit":"parameters","observationPeriod":"2023-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6.7B parameters, per table 2: https://arxiv.org/pdf/2302.13971.pdf"},{"id":"epoch:27caf645f7347afb9a6e:training-compute","entityId":"epoch:27caf645f7347afb9a6e","metric":"training-compute","value":4.00000001e+22,"unit":"FLOP","observationPeriod":"2023-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1T tokens * 6.7B parameters * 6 FLOP/token/parameter = 4e22 FLOP\n"},{"id":"epoch:7a3fb47a7713a652c324:parameters","entityId":"epoch:7a3fb47a7713a652c324","metric":"parameters","value":125000000.0,"unit":"parameters","observationPeriod":"2023-02-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:73fe53d51d569ea73f2d:parameters","entityId":"epoch:73fe53d51d569ea73f2d","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2023-02-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:73fe53d51d569ea73f2d:training-compute","entityId":"epoch:73fe53d51d569ea73f2d","metric":"training-compute","value":4.76e+19,"unit":"FLOP","observationPeriod":"2023-02-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.3e19 for the 355M version at 5B tokens. Scaling linearly to 1.3B parameters would be 4.76e19 FLOP"},{"id":"epoch:0a035418354462d8b26f:parameters","entityId":"epoch:0a035418354462d8b26f","metric":"parameters","value":355000000.0,"unit":"parameters","observationPeriod":"2023-02-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0a035418354462d8b26f:training-compute","entityId":"epoch:0a035418354462d8b26f","metric":"training-compute","value":3.93e+19,"unit":"FLOP","observationPeriod":"2023-02-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 3: 3.93*10^19 FLOP\n\n6 FLOP / parameter / token * 355 * 10^6 parameters * 15 * 10^9 tokens = 3.195e+19 FLOP"},{"id":"epoch:7638b26b4298914eab67:parameters","entityId":"epoch:7638b26b4298914eab67","metric":"parameters","value":153000000.0,"unit":"parameters","observationPeriod":"2023-02-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:7638b26b4298914eab67:training-compute","entityId":"epoch:7638b26b4298914eab67","metric":"training-compute","value":1.87e+19,"unit":"FLOP","observationPeriod":"2023-02-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 3: 1.87*10^19 FLOP\n\n6 FLOP / parameter / token * 153 * 10^6 parameters * 15 * 10^9 tokens = 1.377e+19 FLOP"},{"id":"epoch:86c2da44dd3d6d056fc8:parameters","entityId":"epoch:86c2da44dd3d6d056fc8","metric":"parameters","value":243000000.00000003,"unit":"parameters","observationPeriod":"2023-02-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6a8e71de486a08dd83b7:parameters","entityId":"epoch:6a8e71de486a08dd83b7","metric":"parameters","value":175000000000.0,"unit":"parameters","observationPeriod":"2023-02-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"175B"},{"id":"epoch:24ced54a8a474da336cf:parameters","entityId":"epoch:24ced54a8a474da336cf","metric":"parameters","value":3070000000.0,"unit":"parameters","observationPeriod":"2023-02-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"parameter count of original BASIC-L"},{"id":"epoch:bdc82c61fc3a79e259d7:parameters","entityId":"epoch:bdc82c61fc3a79e259d7","metric":"parameters","value":21743000000.0,"unit":"parameters","observationPeriod":"2023-02-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"21.743B, Table 1"},{"id":"epoch:bdc82c61fc3a79e259d7:training-compute","entityId":"epoch:bdc82c61fc3a79e259d7","metric":"training-compute","value":1.93248e+23,"unit":"FLOP","observationPeriod":"2023-02-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"ViT-22B was trained using 256 visual tokens per image, where each token represents a 14 \u00d7 14 patch extracted from 224 \u00d7 224 sized images. ViT-22B is trained for 177k steps with batch size of 65k: approximately 3 epochs\"\n\n\"ViT-22B was trained on 1024 TPU V4 chips for 177K steps\"\n\n\"Using these techniques, ViT-22B processes 1.15k tokens per second per core during training (forward and backward pass) on TPUv4 [...] ViT-22B\u2019s model flops utilization (MFU) is 54.9%\"\n\n256 * 177k * 65k = 2.945T tokens\n\nSo training time is 2.945T tokens / (1.15k * 2 * 1024) tokens/s = 1.25M seconds = 347.4 hours\n\nSo 1024 TPUv4 chips for 1.25M seconds at 54.9% MFU:\n1024 * 2.75e14 * 1.25M * 0.549 = 1.93248e23"},{"id":"epoch:0c984351118aa63b9e5b:parameters","entityId":"epoch:0c984351118aa63b9e5b","metric":"parameters","value":6700000000.0,"unit":"parameters","observationPeriod":"2023-02-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Fine-tune of GPT-J-6B"},{"id":"epoch:c98910d886196ae6299b:training-compute","entityId":"epoch:c98910d886196ae6299b","metric":"training-compute","value":3.024e+19,"unit":"FLOP","observationPeriod":"2023-02-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The model is trained with a single NVIDIA V100 GPU using a batch size of 8 and learning rate 1\u00d710\u22124 for 2 million iterations, which consumes approximately 7 days.\" Assume FP16 precision and 40% utilization. \n1.3e+14*0.4*604800s=3.0e+19"},{"id":"epoch:d8ef69a1b4fdfdd51b97:parameters","entityId":"epoch:d8ef69a1b4fdfdd51b97","metric":"parameters","value":6846080.0,"unit":"parameters","observationPeriod":"2023-02-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"They provide the model initialization code in Appendix D. Running it  gives the parameter count."},{"id":"epoch:39b287947934302e3f4d:parameters","entityId":"epoch:39b287947934302e3f4d","metric":"parameters","value":12100000000.0,"unit":"parameters","observationPeriod":"2023-01-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12.1B, per Table 2. \n\nonly 108M trainable params (i.e. params trained during the BLIP process)"},{"id":"epoch:000a247c9b6e909358ee:parameters","entityId":"epoch:000a247c9b6e909358ee","metric":"parameters","value":1480000000.0,"unit":"parameters","observationPeriod":"2023-01-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Q-Former has 188M params. The BLIP-2 system overall has \"54x fewer trainable parameters\" than Flamingo80B."},{"id":"epoch:000a247c9b6e909358ee:training-compute","entityId":"epoch:000a247c9b6e909358ee","metric":"training-compute","value":1.20000000001e+21,"unit":"FLOP","observationPeriod":"2023-01-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.wolframalpha.com/input?i=312+teraFLOPS+*+16+*+200+hours+*+0.33"},{"id":"epoch:19a9e602185199a4d016:parameters","entityId":"epoch:19a9e602185199a4d016","metric":"parameters","value":4100000.0,"unit":"parameters","observationPeriod":"2023-01-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"RFDiffusion also contains around 14 times more parameters than Genie (59.8M versus 4.1M).\""},{"id":"epoch:19a9e602185199a4d016:training-compute","entityId":"epoch:19a9e602185199a4d016","metric":"training-compute","value":1.81149696e+21,"unit":"FLOP","observationPeriod":"2023-01-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For Genie-SCOPe, we train the model using data parallelism on 12 A100\nNvidia GPUs with an effective batch size of 48. \"\n12*3.1e+14*0.4*1209600s=1.8e+21"},{"id":"epoch:d2181542376161c1406c:training-compute","entityId":"epoch:d2181542376161c1406c","metric":"training-compute","value":1.4000000000000119e+19,"unit":"FLOP","observationPeriod":"2023-01-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 4x NVIDIA Tesla V100 GPUs (1.30 x 10^14 FLOP/s per GPU)\n\n2. Training duration: 69,131 seconds (\u224819.2 hours) - estimated from 20 epochs, 553,052 samples, batch size 16, assuming 0.1s per step\n\n3. Utilization rate: 40%\n\n4. Final calculation:\n1.30 x 10^14 FLOP/s \u00d7 4 GPUs \u00d7 69,131s \u00d7 0.4 = 1.44 x 10^19 FLOPs"},{"id":"epoch:f1f3807bf013c264ab7b:parameters","entityId":"epoch:f1f3807bf013c264ab7b","metric":"parameters","value":295000000.0,"unit":"parameters","observationPeriod":"2023-01-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"295M for CelebA model, per Table 9"},{"id":"epoch:f1f3807bf013c264ab7b:training-compute","entityId":"epoch:f1f3807bf013c264ab7b","metric":"training-compute","value":3.5e+20,"unit":"FLOP","observationPeriod":"2023-01-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use Pytorch 1.8 (Paszke et al., 2019) and trained all the models on different NVIDIA Tesla V100s (16G memory). In\nmore detail, we use 2 GPUs to train the models on CIFAR10 for 2 days, and 4 GPUs to train the models on ImageNet 32\u00d732\nfor 34 days. For LSUN tower 64\u00d764, CelebA 64\u00d764 and FFHQ 128\u00d7128, we used 16 GPUs to train the models for 3 days,\n5 days and 4 days, respectively\"\n\n5*16 V100-days for CelebA.\n\n5 * 16 * 24 * 3600 * 125 teraflops * 0.4 ~= 3.5e20"},{"id":"epoch:fd4cf9ad929269fe9ac9:parameters","entityId":"epoch:fd4cf9ad929269fe9ac9","metric":"parameters","value":860000000.0,"unit":"parameters","observationPeriod":"2023-01-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use decoder-only Transformers for modeling the semantic stage and the acoustic stages of AudioLM. The models\nshare the same architecture, composed of 24 layers, 16 attention heads, an embedding dimension of 1024, feed-forward\nlayers of dimensionality 4096, dropout of 0.1, and relative\npositional embeddings (Raffel et al., 2020), resulting in\n430M parameters per stage.\"\n\n\"stage\" seems to mean semantic + acoustic, so 860M total"},{"id":"epoch:2821dc129665df6cf790:training-compute","entityId":"epoch:2821dc129665df6cf790","metric":"training-compute","value":4.509e+20,"unit":"FLOP","observationPeriod":"2023-01-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"125000000000000 FLOP / GPU / sec *(208 hours * 16 GPUs +1 GPU * 12 hours) [see training time notes] * 3600 sec / hour * 0.3 [assumed utilization] = 4.509e+20 FLOP"},{"id":"epoch:6a8bf224f64d5dd3d06a:parameters","entityId":"epoch:6a8bf224f64d5dd3d06a","metric":"parameters","value":3500000000.0,"unit":"parameters","observationPeriod":"2023-01-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"exceeds 3.5 billion\""},{"id":"epoch:9750f2d555aa2e6bca54:parameters","entityId":"epoch:9750f2d555aa2e6bca54","metric":"parameters","value":124000000.0,"unit":"parameters","observationPeriod":"2023-01-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9750f2d555aa2e6bca54:training-compute","entityId":"epoch:9750f2d555aa2e6bca54","metric":"training-compute","value":2.976e+17,"unit":"FLOP","observationPeriod":"2023-01-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 124000000.00 parameters * 2000000 tokens * 200 epochs [assumed: not reported for WT2, but reported for WT103 minigpt2 training -> \"Likely\" confidence] = 2.976e+17 FLOP\n\n"},{"id":"epoch:a80292738f2f404e72d5:parameters","entityId":"epoch:a80292738f2f404e72d5","metric":"parameters","value":533000000.0,"unit":"parameters","observationPeriod":"2023-01-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table D.9"},{"id":"epoch:a80292738f2f404e72d5:training-compute","entityId":"epoch:a80292738f2f404e72d5","metric":"training-compute","value":2.8e+21,"unit":"FLOP","observationPeriod":"2023-01-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"AdA was implemented using JAX (Bradbury et al., 2018) and the DeepMind JAX Ecosystem (Babuschkin et al., 2020) and trained on 64 Google TPUv3 devices. The wall-clock time for training this version of AdA from scratch was approximately 5 weeks: 1 week to train the teacher, and 4 weeks to train AdA\"\n\n64 * 123 teraflop/s * 35 days * 24 * 3600 * 0.4 = 9.5e21\n\nThis might be for all single-agent experiments in the paper, or just for the 76M model in Table D.1, I'm not sure.\n\nIn Table E.2, the 533M-param model takes 2e20 FLOP to go through 5B learner steps, and was trained on 70B steps in total (Table 1). That would be 2.8e21 for 70B steps. That might be an underestimate because there are also teacher(?) steps."},{"id":"epoch:6125d25fb6f2ca2fe42e:parameters","entityId":"epoch:6125d25fb6f2ca2fe42e","metric":"parameters","value":1900000000.0,"unit":"parameters","observationPeriod":"2023-01-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 1 indicates 1.15B parameters, but both the huggingface model and a replication (https://huggingface.co/ElnaggarLab/ankh-large and https://www.biorxiv.org/content/10.1101/2023.07.05.547496v1.full.pdf) indicate 1.9B parameters.\nNotebook for counting params: https://colab.research.google.com/drive/1EGI5_vDl4pOBUukJexMHQR16BFKJe4a5?usp=sharing"},{"id":"epoch:6125d25fb6f2ca2fe42e:training-compute","entityId":"epoch:6125d25fb6f2ca2fe42e","metric":"training-compute","value":6.5e+21,"unit":"FLOP","observationPeriod":"2023-01-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 9 from here: https://www.biorxiv.org/content/10.1101/2023.07.05.547496v1.full.pdf\n\nCan also be manually estimated based on the details in Table 11 and 4.6.1 Exp 4. 14B residues * 68 epochs = 952B tokens seen in forward passes. However, only 20% of tokens are masked as individual targets; other tokens in consecutive spans are collapsed into single-token targets to reduce computations. For masking rate of 20%, the average sequence will have 36% as many targets as input tokens under this strategy. This is the relevant number of backward passes:\n(2 * 952B * 19B) + (4 * 952B * 0.36 * 19B) = 6.22e22\n\n36% figure verified here: https://colab.research.google.com/drive/1ETsmp_KRMK8kIRA5kdfcO9QiPK28cBQ6?usp=sharing "},{"id":"epoch:1b3eee4e8c54f35b8a59:parameters","entityId":"epoch:1b3eee4e8c54f35b8a59","metric":"parameters","value":740000000.0,"unit":"parameters","observationPeriod":"2023-01-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 1 indicates 450M, but the model on huggingface https://huggingface.co/ElnaggarLab/ankh-base as well as Table 9 from https://www.biorxiv.org/content/10.1101/2023.07.05.547496v1.full.pdf \neach indicate 740M parameters.\nNotebook for counting params: https://colab.research.google.com/drive/1EGI5_vDl4pOBUukJexMHQR16BFKJe4a5?usp=sharing"},{"id":"epoch:1b3eee4e8c54f35b8a59:training-compute","entityId":"epoch:1b3eee4e8c54f35b8a59","metric":"training-compute","value":2.6e+21,"unit":"FLOP","observationPeriod":"2023-01-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 9 from here: https://www.biorxiv.org/content/10.1101/2023.07.05.547496v1.full.pdf"},{"id":"epoch:81dbf71ff2f79eb7f6de:parameters","entityId":"epoch:81dbf71ff2f79eb7f6de","metric":"parameters","value":2500000000.0,"unit":"parameters","observationPeriod":"2023-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We built four distinct foundation language models of different sizes, ranging from 500M up to 2.5B parameters\""},{"id":"epoch:81dbf71ff2f79eb7f6de:training-compute","entityId":"epoch:81dbf71ff2f79eb7f6de","metric":"training-compute","value":8.08e+21,"unit":"FLOP","observationPeriod":"2023-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Training the largest parameter model required a total of 128 GPUs across 16 compute nodes for 28 days\"\n\nIn repo, they default to jnp.float32, but recommend fp16 or bp16 for activations in the docstring. JAX defaults to TF32 so they should be utilizing tensor cores.\n\nAssuming 1.56e14 FLOP/s for 32-bit calculations and 0.3 utilization rate\n\nEstimate: 1.56e14 FLOP/s * 128 GPUs * 28 days * 24 h/day * 3600 s/h * 0.3 utilization rate = 1.45e22\n\nOr, with 6ND:\n\"the model processed a total of 300B tokens during training\"\n6 * 300B * 2.5B = 4.5e21\n\nGeometric mean: sqrt(1.45e22 * 4.5e21) = 8.08e21"},{"id":"epoch:69bfe4f9130be6b88680:parameters","entityId":"epoch:69bfe4f9130be6b88680","metric":"parameters","value":200000000.0,"unit":"parameters","observationPeriod":"2023-01-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table B1"},{"id":"epoch:69bfe4f9130be6b88680:training-compute","entityId":"epoch:69bfe4f9130be6b88680","metric":"training-compute","value":2.2032e+20,"unit":"FLOP","observationPeriod":"2023-01-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"16 environment instances, each with 1 V100 running for 17 days (table A1) - it's not entirely clear if the GPU days already account for multiple environment instances.\n\nAssuming no:\nCompute: 17*24*60*60*125000000000000*0.3=5.508e+19\nAssuming yes:\nCompute: 17*24*60*60*16*125000000000000*0.3=8.8128e+20\n\nGeometric mean: 220320000000000000000"},{"id":"epoch:7023650db951ad7309a2:parameters","entityId":"epoch:7023650db951ad7309a2","metric":"parameters","value":1100000000.0,"unit":"parameters","observationPeriod":"2023-01-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.1B"},{"id":"epoch:7023650db951ad7309a2:training-compute","entityId":"epoch:7023650db951ad7309a2","metric":"training-compute","value":2.1e+21,"unit":"FLOP","observationPeriod":"2023-01-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Each training run takes 3.1 days to complete on 96 Tesla V100 GPUs for a total of 1.05 \u00d7 10^21 FLOPs. The final model described in Section 6.2 uses twice the amount of compute."},{"id":"epoch:5e3c6187e82022d188ff:parameters","entityId":"epoch:5e3c6187e82022d188ff","metric":"parameters","value":353000000.0,"unit":"parameters","observationPeriod":"2023-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Both the AR model and the NAR model have the same transformer architecture with 12\nlayers, 16 attention heads, an embedding dimension of 1024, a feed-forward layer dimension of 4096, and a dropout of 0.1\"\n\nBen's script says that's 353M parameters, using n_block 12, d_model 1024, d_ff 4096, encoder only False\n\nhttps://github.com/bencottier/ml-parameter-count/blob/main/parameter_count.py"},{"id":"epoch:5e3c6187e82022d188ff:training-compute","entityId":"epoch:5e3c6187e82022d188ff","metric":"training-compute","value":1.01e+19,"unit":"FLOP","observationPeriod":"2023-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The models are trained using 16 NVIDIA TESLA V100 32GB GPUs with a batch size of 6k acoustic\ntokens per GPU for 800k steps\"\n\n353M * 800k * 6k * 6 = 1.01e19\n\n16 V100s is 2080 teraFLOP or 2e15 FLOP so 1e19 would take 1.5 hours at 100% utilization or ~5 hours at 30%. Is that plausible?"},{"id":"epoch:715fd5655e87dde49d04:parameters","entityId":"epoch:715fd5655e87dde49d04","metric":"parameters","value":87500000000.0,"unit":"parameters","observationPeriod":"2023-01-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:715fd5655e87dde49d04:training-compute","entityId":"epoch:715fd5655e87dde49d04","metric":"training-compute","value":1.58e+23,"unit":"FLOP","observationPeriod":"2023-01-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:546645a51bbce9def3cc:parameters","entityId":"epoch:546645a51bbce9def3cc","metric":"parameters","value":66000000000.0,"unit":"parameters","observationPeriod":"2023-01-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:451d71ab816b0f929229:parameters","entityId":"epoch:451d71ab816b0f929229","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2023-01-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:8e7f8acf4190e92f1eeb:parameters","entityId":"epoch:8e7f8acf4190e92f1eeb","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2023-01-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:02d17a24d4f3bfff11d5:training-compute","entityId":"epoch:02d17a24d4f3bfff11d5","metric":"training-compute","value":3.5e+18,"unit":"FLOP","observationPeriod":"2023-01-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The pre-training of Human genome language model is\nresource-intensive (about 7-10 days on 2 NVIDIA TITAN X GPU).\"\n\nAssuming FP32 and 30% utilization\n\nEstimate: (10*24*3600) s * 6.7e12 FLOP/s * 2 * 0.3 = 3.5e18"},{"id":"epoch:c3b3d7a6874142247d42:parameters","entityId":"epoch:c3b3d7a6874142247d42","metric":"parameters","value":2700000000.0,"unit":"parameters","observationPeriod":"2022-12-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.7B"},{"id":"epoch:c3b3d7a6874142247d42:training-compute","entityId":"epoch:c3b3d7a6874142247d42","metric":"training-compute","value":6.48e+21,"unit":"FLOP","observationPeriod":"2022-12-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/token/parameter * 400000000000 training tokens * 2700000000 parameters = 6.48e+21 FLOP\n\n___________________\nin the algorithmic progress paper the estimation was 8.49 \u00d7 10^20 based on the assumption of WT-103 dataset and 509 epochs"},{"id":"epoch:23d821c7e55701e3ebd2:parameters","entityId":"epoch:23d821c7e55701e3ebd2","metric":"parameters","value":125000000.0,"unit":"parameters","observationPeriod":"2022-12-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b20fac4e5673b80208e8:parameters","entityId":"epoch:b20fac4e5673b80208e8","metric":"parameters","value":355000000.0,"unit":"parameters","observationPeriod":"2022-12-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:7f43478679e5e76e07ea:parameters","entityId":"epoch:7f43478679e5e76e07ea","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2022-12-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:063ffdecb28b89c3f1ba:parameters","entityId":"epoch:063ffdecb28b89c3f1ba","metric":"parameters","value":175000000000.0,"unit":"parameters","observationPeriod":"2022-12-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:063ffdecb28b89c3f1ba:training-compute","entityId":"epoch:063ffdecb28b89c3f1ba","metric":"training-compute","value":4.3e+23,"unit":"FLOP","observationPeriod":"2022-12-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"fine-tuned from OPT-175B (4.3e23) with an estimate 2.1e21 FLOP for fine-tuning. \n\"During fine-tuning, our models saw approximately 2 billion tokens, which is only 0.6% of the pre-training budget of OPT\""},{"id":"epoch:2db664e4bb357a7148c5:parameters","entityId":"epoch:2db664e4bb357a7148c5","metric":"parameters","value":86000000.0,"unit":"parameters","observationPeriod":"2022-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We trained a large language model with 86M parameters\""},{"id":"epoch:2db664e4bb357a7148c5:training-compute","entityId":"epoch:2db664e4bb357a7148c5","metric":"training-compute","value":2.9e+19,"unit":"FLOP","observationPeriod":"2022-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"4 NVIDIA Quadro RTX4000 GPUs for 40 days\"\n\nCalculation assuming FP32, utilization 30%:\n= (40 * 24 * 3600) s * 7.1e12 FLOP/s * 0.3 * 4 GPU = 2.999808e+19\n\nalternative calculation:\n\"Gradients were accumulated to an effective batch size of 1,000 examples, or approximately 256,000 tokens. \"\n\"(66,000 gradient steps, 14 full epochs)\"\n\n256000*66000*14*86000000*6=1.220567e+20"},{"id":"epoch:9103dae4a4df04ebabd2:parameters","entityId":"epoch:9103dae4a4df04ebabd2","metric":"parameters","value":35000000.0,"unit":"parameters","observationPeriod":"2022-12-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we also limit the size of the model compared to\nthe original publication, which was 1.2B parameters (resulting in on robot inference time of 1.9s),\nto be of similar size to RT-1 (37M parameters for Gato vs. 35M for RT-1\"\n\n16M params for image tokenizer, 19M for the transformer"},{"id":"epoch:eaa8a96e87efc6ad58da:training-compute","entityId":"epoch:eaa8a96e87efc6ad58da","metric":"training-compute","value":6.7392e+22,"unit":"FLOP","observationPeriod":"2022-12-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000 FLOP / GPU / sec [A100 reported, bf16 assumed] * 200000 GPU-hours * 3600 sec / hour * 0.3 [assumed utilization] = 6.7392e+22 FLOP"},{"id":"epoch:f491046ed9555f9d13d8:parameters","entityId":"epoch:f491046ed9555f9d13d8","metric":"parameters","value":256999999.99999997,"unit":"parameters","observationPeriod":"2022-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For Transformer-XL,\nwe use the large model (257M parameters).\""},{"id":"epoch:25dae3f8de98cdc2cb1b:parameters","entityId":"epoch:25dae3f8de98cdc2cb1b","metric":"parameters","value":1550000000.0,"unit":"parameters","observationPeriod":"2022-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1550M"},{"id":"epoch:25dae3f8de98cdc2cb1b:training-compute","entityId":"epoch:25dae3f8de98cdc2cb1b","metric":"training-compute","value":1.1e+23,"unit":"FLOP","observationPeriod":"2022-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Compared to the Whisper large model, the large-v2 model is trained for 2.5x more epochs with added regularization for improved performance.\"\n\nWe (roughly) estimated Whisper v1 as 4.65e22. 2.5x that is 1.16e23 or ~1.1e23"},{"id":"epoch:c703bcf164b8a9a21c95:parameters","entityId":"epoch:c703bcf164b8a9a21c95","metric":"parameters","value":6000000000.0,"unit":"parameters","observationPeriod":"2022-12-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c703bcf164b8a9a21c95:training-compute","entityId":"epoch:c703bcf164b8a9a21c95","metric":"training-compute","value":7.76e+22,"unit":"FLOP","observationPeriod":"2022-12-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Pretraining took 1 month on 320 A100 (Section 3.1)\n720*60*60*320*312000000000000*0.3=7.7635584e+22"},{"id":"epoch:06ae871037a94514f103:parameters","entityId":"epoch:06ae871037a94514f103","metric":"parameters","value":185200000.0,"unit":"parameters","observationPeriod":"2022-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"185.2M (Table 4)\n\n\"We implemented our methods based on fairseq (Ott et al. 2019). The original transformer model follows the architectural choice described in Baevski and Auli (2018), which includes 16 decoder blocks and sinusoidal position embeddings in the input layer. Each MHSA module has 8 heads and adaptive input representations have three bands of size 20K, 40K and 200K. The embedding layer and FFN\u2019s hidden-state have dimensions of 1024 and\n4096, respectively. We sampled 4K sentences to form the proxy dataset D. Then, we reduced the parameters by 15%, 20% and 25%.\""},{"id":"epoch:06ae871037a94514f103:training-compute","entityId":"epoch:06ae871037a94514f103","metric":"training-compute","value":7.7369416e+18,"unit":"FLOP","observationPeriod":"2022-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.30e18 FLOP [base transformer] + 4.3694162e+17 FLOP [GFM] = 7.7369416e+18 FLOP\n\n_______________\nEstimations from the Algorithmic progress paper (upd - a100 gpu were assumed while the paper reports 3090 gpus): \n\nSOURCE: Compression of Baevski et al. transformer, impute\n\nDuring the GFM process, we removed layer dropout and\ntrained on 8 GPUs. We limited the number of tokens per\nGPU to a maximum threshold 1536, which means each GPU\nprocesses 1536 tokens using the same model parameters. We\naccumulated gradient updates over 8 batches before committing a parameter update following Ott et al. (2018). We set\nthe batch size as 32 and trained 1000 updates. \n\nEstimate based on Baevski: \nThey had 7.30E+18 FLOP, this has an additional training process using GFM\n1000 steps, with approx 1 sec/step = 1000s\n\nhttps://epochai.org/blog/estimating-training-compute\n0.277777 hours, 8 GPUs, NVIDIA A100 (guess), FP16, 30% = 7.4e17 FLOP\n\n7.30e18 + 7.4e17 = 8039999999999999000.00 FLOP "},{"id":"epoch:e1b318ad26dd9b9a4360:parameters","entityId":"epoch:e1b318ad26dd9b9a4360","metric":"parameters","value":738000000.0,"unit":"parameters","observationPeriod":"2022-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"ZymCTRL contains 36 layers totalling 738M parameters\""},{"id":"epoch:e1b318ad26dd9b9a4360:training-compute","entityId":"epoch:e1b318ad26dd9b9a4360","metric":"training-compute","value":5.05e+21,"unit":"FLOP","observationPeriod":"2022-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We trained for 179,000 steps on 48 NVIDIA A100s 80GB for about 15,000 GPU hours\"\n\n15000  * 3600 * 312 teraFLOPS * 0.3 (utilization assumption) = 5.05e21"},{"id":"epoch:59066b70355c90f989c1:parameters","entityId":"epoch:59066b70355c90f989c1","metric":"parameters","value":335000000.0,"unit":"parameters","observationPeriod":"2022-11-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"335M parameters: https://github.com/FlagAI-Open/FlagAI/blob/master/examples/ALM/README.md"},{"id":"epoch:296adaffe3e5361d0a4e:training-compute","entityId":"epoch:296adaffe3e5361d0a4e","metric":"training-compute","value":2.1570000001e+20,"unit":"FLOP","observationPeriod":"2022-11-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"481 hours * 312 TFLOPS (A100) * 40% utilization"},{"id":"epoch:6c23db050f8b67f5c18a:parameters","entityId":"epoch:6c23db050f8b67f5c18a","metric":"parameters","value":160000000.0,"unit":"parameters","observationPeriod":"2022-11-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"160m"},{"id":"epoch:6c23db050f8b67f5c18a:training-compute","entityId":"epoch:6c23db050f8b67f5c18a","metric":"training-compute","value":3.456e+18,"unit":"FLOP","observationPeriod":"2022-11-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"using graph reader on Figure 2 b for the upper right point: \n0.04 PF-days * 10^15 FLOPS/s* 3600s*24h=3456000000000000000\n\ncounting operations: =160000000.00*50743870312*6=4.871411549952 \u00d7 10^19\n\n\"The computation (PF-days) is calculated as 6 \u00d7 # of parameters \u00d7 batch size \u00d7 # of training steps \u00d7 sequence length divided by one PF-day = 8.64 \u00d7 10^19. We train all models for 100,000 steps.\""},{"id":"epoch:a7ca0f8bb69515d0c1b3:parameters","entityId":"epoch:a7ca0f8bb69515d0c1b3","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2022-11-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:a7ca0f8bb69515d0c1b3:training-compute","entityId":"epoch:a7ca0f8bb69515d0c1b3","metric":"training-compute","value":5.1e+20,"unit":"FLOP","observationPeriod":"2022-11-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8 NVIDIA A100 GPUs for 8 days"},{"id":"epoch:a1e534abaef6b095c719:parameters","entityId":"epoch:a1e534abaef6b095c719","metric":"parameters","value":330000000.0,"unit":"parameters","observationPeriod":"2022-11-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"330M"},{"id":"epoch:a1e534abaef6b095c719:training-compute","entityId":"epoch:a1e534abaef6b095c719","metric":"training-compute","value":1.3e+20,"unit":"FLOP","observationPeriod":"2022-11-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The experiments were run on 8x80GB Nvidia A100s with 800GB RAM and 4x32-core CPUs, and each experiment took around 1 day for NQ and 2 days for TriviaQA with large models. Inference was run on the same system, and took 2 minutes.\"\n\n2 days * 24 * 3600 * 8 * 312 teraflop/s * 0.3 utilization = 1.3e20"},{"id":"epoch:d2e891942726bcd9119b:parameters","entityId":"epoch:d2e891942726bcd9119b","metric":"parameters","value":120000000000.0,"unit":"parameters","observationPeriod":"2022-11-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The largest 120B model we train runs on a single NVIDIA A100 node\""},{"id":"epoch:d2e891942726bcd9119b:training-compute","entityId":"epoch:d2e891942726bcd9119b","metric":"training-compute","value":3.24e+23,"unit":"FLOP","observationPeriod":"2022-11-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Authors state the model is trained on 450b tokens. Using 6 FLOP/token/parameter, this is 6*120b*450b = 3.24e23"},{"id":"epoch:6dccc88ff0b3c056f0d8:parameters","entityId":"epoch:6dccc88ff0b3c056f0d8","metric":"parameters","value":2600000000.0,"unit":"parameters","observationPeriod":"2022-11-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:bea16d135a7e4ee85876:parameters","entityId":"epoch:bea16d135a7e4ee85876","metric":"parameters","value":1011000000.0,"unit":"parameters","observationPeriod":"2022-11-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1011M from table 3"},{"id":"epoch:bea16d135a7e4ee85876:training-compute","entityId":"epoch:bea16d135a7e4ee85876","metric":"training-compute","value":1.501e+22,"unit":"FLOP","observationPeriod":"2022-11-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"flops = (128) * (3.12e14) * (14.5 * 24 * 3600) * (0.3) = 1.501e22\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)\n\nfrom Table 3, time and num gpus, GPU model is on page 4 (A100), precision is fp16 and likely utilizes tensor cores"},{"id":"epoch:2fa9f21e8deaeac0e651:parameters","entityId":"epoch:2fa9f21e8deaeac0e651","metric":"parameters","value":1080000000.0,"unit":"parameters","observationPeriod":"2022-11-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.08B, table 1"},{"id":"epoch:2fa9f21e8deaeac0e651:training-compute","entityId":"epoch:2fa9f21e8deaeac0e651","metric":"training-compute","value":2.408e+21,"unit":"FLOP","observationPeriod":"2022-11-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"InternImage-H is pre-trained on a 427 million joint dataset of public Laion-400M [61], YFCC-15M [62], and CC12M [63] for 30 epochs, and then fine-tuned the model on ImageNet-1K for 20 epochs. ImageNet-1K has 1,281,167 images.\n\nTable 2 says InternImage-H uses 188 GFLOP per forward pass at 224 resolution, and 1478 GFLOP at 640\n\nTable 7 indicates training InternImage-H was done at a scale of \"224/640\" so presumably there was pretraining at 224x224 resolution and then some fine-tuning at 640x640. It's not clear how much training was done at each resolution, but typically this is a small fraction of total training (e.g. Noisy Student finds it sufficient to train for 350 epochs at smaller resolution, and then fine-tune at the higher resolution for 1.5 epochs). We'll ignore the additional FLOPs from high resolution training.\n\nTotal training FLOPs:\n188e9 FLOP/image * (427M images * 30 epochs) + (1.281M images * 20 epochs) * 3 (additional FLOPs for backward pass) = 2.408e21"},{"id":"epoch:f712ad6b3ffc8a9c7bba:parameters","entityId":"epoch:f712ad6b3ffc8a9c7bba","metric":"parameters","value":35000000.0,"unit":"parameters","observationPeriod":"2022-11-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:f712ad6b3ffc8a9c7bba:training-compute","entityId":"epoch:f712ad6b3ffc8a9c7bba","metric":"training-compute","value":1.4e+17,"unit":"FLOP","observationPeriod":"2022-11-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*35000000*2666667*250 = 1.4000002e+17"},{"id":"epoch:1617617569438323317c:parameters","entityId":"epoch:1617617569438323317c","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2022-11-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:1617617569438323317c:training-compute","entityId":"epoch:1617617569438323317c","metric":"training-compute","value":7.1347219e+16,"unit":"FLOP","observationPeriod":"2022-11-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*24000000*1238667*400 = 7.1347219e+16"},{"id":"epoch:142a651e532613e8a003:parameters","entityId":"epoch:142a651e532613e8a003","metric":"parameters","value":176000000000.0,"unit":"parameters","observationPeriod":"2022-11-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"176B"},{"id":"epoch:012d489cd5477f24279c:parameters","entityId":"epoch:012d489cd5477f24279c","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2022-11-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:638b20a4ff6166bb4601:parameters","entityId":"epoch:638b20a4ff6166bb4601","metric":"parameters","value":9100000000.0,"unit":"parameters","observationPeriod":"2022-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"9.1B for config D, Table 1"},{"id":"epoch:638b20a4ff6166bb4601:training-compute","entityId":"epoch:638b20a4ff6166bb4601","metric":"training-compute","value":5.46e+19,"unit":"FLOP","observationPeriod":"2022-11-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6*9100000000*1000000000=5.46e+19 (likely, might change because of several epochs / dataset division)\n\n\"The base model was trained using 256 NVIDIA A100 GPUs, while the two super-resolution models were trained with 128 NVIDIA A100 GPUs each\" \nno info on duration"},{"id":"epoch:3dd21ae3a20b8096cf11:parameters","entityId":"epoch:3dd21ae3a20b8096cf11","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2022-10-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:3dd21ae3a20b8096cf11:training-compute","entityId":"epoch:3dd21ae3a20b8096cf11","metric":"training-compute","value":5.1e+22,"unit":"FLOP","observationPeriod":"2022-10-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:4e5c01cec4cee8e9630e:training-compute","entityId":"epoch:4e5c01cec4cee8e9630e","metric":"training-compute","value":1.3150368e+18,"unit":"FLOP","observationPeriod":"2022-10-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"35580000000000 FLOP / sec [fp16 precision assumed] * 15400 sec [see training time notes] * 8 GPUs * 0.3 [assumed utilization] = 1.3150368e+18 FLOP \n\n_______\n\nThey used similar estimation (4.14 \u00d7 10^17 FLOP) in the Algorithmic Progress paper - not sure how exactly it was calculated "},{"id":"epoch:2cf592f30732a1f9e9ca:parameters","entityId":"epoch:2cf592f30732a1f9e9ca","metric":"parameters","value":247000000.00000003,"unit":"parameters","observationPeriod":"2022-10-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:59f831ffd6061c6dc9de:parameters","entityId":"epoch:59f831ffd6061c6dc9de","metric":"parameters","value":247000000.0,"unit":"parameters","observationPeriod":"2022-10-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"247M (Table 2)"},{"id":"epoch:59f831ffd6061c6dc9de:training-compute","entityId":"epoch:59f831ffd6061c6dc9de","metric":"training-compute","value":1.49682e+17,"unit":"FLOP","observationPeriod":"2022-10-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6NC = 6 * 247000000 * 101000000 = 1.49682e+17\n(no information about epochs -> speculative confidence)"},{"id":"epoch:ce4e6fe05661e41a9b57:training-compute","entityId":"epoch:ce4e6fe05661e41a9b57","metric":"training-compute","value":2.50000000000002e+22,"unit":"FLOP","observationPeriod":"2022-10-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 200x NVIDIA V100 GPUs with 1.30\u00d710^14 FLOP/s per GPU\n\n2. Training duration: Estimated 4 weeks based on \"multiple weeks\" phrase\n(4 weeks \u00d7 7 days \u00d7 24 hours \u00d7 3600 seconds = 2,419,200 seconds)\n\n3. Utilization rate: 40%\n\n4. Calculation:\n1.30\u00d710^14 FLOP/s \u00d7 200 GPUs \u00d7 2.4192\u00d710^6 seconds \u00d7 0.4 = 2.5\u00d710^22 FLOP"},{"id":"epoch:bd022d01c3024df46b8f:parameters","entityId":"epoch:bd022d01c3024df46b8f","metric":"parameters","value":2000000000.0,"unit":"parameters","observationPeriod":"2022-10-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 1"},{"id":"epoch:bd022d01c3024df46b8f:training-compute","entityId":"epoch:bd022d01c3024df46b8f","metric":"training-compute","value":7.2e+21,"unit":"FLOP","observationPeriod":"2022-10-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 2*10^9 parameters [figure 1] * 600 * 10^9 tokens [see dataset size notes] = 7.2e+21 FLOP"},{"id":"epoch:2ddacff548313e0c20d8:parameters","entityId":"epoch:2ddacff548313e0c20d8","metric":"parameters","value":182000000.0,"unit":"parameters","observationPeriod":"2022-10-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 3"},{"id":"epoch:1b24d7e68f9edb3e7080:parameters","entityId":"epoch:1b24d7e68f9edb3e7080","metric":"parameters","value":117000000.0,"unit":"parameters","observationPeriod":"2022-10-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 3"},{"id":"epoch:1f36bdc5ca6c9001dfc8:parameters","entityId":"epoch:1f36bdc5ca6c9001dfc8","metric":"parameters","value":107700000.0,"unit":"parameters","observationPeriod":"2022-10-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 3"},{"id":"epoch:1f36bdc5ca6c9001dfc8:training-compute","entityId":"epoch:1f36bdc5ca6c9001dfc8","metric":"training-compute","value":9.864288e+17,"unit":"FLOP","observationPeriod":"2022-10-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"RTX 8000 FP16 FLOPs: 32620000000000\nAssumed utilization for single GPU: 0.7\nEstimate: 12*60*60*32620000000000*0.7=9.864288e+17"},{"id":"epoch:1f33408dac67362321a9:parameters","entityId":"epoch:1f33408dac67362321a9","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2022-10-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"11B"},{"id":"epoch:c8b8d2188d9df20488cc:training-compute","entityId":"epoch:c8b8d2188d9df20488cc","metric":"training-compute","value":2.69568e+20,"unit":"FLOP","observationPeriod":"2022-10-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\" For\nCrossDocked, 100 training epochs take approximately 6 h/8 h in the C\u03b1 case and 48 h/60 h per 100\nepochs on a single NVIDIA A100 GPU with all atom pocket representation\"\n\n312000000000000*0.4*60 hours*10=2.695680e+20"},{"id":"epoch:a923052564d296e096bc:parameters","entityId":"epoch:a923052564d296e096bc","metric":"parameters","value":540000000000.0,"unit":"parameters","observationPeriod":"2022-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a923052564d296e096bc:training-compute","entityId":"epoch:a923052564d296e096bc","metric":"training-compute","value":2.53e+24,"unit":"FLOP","observationPeriod":"2022-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The total number of extra tokens we train on for the 540B\nmodel is approximately 1.3 Billion which constitutes 0.16% extra computation... Training an U-PaLM 540B model only consumes 512 TPUv4 chips and finishes in about 5 days which is considered to be lightweight.\"\n\noriginal PaLM was 2.527e+24. adding 0.16% is ~2.53e24"},{"id":"epoch:b4d84038ceba962e36ab:parameters","entityId":"epoch:b4d84038ceba962e36ab","metric":"parameters","value":540000000000.0,"unit":"parameters","observationPeriod":"2022-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"540B"},{"id":"epoch:914896617440aaa755f6:parameters","entityId":"epoch:914896617440aaa755f6","metric":"parameters","value":540000000000.0,"unit":"parameters","observationPeriod":"2022-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"540B"},{"id":"epoch:914896617440aaa755f6:training-compute","entityId":"epoch:914896617440aaa755f6","metric":"training-compute","value":2.540000000001e+24,"unit":"FLOP","observationPeriod":"2022-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"0.2% greater than Palm 540B, which used 2.5e24"},{"id":"epoch:64ea93a0805dbe3f82b5:parameters","entityId":"epoch:64ea93a0805dbe3f82b5","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2022-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"11B"},{"id":"epoch:64ea93a0805dbe3f82b5:training-compute","entityId":"epoch:64ea93a0805dbe3f82b5","metric":"training-compute","value":3.3e+22,"unit":"FLOP","observationPeriod":"2022-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2: 0.2% greater than T5 xxl, which used 3.3e22 FLOP"},{"id":"epoch:d958c2cbd6bbf77705b3:parameters","entityId":"epoch:d958c2cbd6bbf77705b3","metric":"parameters","value":31000000.0,"unit":"parameters","observationPeriod":"2022-10-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2\nThey compressed GPT-2 (124M parameters) to 31M parameters"},{"id":"epoch:d958c2cbd6bbf77705b3:training-compute","entityId":"epoch:d958c2cbd6bbf77705b3","metric":"training-compute","value":7.936e+20,"unit":"FLOP","observationPeriod":"2022-10-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"base model training compute estimation: 7.936 \u00d7 10^20 FLOP\n\n__________\nIn the Algorithmic progress paper the estimation was 6.2 \u00d7 10^19 FLOP"},{"id":"epoch:da048b33ea4b147e4544:parameters","entityId":"epoch:da048b33ea4b147e4544","metric":"parameters","value":25000000000.0,"unit":"parameters","observationPeriod":"2022-10-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 3"},{"id":"epoch:da048b33ea4b147e4544:training-compute","entityId":"epoch:da048b33ea4b147e4544","metric":"training-compute","value":1.42e+21,"unit":"FLOP","observationPeriod":"2022-10-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 3\nOverall ZettaFlops 1.42"},{"id":"epoch:071aa4d7c3e8ea0cc7b2:parameters","entityId":"epoch:071aa4d7c3e8ea0cc7b2","metric":"parameters","value":176500000000.0,"unit":"parameters","observationPeriod":"2022-10-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Two models: a LM that converts the input to code for a physics simulator, and a foundation model (InstructGPT)\n\n\"the resulting models have 0.3B and 1.5B parameters (used as default)\"\n\nInstructGPT is 175B. 175B+1.5B = 176.5B"},{"id":"epoch:e94d847b96aeb67123f3:parameters","entityId":"epoch:e94d847b96aeb67123f3","metric":"parameters","value":1120000000.0,"unit":"parameters","observationPeriod":"2022-10-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5326f27b63885ef623a2:parameters","entityId":"epoch:5326f27b63885ef623a2","metric":"parameters","value":242000000.0,"unit":"parameters","observationPeriod":"2022-10-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We fine-tune starting from the checkpoint provided by Baevski\nand Auli (2019)8, which has 242M parameters\""},{"id":"epoch:5326f27b63885ef623a2:training-compute","entityId":"epoch:5326f27b63885ef623a2","metric":"training-compute","value":7.9e+19,"unit":"FLOP","observationPeriod":"2022-10-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Pre-trained model is Transformer (Adaptive Input Embeddings) which was 7.3e19. This is from 8 * 67 V100-hours. \n\nfine-tuning:\n\n\"Training was performed on a single NVIDIA A40 GPU for 35 hours\"\n\n35h, 1 GPU, 149.7e12, 30% = 5.7e18 FLOP\"\n\n5.7e18 + 7.3e19 is 7.9e19"},{"id":"epoch:be520cca5fb3b1dc297e:parameters","entityId":"epoch:be520cca5fb3b1dc297e","metric":"parameters","value":1800000000.0,"unit":"parameters","observationPeriod":"2022-10-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Unless specified otherwise, we train a 1.8B parameter Phenaki model on a corpus of \u223c15M textvideo pairs at 8 FPS mixed with \u223c50M text-images plus \u223c400M pairs of LAION-400M [41] (more\ndetails in Appendix B.3). The model used in the visualisations in this paper was trained for 1 million\nsteps at a batch size of 512, which took less than 5 days. In this setup 80% of the training data came\nfrom the video dataset and each image dataset contributed 10%."},{"id":"epoch:982db4ad746bbeb9908e:parameters","entityId":"epoch:982db4ad746bbeb9908e","metric":"parameters","value":11600000000.0,"unit":"parameters","observationPeriod":"2022-10-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 6 summarizes the entire cascading pipeline of Imagen Video. In total, we have 1 frozen text encoder, 1 base video diffusion model, 3 SSR (spatial super-resolution), and 3 TSR (temporal superresolution) models \u2013 for a total of 7 video diffusion models, with a total of 11.6B diffusion model parameters"},{"id":"epoch:4b2943647fb6beffe18d:training-compute","entityId":"epoch:4b2943647fb6beffe18d","metric":"training-compute","value":7.1414784e+20,"unit":"FLOP","observationPeriod":"2022-10-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Compute: 0.3 [assumed utilization] * (64 cores / 2 cores per chip )*123000000000000 FLOP / TPU v3 chip / sec * 7 days * 24 hours / day * 3600 sec / hour = 7.1414784e+20 FLOP"},{"id":"epoch:247a7fd6ae7b1ad465f4:parameters","entityId":"epoch:247a7fd6ae7b1ad465f4","metric":"parameters","value":1500000.0,"unit":"parameters","observationPeriod":"2022-10-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e248a8397a01373a7dad:parameters","entityId":"epoch:e248a8397a01373a7dad","metric":"parameters","value":20240000.0,"unit":"parameters","observationPeriod":"2022-10-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For determining the hyperparameters of DIFFDOCK\u2019s score model, we trained\nsmaller models (3.97 million parameters) that fit into 48GB of GPU RAM before scaling it up to the final model (20.24 million parameters) that was trained on four 48GB GPUs\"\n\nThere's a separate 4.77M \"confidence model\" that helps make predictions along with the score model"},{"id":"epoch:e248a8397a01373a7dad:training-compute","entityId":"epoch:e248a8397a01373a7dad","metric":"training-compute","value":7.2e+19,"unit":"FLOP","observationPeriod":"2022-10-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We trained our final score model on four 48GB RTX A6000 GPUs for 850 epochs (around 18 days).\"\n\n4 * 38.7 teraflops * 18 days * 24 * 3600 * 0.3 = 7.2e19\n\nhttps://www.techpowerup.com/gpu-specs/rtx-a6000.c3686"},{"id":"epoch:d919944edbaaaa98f145:parameters","entityId":"epoch:d919944edbaaaa98f145","metric":"parameters","value":124000000.0,"unit":"parameters","observationPeriod":"2022-10-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"124M (the same as GPT-2)"},{"id":"epoch:d919944edbaaaa98f145:training-compute","entityId":"epoch:d919944edbaaaa98f145","metric":"training-compute","value":1.20380928e+20,"unit":"FLOP","observationPeriod":"2022-10-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"assuming that GPT-2 (124M) compute is similar to GPT-2 (117M) compute, which is 120000000000000000000 FLOP = 1.2 \u00d7 10^20 FLOP\n\n120000000000000000000 FLOP + 380928000000000000 FLOP = 1.204\u00d710^20 FLOP "},{"id":"epoch:9c47673da5474ea33d97:training-compute","entityId":"epoch:9c47673da5474ea33d97","metric":"training-compute","value":5.4e+20,"unit":"FLOP","observationPeriod":"2022-09-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"125000000000000 FLOP / GPU / sec [bf16 assumed] * 4000 GPU-hours [inferring from Figure 2] * 3600 sec / hour * 0.3 [asssumed utilization] = 5.4e+20 FLOP"},{"id":"epoch:1023fb967191b3443ce5:parameters","entityId":"epoch:1023fb967191b3443ce5","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2022-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"70B"},{"id":"epoch:41afa5636229ce6c1665:parameters","entityId":"epoch:41afa5636229ce6c1665","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2022-09-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"CPM-Ant is an open-source Chinese pre-trained language model (PLM) with 10B parameters.\""},{"id":"epoch:41afa5636229ce6c1665:training-compute","entityId":"epoch:41afa5636229ce6c1665","metric":"training-compute","value":2.004e+22,"unit":"FLOP","observationPeriod":"2022-09-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 10*10e9 parameters * 3.34e10 tokens [see training dataset size notes] = 2.004e+22 FLOP"},{"id":"epoch:27482b52c5711ccbc3c9:parameters","entityId":"epoch:27482b52c5711ccbc3c9","metric":"parameters","value":1550000000.0,"unit":"parameters","observationPeriod":"2022-09-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:27482b52c5711ccbc3c9:training-compute","entityId":"epoch:27482b52c5711ccbc3c9","metric":"training-compute","value":4.2072663e+21,"unit":"FLOP","observationPeriod":"2022-09-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See figure 9"},{"id":"epoch:7599287795c918b70b05:parameters","entityId":"epoch:7599287795c918b70b05","metric":"parameters","value":230000000.0,"unit":"parameters","observationPeriod":"2022-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we were able to reduce the number of DistilProtBert parameters by almost half, to 230 M\""},{"id":"epoch:7599287795c918b70b05:training-compute","entityId":"epoch:7599287795c918b70b05","metric":"training-compute","value":1.9e+20,"unit":"FLOP","observationPeriod":"2022-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Pretraining was done on five v100 32-GB Nvidia GPUs from a DGX\ncluster with a local batch size of 16 examples... The model was trained for three epochs using mixed precision and dynamic padding. Every epoch run took approximately 4 days, resulting in total pretraining time of 12 days\"\n\n5 * 125 teraFLOP/s * 12 * 24 * 3600 * 0.3 (assumed utilization) = 1.9e20"},{"id":"epoch:4725c5978bb6730e0602:parameters","entityId":"epoch:4725c5978bb6730e0602","metric":"parameters","value":986000000.0,"unit":"parameters","observationPeriod":"2022-09-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"986M params from https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K"},{"id":"epoch:4725c5978bb6730e0602:training-compute","entityId":"epoch:4725c5978bb6730e0602","metric":"training-compute","value":7.8410000000001e+22,"unit":"FLOP","observationPeriod":"2022-09-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"A CLIP ViT-H/14 model trained with the LAION-2B English subset of LAION-5B (https://laion.ai/blog/laion-5b/) using OpenCLIP (https://github.com/mlfoundations/open_clip).\"\nPer https://laion.ai/blog/large-openclip/ H/14 used 824 A100s, trained 42 samples per gpu-second, saw 32B samples\ncompute = 33% * 311.84 TFLOPS * (32 billion / 42) seconds = 7.841e22 FLOP"},{"id":"epoch:f44e043900dc06b35238:parameters","entityId":"epoch:f44e043900dc06b35238","metric":"parameters","value":20000000000.0,"unit":"parameters","observationPeriod":"2022-09-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"20B"},{"id":"epoch:fc8a0df87b5dc8a2bbff:parameters","entityId":"epoch:fc8a0df87b5dc8a2bbff","metric":"parameters","value":16900000000.0,"unit":"parameters","observationPeriod":"2022-09-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3.9b Image Encoder, \n14b Multimodal Encoder-Decoder"},{"id":"epoch:fc8a0df87b5dc8a2bbff:training-compute","entityId":"epoch:fc8a0df87b5dc8a2bbff","metric":"training-compute","value":1.69e+23,"unit":"FLOP","observationPeriod":"2022-09-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Pre-training the ViT component involved 1.1 million steps (they train over 1M steps but run the last 100k twice and then average the two resulting models). Batch size is 16384 and the inputs are 224x224. Table 8 indicates a forward pass with ViT-e/14 on a 224 image takes 1980 GFLOPs, so total training compute for the ViT-e/14 model is:\n1980e9 * 16384 * 1.1 million * 3 (account for backward passes) = 1.07e23\n\nIn the \"overal model\" section, they then say: \"The largest model, PaLI-17B, is pretrained using 1,024 GCP-TPUv4 chips for 7 days\". It is then trained for another 3 days on 512 chips at higher resolution. \n\nI assume the stated TPUv4 training does not include the ViT pretraining, since it amounts to fewer FLOPs than we estimate above for the ViT.\n\n275 teraFLOP/s * ((1024 * 7) + (512 * 3)) * 24 * 3600 * 0.3 (utilization assumption) = 6.2e22\n\nTotal: 1.07e23 + 6.2e22 = 1.69e23"},{"id":"epoch:24e0fbe3822fc0801e66:parameters","entityId":"epoch:24e0fbe3822fc0801e66","metric":"parameters","value":1900000000.0,"unit":"parameters","observationPeriod":"2022-08-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.9B from Table 2"},{"id":"epoch:24e0fbe3822fc0801e66:training-compute","entityId":"epoch:24e0fbe3822fc0801e66","metric":"training-compute","value":7e+19,"unit":"FLOP","observationPeriod":"2022-08-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from Table 11, 1M training steps with batch size 6144. \nFrom Table 2 we have that model have 1.9B parameters.\nModel is VIT"},{"id":"epoch:a4e69d9483133b88f7ed:training-compute","entityId":"epoch:a4e69d9483133b88f7ed","metric":"training-compute","value":5.0000000000000004e+22,"unit":"FLOP","observationPeriod":"2022-08-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000 FLOP / GPU / sec * 150000 GPU-hours * 3600 sec / hour * 0.3 [assumed utilization] = 5.0544e+22 FLOP\n\nhttps://github.com/CompVis/stable-diffusion/blob/main/Stable_Diffusion_v1_Model_Card.md"},{"id":"epoch:abb469a0e09322ae7b15:parameters","entityId":"epoch:abb469a0e09322ae7b15","metric":"parameters","value":540000000000.0,"unit":"parameters","observationPeriod":"2022-08-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"540B from PaLM-540B.\n\n"},{"id":"epoch:67bfd5e56b37747cf53a:parameters","entityId":"epoch:67bfd5e56b37747cf53a","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2022-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"~70B\""},{"id":"epoch:67bfd5e56b37747cf53a:training-compute","entityId":"epoch:67bfd5e56b37747cf53a","metric":"training-compute","value":3.5461e+23,"unit":"FLOP","observationPeriod":"2022-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"~839000h\" GPU-hours on A100s, per Environmental Impact section of model card.\n\n312 trillion * 839000 * 3600 * 0.3 = 2.8e23\n\n6ND = 6*70B*1069.30B = 4.49106e+23\n\nsqrt(2.8e23*4.49106e+23) = 3.54612... \u00d7 10^23\n\nreported here:\n167TFLOPS\nhttps://docs.aleph-alpha.com/docs/Deprecated%20Luminous/Deprecated-Luminous/model-card/"},{"id":"epoch:6a6359c298852b46197f:parameters","entityId":"epoch:6a6359c298852b46197f","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2022-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"~30B (~42B with multi-modality)"},{"id":"epoch:6a6359c298852b46197f:training-compute","entityId":"epoch:6a6359c298852b46197f","metric":"training-compute","value":1.0019457e+23,"unit":"FLOP","observationPeriod":"2022-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"311840000000000*360000*3600*0.3 = 1.2124339e+23\n\n6ND = 6*30*10^9*460000000000 = 8.28e+22\n\nsqrt(8.28e+22*1.2124339e+23) = 1.0019457e+23"},{"id":"epoch:8ac70764a8eaaa9c5e40:parameters","entityId":"epoch:8ac70764a8eaaa9c5e40","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2022-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"~13B (~15B with multi-modality)"},{"id":"epoch:8ac70764a8eaaa9c5e40:training-compute","entityId":"epoch:8ac70764a8eaaa9c5e40","metric":"training-compute","value":3.1673782e+22,"unit":"FLOP","observationPeriod":"2022-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"311840000000000*95000*3600*0.3 = 3.1994784e+22\n\n6ND = 6*13*10^9*402000000000 = 3.1356e+22\n\nsqrt(3.1356e+22* 3.1994784e+22) = 3.1673782e+22"},{"id":"epoch:f442682820d5e93b10b3:parameters","entityId":"epoch:f442682820d5e93b10b3","metric":"parameters","value":40000000.0,"unit":"parameters","observationPeriod":"2022-08-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f442682820d5e93b10b3:training-compute","entityId":"epoch:f442682820d5e93b10b3","metric":"training-compute","value":1e+20,"unit":"FLOP","observationPeriod":"2022-08-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 8x NVIDIA Tesla V100 SXM2 32GB GPUs (1.25 x 10^14 FLOP/s per GPU)\n\n2. Training duration: 70 hours (directly provided) = 252,000 seconds\n\n3. Utilization rate: 40%\n\n4. Final calculation:\n1.25 x 10^14 FLOP/s/GPU \u00d7 8 GPUs \u00d7 252,000 seconds \u00d7 0.4 = 1.0 x 10^20 FLOPs"},{"id":"epoch:a70a6e122ef72b60101e:parameters","entityId":"epoch:a70a6e122ef72b60101e","metric":"parameters","value":175000000000.0,"unit":"parameters","observationPeriod":"2022-08-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a70a6e122ef72b60101e:training-compute","entityId":"epoch:a70a6e122ef72b60101e","metric":"training-compute","value":4.3e+23,"unit":"FLOP","observationPeriod":"2022-08-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(taken from OPT-175 base)"},{"id":"epoch:5c9d1bbdb2a55c1165c9:training-compute","entityId":"epoch:5c9d1bbdb2a55c1165c9","metric":"training-compute","value":2.59999999999998e+21,"unit":"FLOP","observationPeriod":"2022-08-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 8x NVIDIA A100 GPUs (3.12E14 FLOP/s per GPU)\n\n2. Training duration: 30 days estimated (2,592,000 seconds)\n\n3. Utilization rate: 40%\n\n4. Calculation: 3.12E14 FLOP/s \u00d7 8 GPUs \u00d7 2,592,000 seconds \u00d7 0.4 = 2.59E21 FLOPs"},{"id":"epoch:3c54ee6d024b87967ffd:parameters","entityId":"epoch:3c54ee6d024b87967ffd","metric":"parameters","value":27000000.0,"unit":"parameters","observationPeriod":"2022-08-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"27M"},{"id":"epoch:3c54ee6d024b87967ffd:training-compute","entityId":"epoch:3c54ee6d024b87967ffd","metric":"training-compute","value":2.2977e+18,"unit":"FLOP","observationPeriod":"2022-08-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"125000000000000 FLOP / GPU / sec * 17.02 GPU-hours * 3600 sec / hour * 0.3 [assumed utilization] = 2.2977e+18 FLOP"},{"id":"epoch:16c34f9da8627ca131b7:parameters","entityId":"epoch:16c34f9da8627ca131b7","metric":"parameters","value":5800000000.0,"unit":"parameters","observationPeriod":"2022-08-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"5.8B"},{"id":"epoch:01a3f8d13fbc150dc6d0:parameters","entityId":"epoch:01a3f8d13fbc150dc6d0","metric":"parameters","value":130000000000.0,"unit":"parameters","observationPeriod":"2022-08-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Dense model"},{"id":"epoch:01a3f8d13fbc150dc6d0:training-compute","entityId":"epoch:01a3f8d13fbc150dc6d0","metric":"training-compute","value":3.5490054945e+23,"unit":"FLOP","observationPeriod":"2022-08-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"96 NVIDIA A100 (40G * 8) servers for 2 months\"\n\n312 TFLOPS/GPU * 96 servers * 8 GPU/server * 2 months * 32.5% utilization = 4.037e23\n\nutilization rate - citation from the paper: \"we report hardware FLOPs utilization (HFU) of 43.3% and model FLOPs utilization (MFU) of 32.5% due to re-materialization.\"\n\nAligns pretty well with 6ND:\n6 * 400B * 130B = 3.12E23\n\nGeometric mean: sqrt(4.037e23 * 3.12e23) = 3.549e23"},{"id":"epoch:6d8e7837926bcb0ee5e0:parameters","entityId":"epoch:6d8e7837926bcb0ee5e0","metric":"parameters","value":19750000000.0,"unit":"parameters","observationPeriod":"2022-08-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 1 on p.3 of the paper"},{"id":"epoch:6d8e7837926bcb0ee5e0:training-compute","entityId":"epoch:6d8e7837926bcb0ee5e0","metric":"training-compute","value":2.04374016e+23,"unit":"FLOP","observationPeriod":"2022-08-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training throughput is reported as 154 TFLOP/s - see p.5 of the paper.\n\"We relied on an internal and optimized version of DeepSpeed that we have since open-sourced (Chiu & Zheng, 2022) to obtain training throughput of up to 154 TFLOPS/GPU on 16 AWS p4d.24xlarge compute instances.\"\n\nAccelerator compute days are reported as 15,360 days - see Table 17 on p.18 of the paper."},{"id":"epoch:7a3fa2d7e2588672e88f:parameters","entityId":"epoch:7a3fa2d7e2588672e88f","metric":"parameters","value":738000000.0,"unit":"parameters","observationPeriod":"2022-07-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Here, we introduce ProtGPT2, an autoregressive Transformer model with 738 million parameters capable of generating de novo protein sequences in a high-throughput fashion.\""},{"id":"epoch:7a3fa2d7e2588672e88f:training-compute","entityId":"epoch:7a3fa2d7e2588672e88f","metric":"training-compute","value":4.1e+21,"unit":"FLOP","observationPeriod":"2022-07-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The model trained on 128 NVIDIA A100s in 4 days\"\n\n128 * 4 * 24 * 3600 * 312 trillion FLOP/s * 0.3 = 4.1e21\n\n5.3E+21 from Table 9 https://www.biorxiv.org/content/10.1101/2023.07.05.547496v1"},{"id":"epoch:678fc4ffa5c953ebbe13:parameters","entityId":"epoch:678fc4ffa5c953ebbe13","metric":"parameters","value":2700000000.0,"unit":"parameters","observationPeriod":"2022-07-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.7B"},{"id":"epoch:d61f8cccc10584535ca7:parameters","entityId":"epoch:d61f8cccc10584535ca7","metric":"parameters","value":670000000.0,"unit":"parameters","observationPeriod":"2022-07-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our model contains 66 layers with around 670 million parameters without sharing parameters, which doubles the layer count of ESM-1b but roughly retains the parameter count.\""},{"id":"epoch:d61f8cccc10584535ca7:training-compute","entityId":"epoch:d61f8cccc10584535ca7","metric":"training-compute","value":1.03514112e+22,"unit":"FLOP","observationPeriod":"2022-07-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"OmegaPLM is implemented in PyTorch (44) and trained for 2,560 GPU Nvidia A100 80G days.\" \n\"Default precision format in Nvidia A100 GPUs is set to TensorFloat-32 for matrix operations.\"\n\nAssume 0.3 utilization for language model\n\nEstimate: (2560 * 24 * 3600) s * 156e12 FLOP/s * 0.3 * = 1.04e22"},{"id":"epoch:b136b42f7f24f8f74999:parameters","entityId":"epoch:b136b42f7f24f8f74999","metric":"parameters","value":15000000000.0,"unit":"parameters","observationPeriod":"2022-07-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we train models up to 15B parameters\""},{"id":"epoch:b136b42f7f24f8f74999:training-compute","entityId":"epoch:b136b42f7f24f8f74999","metric":"training-compute","value":7.35000000001e+22,"unit":"FLOP","observationPeriod":"2022-07-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from xTrimoPGLM paper Table 9 (https://www.biorxiv.org/content/10.1101/2023.07.05.547496v1): 5.1e22 FLOP\n\nfrom Arb Research (https://arbresearch.com/files/gen_bio.pdf): \"ESM-2-15B: 270000 updates x 3.2M batch size x 15 B \u201cconnections\u201d x 6. : 7.8e22 FLOP\n\nfrom the paper's Supplementary Materials: \n\"We trained each model over 512 NVIDIA V100 GPUs. ESM2 700M took 8 days to train. The 3B parameter LM took 30 days. The 15B model took 60 days.\"\n60 days x 512 V100s x an imputed 30% utilization\": 1e23 FLOP\n\nGeometric mean: 7.35e22"},{"id":"epoch:1d90515f16d445f99ed1:parameters","entityId":"epoch:1d90515f16d445f99ed1","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2022-07-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"In the name\n"},{"id":"epoch:1d90515f16d445f99ed1:training-compute","entityId":"epoch:1d90515f16d445f99ed1","metric":"training-compute","value":3.000000001e+22,"unit":"FLOP","observationPeriod":"2022-07-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from xTrimoPGLM paper Table 9 (https://www.biorxiv.org/content/10.1101/2023.07.05.547496v1): 1.8e22 FLOP\n\nfrom the paper's Supplementary Materials: \n\"We trained each model over 512 NVIDIA V100 GPUs. ESM2 700M took 8 days to train. The 3B parameter LM took 30 days. The 15B model took 60 days.\"\n30 days x 512 V100s x an imputed 30% utilization\": 5e22 FLOP\n\nGeometric mean: 3e22"},{"id":"epoch:67634548355656e5c40b:parameters","entityId":"epoch:67634548355656e5c40b","metric":"parameters","value":650000000.0,"unit":"parameters","observationPeriod":"2022-07-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"In the name"},{"id":"epoch:67634548355656e5c40b:training-compute","entityId":"epoch:67634548355656e5c40b","metric":"training-compute","value":7.560000000001e+21,"unit":"FLOP","observationPeriod":"2022-07-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from xTrimoPGLM paper Table 9 (https://www.biorxiv.org/content/10.1101/2023.07.05.547496v1): 4.4e21 FLOP\n\nfrom the paper's Supplementary Materials: \n\"We trained each model over 512 NVIDIA V100 GPUs. ESM2 700M took 8 days to train. The 3B parameter LM took 30 days. The 15B model took 60 days.\"\n8 days x 512 V100s x an imputed 30% utilization\": 1.3e22 FLOP\n\nGeometric mean: 7.56e21 FLOP"},{"id":"epoch:d5c26b8185959f392627:parameters","entityId":"epoch:d5c26b8185959f392627","metric":"parameters","value":150000000.0,"unit":"parameters","observationPeriod":"2022-07-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"In the name"},{"id":"epoch:d5c26b8185959f392627:training-compute","entityId":"epoch:d5c26b8185959f392627","metric":"training-compute","value":1.1e+21,"unit":"FLOP","observationPeriod":"2022-07-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from xTrimoPGLM paper Table 9 (https://www.biorxiv.org/content/10.1101/2023.07.05.547496v1): 1.1e21 FLOP"},{"id":"epoch:bf24aab3ee8e0193351b:parameters","entityId":"epoch:bf24aab3ee8e0193351b","metric":"parameters","value":35000000.0,"unit":"parameters","observationPeriod":"2022-07-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"In the name"},{"id":"epoch:bf24aab3ee8e0193351b:training-compute","entityId":"epoch:bf24aab3ee8e0193351b","metric":"training-compute","value":2.1e+20,"unit":"FLOP","observationPeriod":"2022-07-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"All language models were trained for 500K updates, except the 15B language model\" \n\"All models used 2 million tokens as batch size except the 15B model\"\n[Supplementary Materials]\n\nHence: 1000B training tokens (500k steps, 2M tokens/batch)\n\nEstimate: 35M*2*1000B + 35M*4*1000B"},{"id":"epoch:5cb729103d3c98f8031d:parameters","entityId":"epoch:5cb729103d3c98f8031d","metric":"parameters","value":8000000.0,"unit":"parameters","observationPeriod":"2022-07-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"In the name"},{"id":"epoch:5cb729103d3c98f8031d:training-compute","entityId":"epoch:5cb729103d3c98f8031d","metric":"training-compute","value":4.8e+19,"unit":"FLOP","observationPeriod":"2022-07-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"All language models were trained for 500K updates, except the 15B language model\" \n\"All models used 2 million tokens as batch size except the 15B model\"\n[Supplementary Materials]\n\nHence: 1000B training tokens (500k steps, 2M tokens/batch)\n\nEstimate: 8M*2*1000B + 8M*4*1000B"},{"id":"epoch:f3517279d4b3272b73fa:parameters","entityId":"epoch:f3517279d4b3272b73fa","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2022-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://huggingface.co/FUXI/yuyan-11b"},{"id":"epoch:74623e36e431f799c5de:parameters","entityId":"epoch:74623e36e431f799c5de","metric":"parameters","value":247000000.00000003,"unit":"parameters","observationPeriod":"2022-07-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"WikiText-103 experiments use 16-layer Transformers (10 heads, 410 hidden size, 2100 intermediate FF)\""},{"id":"epoch:74623e36e431f799c5de:training-compute","entityId":"epoch:74623e36e431f799c5de","metric":"training-compute","value":6.7839792e+18,"unit":"FLOP","observationPeriod":"2022-07-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 200000 steps * 128 tokens per batch * 60 batches per step * 247000000 parameters= 2.276352e+18 FLOP\n\n312000000000000 FLOP / sec / GPU * 2 GPUs * 30 hours [could be less, upper bound] * 3600 sec / hour * 0.3 [assumed utilization] = 2.02176e+19 FLOP\n\nsqrt(2.276352e+18*2.02176e+19) = 6.7839792e+18 FLOP"},{"id":"epoch:8684289d4c0ef11e2765:parameters","entityId":"epoch:8684289d4c0ef11e2765","metric":"parameters","value":1200000000.0,"unit":"parameters","observationPeriod":"2022-07-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"with up to 1.2 billion parameters\""},{"id":"epoch:8684289d4c0ef11e2765:training-compute","entityId":"epoch:8684289d4c0ef11e2765","metric":"training-compute","value":8.64e+20,"unit":"FLOP","observationPeriod":"2022-07-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The models were trained for a total training time of over 25 thousand Nvidia-V100 GPU hours\"\n\n125 teraFLOP/s (uncertain which V100 model, tensor performance varies from 112-130tFLOP/s) * 25000 * 3600 * 0.3 (utilization) = 3.4e+21\" <- the total compute for several models\n\nFor the biggest (XL) from Figure 1:\n10 PF-days = 10*10^15*24*3600 FLOPs = 8.64e+20 FLOPs <- total compute just for the XL model"},{"id":"epoch:aaf478f87269a2ba3df8:parameters","entityId":"epoch:aaf478f87269a2ba3df8","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2022-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:be9b8a581be359066ea1:parameters","entityId":"epoch:be9b8a581be359066ea1","metric":"parameters","value":176247271424.0,"unit":"parameters","observationPeriod":"2022-07-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See \"Technical Specifications\" on Hugging Face:\nhttps://huggingface.co/bigscience/bloom"},{"id":"epoch:be9b8a581be359066ea1:training-compute","entityId":"epoch:be9b8a581be359066ea1","metric":"training-compute","value":3.65664e+23,"unit":"FLOP","observationPeriod":"2022-07-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://bigscience.huggingface.co/blog/bloom Blog post says 117 days.\n\n384 A100 GPUs * 314 TFLOPS throughput per GPU * 117 days * 0.3 (utilization assumption) = 3.65664e23\nhttps://www.wolframalpha.com/input?i=384+*+314+TFLOPS+*+117+days+*+0.3"},{"id":"epoch:85d1fa8b33f176cc60f5:parameters","entityId":"epoch:85d1fa8b33f176cc60f5","metric":"parameters","value":54500000000.0,"unit":"parameters","observationPeriod":"2022-07-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 8.2.4: \"The model has a total of 54.5B parameters\nand FLOPs similar to that of a 3.3B dense model\""},{"id":"epoch:85d1fa8b33f176cc60f5:training-compute","entityId":"epoch:85d1fa8b33f176cc60f5","metric":"training-compute","value":1.751113728e+22,"unit":"FLOP","observationPeriod":"2022-07-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 8.8:\n\" To train NLLB-200, a cumulative\nof 51968 GPU hours of computation was performed on hardware of type A100-SXM-80GB\"\nSee also Table 48\n\nSection 8.2.4 states they use FP16\n\nNVIDIA Datasheet states 312TFLOPS for FP16\nhttps://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/a100/pdf/nvidia-a100-datasheet-nvidia-us-2188504-web.pdf\n\nAssuming 0.3 utilization:\n\n312e12*3600*51968*0.3\n\nAlso:\n\"Our final model is a Transformer\nencoder-decoder model in which we replace the Feed Forward Network (FFN) layer in\nevery 4th Transformer block with a Sparsely Gated Mixture of Experts layer containing 128\nexperts. We use model dimension 2048, FFN dimension 8192, 16 attention heads, 24 encoder\nlayers and 24 decoder layers. We use Pre-LayerNorm (Xiong et al., 2020) as described in\nSection 6.1.1. We share the embedding weights of the encoder input embedding, decoder\ninput embedding and decoder output embedding layers. We use an overall dropout of 0.3,\nattention dropout 0.1 and EOM with peom=0.2. The model has a total of 54.5B parameters\nand FLOPs similar to that of a 3.3B dense model.\""},{"id":"epoch:5bfd56b50468e2dba11c:parameters","entityId":"epoch:5bfd56b50468e2dba11c","metric":"parameters","value":7070000000.0,"unit":"parameters","observationPeriod":"2022-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5bfd56b50468e2dba11c:training-compute","entityId":"epoch:5bfd56b50468e2dba11c","metric":"training-compute","value":1.5014556e+22,"unit":"FLOP","observationPeriod":"2022-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 7070000000 parameters * 354000000000 tokens = 1.501668 \u00d7 10^22 FLOP"},{"id":"epoch:da614b9af1862f2b97d3:parameters","entityId":"epoch:da614b9af1862f2b97d3","metric":"parameters","value":770000000.0,"unit":"parameters","observationPeriod":"2022-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We pretrain a CodeT5-large model (770M) from scratch following T5-large\u2019s architecture\""},{"id":"epoch:da614b9af1862f2b97d3:training-compute","entityId":"epoch:da614b9af1862f2b97d3","metric":"training-compute","value":2.72e+21,"unit":"FLOP","observationPeriod":"2022-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We perform our experiments on a kubernetes with 16 A100-40G GPUs on Google Cloud Platform and the total pretraining duration is around 21 days\"\n\n16 * 312tFLOP/s * 21 * 24 * 3600 * 0.3 (utilization assumption) = 2.72e21"},{"id":"epoch:c154f7d8f1aba24c90e3:parameters","entityId":"epoch:c154f7d8f1aba24c90e3","metric":"parameters","value":1700000000.0,"unit":"parameters","observationPeriod":"2022-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a3e4037cfca4bb0542e6:parameters","entityId":"epoch:a3e4037cfca4bb0542e6","metric":"parameters","value":560000000.0,"unit":"parameters","observationPeriod":"2022-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ebc6049045f95b3b511d:parameters","entityId":"epoch:ebc6049045f95b3b511d","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2022-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c68e6076e9873974d5c0:parameters","entityId":"epoch:c68e6076e9873974d5c0","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2022-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:23f1fc5ff64792d215b7:parameters","entityId":"epoch:23f1fc5ff64792d215b7","metric":"parameters","value":175000000000.0,"unit":"parameters","observationPeriod":"2022-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"same as the base model"},{"id":"epoch:5f91e7d88eb86ecf7c2c:parameters","entityId":"epoch:5f91e7d88eb86ecf7c2c","metric":"parameters","value":540350000000.0,"unit":"parameters","observationPeriod":"2022-06-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"To further our understanding of the impact of scale on few-shot learning, we trained a 540-billion parameter, densely activated, Transformer language model, which we call Pathways Language Model (PaLM).\"\n\nOur approach is to start with the PaLM pretrained decoder-only transformer language models Chowdhery et al. (2022), and further train (finetune) them on our mathematical dataset using an autoregressive objective.\nTable 2 contains the main model and training hyperparameters.\n\nSee Table 2"},{"id":"epoch:5f91e7d88eb86ecf7c2c:training-compute","entityId":"epoch:5f91e7d88eb86ecf7c2c","metric":"training-compute","value":2.7415e+24,"unit":"FLOP","observationPeriod":"2022-06-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Minerva was fine-tuned from PaLM using the same hardware. Assume the same model FLOPs utilization rate for pre-training and fine-tuning.\n\"the 540B model was trained for 29 days on a v4-1024\"\n\nPaLM pretraining time: 6144 TPU for 1200 hours + 3072 TPU for 336 hours = @8404992 TPU-hours\nMinerva finetuning time: 1024 TPU for 696 hours = 712704 TPU-hours\nSo fine-tuning added 8.5% more compute.\n\nMinerva total compute = PaLM pretraining compute * (712704+8404992)/(8404992) = 2.7415*10^24 FLOP\nhttps://www.wolframalpha.com/input?i=%28712704%2B8404992%29%2F%288404992%29+*+2.5272*10%5E24\n\nPalm pretraining: 2.5272e+24"},{"id":"epoch:3e4908b91c159f2e5019:training-compute","entityId":"epoch:3e4908b91c159f2e5019","metric":"training-compute","value":2.285273088e+22,"unit":"FLOP","observationPeriod":"2022-06-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"flops = (128) * (1.23e14) * (1344 * 3600) * (0.3) = 2.3e22\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)\nfrom https://huggingface.co/dalle-mini/dalle-mega#environmental-impact"},{"id":"epoch:0de8afd45ab5664cf3ac:parameters","entityId":"epoch:0de8afd45ab5664cf3ac","metric":"parameters","value":6400000000.0,"unit":"parameters","observationPeriod":"2022-06-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We introduce a suite of protein language models, named ProGen2, that are scaled up to 6.4B parameters\""},{"id":"epoch:0de8afd45ab5664cf3ac:training-compute","entityId":"epoch:0de8afd45ab5664cf3ac","metric":"training-compute","value":1.35e+22,"unit":"FLOP","observationPeriod":"2022-06-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Estimate 1:\n\"350,000 steps x 1m batch size x 6.4 B \u201cconnections\u201d x 6\" - Arb Research (https://arbresearch.com/files/gen_bio.pdf)\nSteps and batches from Table 1. \nFLOP estimate: 1.3e22\n\nTable 9 from here: https://www.biorxiv.org/content/10.1101/2023.07.05.547496v1.full.pdf\nFLOP estimate: 1.4e22\n\nGeometric mean = 1.35e22 FLOP"},{"id":"epoch:5462d2a7b3583b35b590:parameters","entityId":"epoch:5462d2a7b3583b35b590","metric":"parameters","value":764000000.0,"unit":"parameters","observationPeriod":"2022-06-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"764M (Table1)"},{"id":"epoch:5462d2a7b3583b35b590:training-compute","entityId":"epoch:5462d2a7b3583b35b590","metric":"training-compute","value":1.1e+21,"unit":"FLOP","observationPeriod":"2022-06-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 9 from here: https://www.biorxiv.org/content/10.1101/2023.07.05.547496v1.full.pdf\nFLOP estimate: 1.1E+21\n\n6 FLOP / parameter / token * 764*10^6 parameters * 200*10^9 tokens = 9.168e+20 FLOP"},{"id":"epoch:c323f4dcb307bb217564:parameters","entityId":"epoch:c323f4dcb307bb217564","metric":"parameters","value":3500000000.0,"unit":"parameters","observationPeriod":"2022-06-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c323f4dcb307bb217564:training-compute","entityId":"epoch:c323f4dcb307bb217564","metric":"training-compute","value":1.6663535e+21,"unit":"FLOP","observationPeriod":"2022-06-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From section 4: \"Training was performed on GPU resources from the Berzelius Superpod, which is currently the fastest super\ncomputer in Sweden, equipped with 60 Nvidia DGX\nA100 servers, each of which consists of 8 Nvidia A100\nGPUs with 320 GB Total GPU memory. Our training\nprocess took 2.5 days utilizing 16 of the DGX A100\nservers (in total 128 GPUs).\"\n\n2.5*24*60**2 * 128 * 1.56E+14 * 0.3 = 1.3e21\n\n6 FLOP / parameter / token * 3.5 * 10^9 parameters * 97000 steps * 512 samples per batch * 2048 tokens per sequence = 2.1359493e+21 FLOP\n\nsqrt(1.3e21*2.1359493e+21) = 1.6663535e+21 FLOP"},{"id":"epoch:f16fb17630ce9d321658:parameters","entityId":"epoch:f16fb17630ce9d321658","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2022-06-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"100B"},{"id":"epoch:f16fb17630ce9d321658:training-compute","entityId":"epoch:f16fb17630ce9d321658","metric":"training-compute","value":2.2e+23,"unit":"FLOP","observationPeriod":"2022-06-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"It took us 65 days to train the model on a pool of 800 A100 graphics cards and 1.7 TB of online texts, books, and countless other sources.\""},{"id":"epoch:a0c8320d7538133f17f7:parameters","entityId":"epoch:a0c8320d7538133f17f7","metric":"parameters","value":20000000000.0,"unit":"parameters","observationPeriod":"2022-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Abstract: \"we achieve consistent quality improvements\nby scaling the encoder-decoder Transformer model up to 20B parameters\""},{"id":"epoch:a0c8320d7538133f17f7:training-compute","entityId":"epoch:a0c8320d7538133f17f7","metric":"training-compute","value":5.09607936e+23,"unit":"FLOP","observationPeriod":"2022-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Calculated from architecture. Does not take into account the encoding and decoding of text and images, only the transformer stack.\n\nTable 1 shows for the 20B model\n16 encoder layers\n64 decoder layers\nDmodel = 4096\nDhidden = 16384\nNum heads = 64\n\nJust below table 1:\n\"We use a maximum length of text tokens of 128, and the length of image tokens are fixed to 1024\"\n\nI take the length of the sequence to be 100 for the encoder stack and 1024 for the decoder stack.\n\nSection 3, Training: \"a total\nof 450,000 steps and final ratio of 0.025. We use a global batch size of 8192 during training.\"\n\n6* 20B parameters * (1024+128) sequence length*450000 steps*8192 batch size= 5.096079e+23"},{"id":"epoch:30bcbbcf121ab1168e12:parameters","entityId":"epoch:30bcbbcf121ab1168e12","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2022-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We introduce CodeGeeX, a large-scale multilingual code generation model with 13 billion parameters\""},{"id":"epoch:30bcbbcf121ab1168e12:training-compute","entityId":"epoch:30bcbbcf121ab1168e12","metric":"training-compute","value":6.630000000001e+22,"unit":"FLOP","observationPeriod":"2022-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assume 1 epoch on 850B tokens.\nC=6DN=6*850B*13B\nhttps://www.wolframalpha.com/input?i=6+*+13+billion+*+850+billion"},{"id":"epoch:1bc806f6fca763eb18c8:parameters","entityId":"epoch:1bc806f6fca763eb18c8","metric":"parameters","value":2700000000.0,"unit":"parameters","observationPeriod":"2022-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f17ec7fd1c3d2f3ffc5b:parameters","entityId":"epoch:f17ec7fd1c3d2f3ffc5b","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2022-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:dbaab1977f0f32ba0de3:parameters","entityId":"epoch:dbaab1977f0f32ba0de3","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2022-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:82527e608af0b64acdbe:parameters","entityId":"epoch:82527e608af0b64acdbe","metric":"parameters","value":2700000000.0,"unit":"parameters","observationPeriod":"2022-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:8d3df3c534965b013c48:parameters","entityId":"epoch:8d3df3c534965b013c48","metric":"parameters","value":125000000.0,"unit":"parameters","observationPeriod":"2022-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:3dafb903bb49a6034c51:parameters","entityId":"epoch:3dafb903bb49a6034c51","metric":"parameters","value":6700000000.0,"unit":"parameters","observationPeriod":"2022-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:3dafb903bb49a6034c51:training-compute","entityId":"epoch:3dafb903bb49a6034c51","metric":"training-compute","value":1.2060000000001e+22,"unit":"FLOP","observationPeriod":"2022-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.wolframalpha.com/input?i=6+FLOP+*+6.7+billion+*+300+billion\n\n4.3e+23 (175B exact compute) * 6.7B/175B = 1.6462857e+22"},{"id":"epoch:00eda65e5d1285b030d2:parameters","entityId":"epoch:00eda65e5d1285b030d2","metric":"parameters","value":66000000000.0,"unit":"parameters","observationPeriod":"2022-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:00eda65e5d1285b030d2:training-compute","entityId":"epoch:00eda65e5d1285b030d2","metric":"training-compute","value":1.100000000001e+23,"unit":"FLOP","observationPeriod":"2022-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"OPT-66B was trained for 140k steps, using a batch size of 2M tokens (see the OPT baselines logbook and Table 1 in Zhang et al. (2022), respectively), so training took 140e3 \u2217 2e6 \u2217 66e9 \u2217 6 = 1.1e23 FLOP"},{"id":"epoch:be5019511dfd72a295ea:parameters","entityId":"epoch:be5019511dfd72a295ea","metric":"parameters","value":350000000.0,"unit":"parameters","observationPeriod":"2022-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1791311955d86aee7b72:parameters","entityId":"epoch:1791311955d86aee7b72","metric":"parameters","value":2700000000.0,"unit":"parameters","observationPeriod":"2022-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:053bf3d6cabaec22d0d1:parameters","entityId":"epoch:053bf3d6cabaec22d0d1","metric":"parameters","value":125000000.0,"unit":"parameters","observationPeriod":"2022-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d69cdf3effda6ded5664:parameters","entityId":"epoch:d69cdf3effda6ded5664","metric":"parameters","value":30000000000.0,"unit":"parameters","observationPeriod":"2022-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d69cdf3effda6ded5664:training-compute","entityId":"epoch:d69cdf3effda6ded5664","metric":"training-compute","value":5.4000000000001e+22,"unit":"FLOP","observationPeriod":"2022-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.wolframalpha.com/input?i=6+FLOP+*+30+billion+*+300+billion"},{"id":"epoch:207ebbda5a2f74981cf0:parameters","entityId":"epoch:207ebbda5a2f74981cf0","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2022-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0d9cd359da6a4494b07c:parameters","entityId":"epoch:0d9cd359da6a4494b07c","metric":"parameters","value":2925000000.0,"unit":"parameters","observationPeriod":"2022-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2952M, Table 2"},{"id":"epoch:0d9cd359da6a4494b07c:training-compute","entityId":"epoch:0d9cd359da6a4494b07c","metric":"training-compute","value":3.5e+21,"unit":"FLOP","observationPeriod":"2022-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1M steps, batch size 1024. Sequence length may be 128-256:\n\"We use a maximum of 256 and 128 text tokens for inputs and outputs respectively, and a maximum\nlength of 576 (i.e. 24 \u00d7 24 patch encoding from a 384 \u00d7 384 image) for image inputs and 256 (i.e.\n16 \u00d7 16 latent codes from a 256 \u00d7 256 image) for image outputs\n\n6 * 1 million * 1024 * 128 * 2.9 billion = 2.3e21\n6 * 1 million * 1024 * 256 * 2.9 billion = 4.6e21\naverage is 3.5e21\n\nNo hardware details."},{"id":"epoch:cd4726333c93d4902cc1:parameters","entityId":"epoch:cd4726333c93d4902cc1","metric":"parameters","value":2100000000.0,"unit":"parameters","observationPeriod":"2022-06-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our largest CoCa model (\"CoCa\" in short) follows the ViT-giant setup in [21] with 1B-parameters in the image encoder and 2.1B-parameters altogether with the text decoder\""},{"id":"epoch:cd4726333c93d4902cc1:training-compute","entityId":"epoch:cd4726333c93d4902cc1","metric":"training-compute","value":7.3e+22,"unit":"FLOP","observationPeriod":"2022-06-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Pretraining CoCa takes about 5 days on 2,048 CloudTPUv4 chips\"\n\n275 teraFLOP/s * 2048 * 5 * 24 * 3600 * 0.3 (assumed utilization) = 7.3e22"},{"id":"epoch:4665707e50f970a601bd:parameters","entityId":"epoch:4665707e50f970a601bd","metric":"parameters","value":74000000.0,"unit":"parameters","observationPeriod":"2022-06-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"74M\nTable 3\n"},{"id":"epoch:4665707e50f970a601bd:training-compute","entityId":"epoch:4665707e50f970a601bd","metric":"training-compute","value":2.22e+18,"unit":"FLOP","observationPeriod":"2022-06-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 74000000 parameters * 2000000 tokens * 2500 epochs = 2.22 \u00d7 10^18 FLOP"},{"id":"epoch:fb38a29c2ee0fd140456:parameters","entityId":"epoch:fb38a29c2ee0fd140456","metric":"parameters","value":55000000.0,"unit":"parameters","observationPeriod":"2022-06-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 3\n"},{"id":"epoch:c1c86d4b1d5403910796:parameters","entityId":"epoch:c1c86d4b1d5403910796","metric":"parameters","value":137000000000.0,"unit":"parameters","observationPeriod":"2022-06-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"137B. Table App.1\n"},{"id":"epoch:c1c86d4b1d5403910796:training-compute","entityId":"epoch:c1c86d4b1d5403910796","metric":"training-compute","value":5.6e+23,"unit":"FLOP","observationPeriod":"2022-06-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"BIG-G models were trained at Google. We use 13 dense decoder-only Transformer models (Vaswani\net al., 2017) with gated activation layers (Dauphin et al., 2017) and GELU activations based on the LaMDA\narchitectures (Thoppilan et al., 2022). These models were trained on a dataset consisting of a mixture of web\ndocuments, code, dialog, and Wikipedia data, with approximately three billion documents tokenized to 2.8\ntrillion BPE tokens using a 32k-token SentencePiece vocabulary\"\n\nAppendix:\n\n\"We use a pre-training batch size of 262k tokens for all models...\"\n\n2.6M steps for 137B, per Table App.1. So trained on 2.6M * 262k = 681B tokens (~0.25 epochs)\n681B * 137B * 6 = 5.6e23"},{"id":"epoch:b81b0a6172e55cb0e601:parameters","entityId":"epoch:b81b0a6172e55cb0e601","metric":"parameters","value":5600000000.0,"unit":"parameters","observationPeriod":"2022-06-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 1: \"We scale this up to a large 5.6B parameter LIMoE-H/14\""},{"id":"epoch:b81b0a6172e55cb0e601:training-compute","entityId":"epoch:b81b0a6172e55cb0e601","metric":"training-compute","value":1.8e+22,"unit":"FLOP","observationPeriod":"2022-06-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 3.2: \"The model contains 5.6B parameters in total, but only applies 675M parameters per token\"\n\nFrom Section A.3, \"batch size 21502 with resolution 288 and text sequence length16\". \"The model was trained for 700k steps pre-cooldown. There was one cooldown of length 125k steps\nfrom the final step, and 3 of length 40k steps starting from step 650k\". Patch size 14 for images.\n\nAssume C = 6*N*D. \nC = 6*675e6*21.5e3*1e6*(16+(288/14)**2)/2 = 1.8e22\n\nThis is broadly consistent with ViT-H/14's compute\n\n\nupd Figure 11(c) suggests 1.4*10^4*10^9 FLOPS = 1.4*10^13 FLOPS"},{"id":"epoch:0d30cb418d18f6e9c93e:parameters","entityId":"epoch:0d30cb418d18f6e9c93e","metric":"parameters","value":750000000.0,"unit":"parameters","observationPeriod":"2022-06-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train a Transformer model (750M parameters, similar to GPT-2 Large)\"\n\n\"Specifically, we use a 16-layer Transformer with 8 attention heads, hidden size 1024 and fully-connected dimension 4096.\""},{"id":"epoch:0d30cb418d18f6e9c93e:training-compute","entityId":"epoch:0d30cb418d18f6e9c93e","metric":"training-compute","value":3.31776e+18,"unit":"FLOP","observationPeriod":"2022-06-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 160000 steps * 3 samples per batch * 1536 tokens per sample *  750000000 parameters = 3.31776 \u00d7 10^18 FLOP"},{"id":"epoch:849c434615f20f1e8ef0:training-compute","entityId":"epoch:849c434615f20f1e8ef0","metric":"training-compute","value":2.8e+19,"unit":"FLOP","observationPeriod":"2022-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"192*7 GPU (P100) hours per Table 6\n\n19 TFLOP/s \n\n19 trillion * 192 * 7 * 3600 * 0.3 = 2.76e19\n\n6 FLOP / parameter / token * 247000000  parameters [not sure how it was estimated] * 50000 steps * 1024 tokens / GPU * 192 GPUs = 1.4568653e+19 FLOP "},{"id":"epoch:9443ebb361c485a88e2b:parameters","entityId":"epoch:9443ebb361c485a88e2b","metric":"parameters","value":9400000000.0,"unit":"parameters","observationPeriod":"2022-05-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:aca0e4863f30a42cdff0:parameters","entityId":"epoch:aca0e4863f30a42cdff0","metric":"parameters","value":700000000.0,"unit":"parameters","observationPeriod":"2022-05-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our largest transformer model, Tranception L, has 700M parameters and is trained on UniRef100 (Suzek et al., 2014)\""},{"id":"epoch:aca0e4863f30a42cdff0:training-compute","entityId":"epoch:aca0e4863f30a42cdff0","metric":"training-compute","value":7.24e+21,"unit":"FLOP","observationPeriod":"2022-05-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Trained using 64 A100 GPUs for two weeks.\n64 * 312 teraFLOP/s * 14 days * 24 hours/day * 3600 seconds/hour * 0.3 utilization (assumption)\n= 7.24e21"},{"id":"epoch:455f19bbf6eeba9098e0:parameters","entityId":"epoch:455f19bbf6eeba9098e0","metric":"parameters","value":355000000.0,"unit":"parameters","observationPeriod":"2022-05-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"GPT-2 Medium from here, aka GPT-2 355M: https://huggingface.co/openai-community/gpt2-medium"},{"id":"epoch:455f19bbf6eeba9098e0:training-compute","entityId":"epoch:455f19bbf6eeba9098e0","metric":"training-compute","value":8.9280922e+20,"unit":"FLOP","observationPeriod":"2022-05-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.272e+21 (GPT-2 355M). 3x speedup claimed over the original implementation, but not clear how this interacts with overall training compute.\n\nAlternative data for calculation: 8x A100 80GB * 165.6 hours at 60% utilization (higher than average because FlashAttention-2 paper reports 72% utilization with larger models and FlashAttention-1 runs ~95% as fast as FA-2 on the smallest model then tested: GPT3-1.3B with 2K context, whereas this paper uses GPT2-355M with 1K context. Paper uses mixed precision training for GPT-2 Medium (PyTorch AMP). \n\n\"We train the model on 8\u00d7A100-40GB GPUs, and we measure the wall-clock training time. Training\nGPT-2 small takes between 2.7-9.5 days, and training GPT-2 medium takes between 6.9-21.0 days (Table 2).\"\n\n8 GPUs * 312000000000000 FLOP / GPU / sec * 165.6 hours * 3600 sec / hour * 0.6 [assumed utilization] = 8.9280922e+20 FLOP"},{"id":"epoch:6e60f949cb3221bbeff5:parameters","entityId":"epoch:6e60f949cb3221bbeff5","metric":"parameters","value":247000000.00000003,"unit":"parameters","observationPeriod":"2022-05-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6e60f949cb3221bbeff5:training-compute","entityId":"epoch:6e60f949cb3221bbeff5","metric":"training-compute","value":3.12e+19,"unit":"FLOP","observationPeriod":"2022-05-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 9: \n73728 tokens per update * 286000 updates =21086208000 tokens\n\nUsing 6ND formula:\n6 * 247000000 * 21086208000 = 3.124976e+19"},{"id":"epoch:5712906bbaf999ecdfaa:parameters","entityId":"epoch:5712906bbaf999ecdfaa","metric":"parameters","value":7000000.0,"unit":"parameters","observationPeriod":"2022-05-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5712906bbaf999ecdfaa:training-compute","entityId":"epoch:5712906bbaf999ecdfaa","metric":"training-compute","value":2.06e+17,"unit":"FLOP","observationPeriod":"2022-05-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"24576 tokens per update *  200000 updates = 4915200000 tokens\n\nUsing 6ND formula:\n6*7000000*4915200000=2.064384e+17"},{"id":"epoch:391ff1f0e0910a82008b:parameters","entityId":"epoch:391ff1f0e0910a82008b","metric":"parameters","value":150000000.0,"unit":"parameters","observationPeriod":"2022-05-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:391ff1f0e0910a82008b:training-compute","entityId":"epoch:391ff1f0e0910a82008b","metric":"training-compute","value":6.48e+18,"unit":"FLOP","observationPeriod":"2022-05-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 9: \n36000 tokens per update * 200000 updates = 7200000000 tokens\n\nUsing 6ND formula:\n6*7200000000*150000000=6.48e+18"},{"id":"epoch:a604f87cfe5165970181:parameters","entityId":"epoch:a604f87cfe5165970181","metric":"parameters","value":7762000000.0,"unit":"parameters","observationPeriod":"2022-05-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2B 64x64 generation model, 600M 64->256 super-resolution model, 400M 256->1024 super-resolution model\nUses encodings from a frozen T5-XXL, which should be included in total parameter count. Loading the model directly, there are 4,762,310,656 parameters in the encoder.\n2B + 4.762B + 600M + 400M = 7.762 billion\n\nhere they claim it is 3B parameters: https://arxiv.org/pdf/2407.15811"},{"id":"epoch:a604f87cfe5165970181:training-compute","entityId":"epoch:a604f87cfe5165970181","metric":"training-compute","value":1.46e+22,"unit":"FLOP","observationPeriod":"2022-05-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"256 TPU v4 chips for 64x64, for 4 days\n128 TPU v4 chips for 64->256, for 2 days\n128 TPU v4 chips for 256->1024, for 2 days\n\n256 TPUs * 275 teraFLOPS/TPU * 4 days + 2 * (128 TPUs * 275 teraFLOPS/TPU * 2 days) * 40% utilization = 1.46e+22 FLOP"},{"id":"epoch:bc3e63ceb2c4e755b188:parameters","entityId":"epoch:bc3e63ceb2c4e755b188","metric":"parameters","value":1180000000.0,"unit":"parameters","observationPeriod":"2022-05-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"This section focuses on in-simulation evaluation.\nFigure 10 compares the full 1.18B parameter Gato\" p.10"},{"id":"epoch:bc3e63ceb2c4e755b188:training-compute","entityId":"epoch:bc3e63ceb2c4e755b188","metric":"training-compute","value":4.02e+21,"unit":"FLOP","observationPeriod":"2022-05-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"256 (16x16x) TPUv3 chips x 123e12 FLOPS/chip x 4 days x 86400 seconds/day * 0.4 utilization = 4.35e21 FLOPs\n\nSimilar value by 6NC:\n6 * 524288000000 * 1.18B = 3.71e21\n\nUsing geometric mean:\nsqrt(4.35e21 * 3.71e21) = 4.02e21"},{"id":"epoch:ee5e1bf9814e2f3bbf97:parameters","entityId":"epoch:ee5e1bf9814e2f3bbf97","metric":"parameters","value":20000000000.0,"unit":"parameters","observationPeriod":"2022-05-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Taken from Directory of LLMs"},{"id":"epoch:ee5e1bf9814e2f3bbf97:training-compute","entityId":"epoch:ee5e1bf9814e2f3bbf97","metric":"training-compute","value":1.2e+23,"unit":"FLOP","observationPeriod":"2022-05-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Trained on 1T tokens\n20B * 1T * 6 = 1.2e23 \n\nSecond source: Section 5.1 says model was trained on 512 TPUv4 chips, and took slightly over 1 month\n512 * 2.75e14 * 31 * 24 * 3600 * 0.3 = 1.13e23"},{"id":"epoch:7df68d8ef5e5dd3aaa68:training-compute","entityId":"epoch:7df68d8ef5e5dd3aaa68","metric":"training-compute","value":4.4928e+19,"unit":"FLOP","observationPeriod":"2022-05-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training was done using the Isaac Gym simulator on an NVIDIA V100 GPU. The model was trained on over 10 billion samples, which equates to 10 years of simulated experience time. Training took around 10 days on a single GPU.\n1.3e14 * 10 * 24 * 3600 * 0.4 = 4.49e19"},{"id":"epoch:04c42849544903e122e3:training-compute","entityId":"epoch:04c42849544903e122e3","metric":"training-compute","value":1.296e+21,"unit":"FLOP","observationPeriod":"2022-05-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\" to match the prior state-of-the-art performance of ADM [Dhariwal and Nichol 2021] at a resolution of 512^2 pixels,\ntraining the models on a single NVIDIA Tesla V100 takes 400 days compared to the previously required 1914 V100-days.\"\n\n125000000000000 FLOP / GPU / sec * 9600 GPU-hours [for a 512 resolution SOTA result] * 3600 sec / hour * 0.3 [assumed utilization] = 1.296e+21 FLOP"},{"id":"epoch:249426959eae9bc3c55b:parameters","entityId":"epoch:249426959eae9bc3c55b","metric":"parameters","value":418000000.0,"unit":"parameters","observationPeriod":"2022-05-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"DeBERTaV3-large had 418M params, per Table 2"},{"id":"epoch:0c0b5e91383cf9f57910:parameters","entityId":"epoch:0c0b5e91383cf9f57910","metric":"parameters","value":7000000000.0,"unit":"parameters","observationPeriod":"2022-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2930a517af00d1acab2b:parameters","entityId":"epoch:2930a517af00d1acab2b","metric":"parameters","value":175000000000.0,"unit":"parameters","observationPeriod":"2022-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In line with Meta AI\u2019s commitment to open science, we are sharing Open Pretrained Transformer (OPT-175B), a language model with 175 billion parameters trained on publicly available data sets\""},{"id":"epoch:2930a517af00d1acab2b:training-compute","entityId":"epoch:2930a517af00d1acab2b","metric":"training-compute","value":4.3e+23,"unit":"FLOP","observationPeriod":"2022-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://github.com/facebookresearch/metaseq/blob/main/projects/OPT/chronicles/final_update.md\n\n\"As of yesterday, at 12:46pm PST on January 6, our 175B model finally completed its training run on 300B tokens. This required ~4.30E+23 FLOPs of compute\""},{"id":"epoch:6f2177b03db29ef2b9c7:parameters","entityId":"epoch:6f2177b03db29ef2b9c7","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2022-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6f2177b03db29ef2b9c7:training-compute","entityId":"epoch:6f2177b03db29ef2b9c7","metric":"training-compute","value":3.53e+23,"unit":"FLOP","observationPeriod":"2022-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2851200*992*312e12*0.40 = 3.53e23 assuming 40% utilization"},{"id":"epoch:2e834940b727713e5259:parameters","entityId":"epoch:2e834940b727713e5259","metric":"parameters","value":80000000000.0,"unit":"parameters","observationPeriod":"2022-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We obtain three models, Flamingo-3B, Flamingo-9B and Flamingo-80B\"\n\n\" The Flamingo-80B model builds on top of the frozen Chinchilla 70B language model [42]. Starting from the very first layer and before every seventh transformer blocks, we add a GATED XATTN-DENSE layer attending to the visual inputs; this accounts for 10B additional learned parameters. For simplicity, we refer to this model as simply Flamingo throughout the paper\""},{"id":"epoch:2e834940b727713e5259:training-compute","entityId":"epoch:2e834940b727713e5259","metric":"training-compute","value":2.18972000000001e+23,"unit":"FLOP","observationPeriod":"2022-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1536 TPU v4 chips for 15 days. Assuming 40% utilization:\nC = 1536 TPU * 275*10^12 FLOP/s/TPU * 15 day * 86400 s/day * 0.40 = 2.2*10^23 FLOP\n\n\"All training and evaluation was performed on TPUv4 instances. The largest model containing 80 billion parameters is trained on QUSV chips for 15 days and sharded across 16 devices.\"\n\n\"All trained parameters and optimizer accumulators are stored and updated in float32; all activations and gradients are computed in bfloat16 after downcasting of parameters from float32 to bfloat16\""},{"id":"epoch:22a73e38ce0900f66388:parameters","entityId":"epoch:22a73e38ce0900f66388","metric":"parameters","value":6000000000.0,"unit":"parameters","observationPeriod":"2022-04-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The backbone of our pretrained CogLM is a Transformer with Sandwich LayerNorm [3]. The model has 6 billion parameters (48 layers, hidden size 3072, 48 attention heads)"},{"id":"epoch:22a73e38ce0900f66388:training-compute","entityId":"epoch:22a73e38ce0900f66388","metric":"training-compute","value":2.265e+22,"unit":"FLOP","observationPeriod":"2022-04-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The model has 6 billion parameters (48 layers, hidden size 3072, 48 attention heads), trained for 300,000 iterations in FP16 with batch size 4,096\"\n\nUsing 6ND formula:\n\n6*6000000000*300000*4096*512=2.2649242e+22"},{"id":"epoch:005cad56b62fa71d2efe:parameters","entityId":"epoch:005cad56b62fa71d2efe","metric":"parameters","value":151000000.0,"unit":"parameters","observationPeriod":"2022-04-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:b80051bcf677becfeb1d:parameters","entityId":"epoch:b80051bcf677becfeb1d","metric":"parameters","value":9410000000.0,"unit":"parameters","observationPeriod":"2022-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2: \"In Section 4.4, we run our large model (9.41B parameters)\""},{"id":"epoch:b80051bcf677becfeb1d:training-compute","entityId":"epoch:b80051bcf677becfeb1d","metric":"training-compute","value":5.32224e+20,"unit":"FLOP","observationPeriod":"2022-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"112 hours on 32 V100 GPUs\nassumed 0.33 util rate\n\n112 hours *3600 seconds / hour *0.33 utilization *32 gpus *125000000000000 FLOPs=532224000000000000000\n"},{"id":"epoch:8f11ee0234bae2fbb839:parameters","entityId":"epoch:8f11ee0234bae2fbb839","metric":"parameters","value":1450000000.0,"unit":"parameters","observationPeriod":"2022-04-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 2"},{"id":"epoch:8f11ee0234bae2fbb839:training-compute","entityId":"epoch:8f11ee0234bae2fbb839","metric":"training-compute","value":5.0000000000000004e+22,"unit":"FLOP","observationPeriod":"2022-04-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"I get 5e22 FLOP. 150k hours on A100 [1] gives 150*10^3 hours * 3600 seconds/hour * 3.12E+14 peak performance of A100 * 0.33 utilisation = 5e22  FLOP\"\n\n[1] https://twitter.com/EMostaque/status/1563870674111832066"},{"id":"epoch:dddec47b69d8904eac84:parameters","entityId":"epoch:dddec47b69d8904eac84","metric":"parameters","value":52000000000.0,"unit":"parameters","observationPeriod":"2022-04-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2f105a2ec7a24f81aa41:parameters","entityId":"epoch:2f105a2ec7a24f81aa41","metric":"parameters","value":600000000.0,"unit":"parameters","observationPeriod":"2022-04-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"600M"},{"id":"epoch:2ab0e32abfeeed92b5e0:parameters","entityId":"epoch:2ab0e32abfeeed92b5e0","metric":"parameters","value":110000000.0,"unit":"parameters","observationPeriod":"2022-04-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Base model is BERT base (110M parameters), pre-trained on human reference genome (DNABert: https://academic.oup.com/bioinformatics/article/37/15/2112/6128680)"},{"id":"epoch:2ab0e32abfeeed92b5e0:training-compute","entityId":"epoch:2ab0e32abfeeed92b5e0","metric":"training-compute","value":1.4e+20,"unit":"FLOP","observationPeriod":"2022-04-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See DNABert entry:\n\n\"Since the pre-training of DNABERT model is resource-intensive (about 25\u2009days on 8 NVIDIA 2080Ti GPUs)\"\n\nAssuming FP16 and 30% utilization\n\nCalculation = (25 * 24 *3600) s * 2.7e13 FLOP/s per GPU * 8 GPUs * 0.3 utilization = 1.4e20 FLOP"},{"id":"epoch:fa4f8c70f3a596441ad1:parameters","entityId":"epoch:fa4f8c70f3a596441ad1","metric":"parameters","value":3500000000.0,"unit":"parameters","observationPeriod":"2022-04-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our decoder architecture is the 3.5 billion parameter GLIDE model\""},{"id":"epoch:fa4f8c70f3a596441ad1:training-compute","entityId":"epoch:fa4f8c70f3a596441ad1","metric":"training-compute","value":3.3695784e+23,"unit":"FLOP","observationPeriod":"2022-04-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Decoder architecture is similar to Imagen (1.46E+22), but trained on 1.6e9 datapoints (Table 3) rather than Imagen's 5.1e9 datapoints.\n\nDALL-E 2 uses two models as priors. I estimate the prior model's FLOP as 6*N*D = 6 * 1e9 * 4096 * 1e6 = 2.5e19 FLOP. However, this seems low compared to CLIP.\n\nSo it may be possible to estimate DALL-E 2's compute by analogy to Imagen, but there is a lot of uncertainty and more research would be needed.\n\nhere (https://arxiv.org/pdf/2407.15811) they claim the DALL-E.2 model was trained on the equivalent of 5208.3 days on 8*A100 GPUs:\n\n312000000000000 FLOP / sec / GPU * 8 GPUs * 5208.3 days * 24 hours / day * 3600 sec / hour * 0.3 [assumed utilization] = 3.3695784e+23 FLOP"},{"id":"epoch:44d3df064dc368229df4:parameters","entityId":"epoch:44d3df064dc368229df4","metric":"parameters","value":540350000000.0,"unit":"parameters","observationPeriod":"2022-04-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"To further our understanding of the impact of scale on few-shot learning, we trained a 540-billion parameter, densely activated, Transformer language model, which we call Pathways Language Model (PaLM).\""},{"id":"epoch:44d3df064dc368229df4:training-compute","entityId":"epoch:44d3df064dc368229df4","metric":"training-compute","value":2.5272e+24,"unit":"FLOP","observationPeriod":"2022-04-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 20.\n\n6144 TPUv4 for 1200 hours + 3072 TPUv4 for 336 hours.\nEquivalent to 6144 TPUv4 for 1368 hours.\n\n46.2% model FLOPs utilization\n\n\"The 540B-parameter PaLM model sustained a remarkable 57.8% of the peak hardware floating point performance over 50 days while training on TPU v4 supercomputers.\" https://cloud.google.com/blog/topics/systems/tpu-v4-enables-performance-energy-and-co2e-efficiency-gains"},{"id":"epoch:cee1b982275b6624c509:parameters","entityId":"epoch:cee1b982275b6624c509","metric":"parameters","value":165000000.0,"unit":"parameters","observationPeriod":"2022-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"165M\ntable 2\n\n\"For dense models, we use standard implementations of GPT-2 [86] from Huggingface transformers library\nand from Nvidia\u2019s Megatron-LM repo. We follow the training recipe of the Megatron-LM repo.\nThe Monarch version of these models simply swap out the dense weight matrices in the attention blocks\n(projection matrices) and in the FFN block (linear layers) with Monarch matrices.\""},{"id":"epoch:cee1b982275b6624c509:training-compute","entityId":"epoch:cee1b982275b6624c509","metric":"training-compute","value":4.36e+20,"unit":"FLOP","observationPeriod":"2022-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use an effective batch size of 512\"\nGPT-2-Medium: 10+100 epochs on wikitext-103 + 4k/400k iterations on openwebtext \n\n6 FLOP / parameter / token * 165 * 10^6 parameters * (103000000 tokens * 110 epochs + 512 * 404000) = 1.142147952 \u00d7 10^19 FLOP\n"},{"id":"epoch:0f68ec7229da0e3a7cb8:parameters","entityId":"epoch:0f68ec7229da0e3a7cb8","metric":"parameters","value":72000000.0,"unit":"parameters","observationPeriod":"2022-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"GPPT-2 small has 124M parameters: https://huggingface.co/openai-community/gpt2\n\nupd see Table 2"},{"id":"epoch:0f68ec7229da0e3a7cb8:training-compute","entityId":"epoch:0f68ec7229da0e3a7cb8","metric":"training-compute","value":9e+19,"unit":"FLOP","observationPeriod":"2022-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"400k iterations with batch size 512 (Appendix E)\nGPT 2 small on Huggingface has a context length of 1024\nFLOP estimate using 6ND: \n6*400000*512*1024*72000000=9.0596966e+19"},{"id":"epoch:167f9b008e46d9932287:parameters","entityId":"epoch:167f9b008e46d9932287","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2022-03-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.3B\n\n\"The baseline model in this setting follows the 1.3B parameter architecture of Brown et al. (2020), also known as GPT-3 XL: 24 transformer layers with 2048 model dimensions, 8192 feed-forward dimensions, and 32 attention heads.\""},{"id":"epoch:167f9b008e46d9932287:training-compute","entityId":"epoch:167f9b008e46d9932287","metric":"training-compute","value":2.09664e+19,"unit":"FLOP","observationPeriod":"2022-03-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 1.3 * 10^9 parameters * 256 000 tokens per batch [table 5] * 10500 steps [table 5] = 2.09664e+19 FLOP "},{"id":"epoch:e2e066fffe636ddf8ae0:parameters","entityId":"epoch:e2e066fffe636ddf8ae0","metric":"parameters","value":70000000000.0,"unit":"parameters","observationPeriod":"2022-03-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We test this hypothesis by training a predicted compute-optimal model, \\chinchilla, that uses the same compute budget as \\gopher but with 70B parameters and 4\u00d7 more more data. \\chinchilla uniformly and significantly outperforms \\Gopher (280B), GPT-3 (175B), Jurassic-1 (178B), and Megatron-Turing NLG (530B) on a large range of downstream evaluation tasks.\""},{"id":"epoch:e2e066fffe636ddf8ae0:training-compute","entityId":"epoch:e2e066fffe636ddf8ae0","metric":"training-compute","value":5.76e+23,"unit":"FLOP","observationPeriod":"2022-03-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Both Chinchilla and Gopher have been trained for the same number of FLOPs but differ in the size of the model and the number of training tokens.\"\n\nWe see the number of flops in table 3"},{"id":"epoch:5252c7291a038325731e:parameters","entityId":"epoch:5252c7291a038325731e","metric":"parameters","value":173900000000000.0,"unit":"parameters","observationPeriod":"2022-03-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 3, MoDa-174T has 173.9 trillion parameters"},{"id":"epoch:6ad06a972ade9738d44b:training-compute","entityId":"epoch:6ad06a972ade9738d44b","metric":"training-compute","value":3.7999999999999995e+18,"unit":"FLOP","observationPeriod":"2022-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 2x TITAN Xp GPUs (1.10\u00d710\u00b9\u00b3 FLOP/s each)\n2. Training duration: Estimated 5 days (432,000 seconds) based on \"several days\" mention -> \"Likely\" confidence\n3. Utilization rate: 40%\n4. Calculation: 2 GPUs \u00d7 1.10\u00d710\u00b9\u00b3 FLOP/s \u00d7 432,000s \u00d7 0.40 = 3.8\u00d710\u00b9\u2078 FLOPs"},{"id":"epoch:a31c72bc0122c09aed2c:parameters","entityId":"epoch:a31c72bc0122c09aed2c","metric":"parameters","value":4000000000.0,"unit":"parameters","observationPeriod":"2022-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Experiments were performed with a 4 billion parameter\ntransformer, generating a sequence of 256 text tokens, 256\nscene tokens, and 1024 image tokens, that are then decoded\ninto an image with a resolution of 256 \u00d7 256 or 512 \u00d7 512\npixels (depending on the model of choice)\""},{"id":"epoch:a31c72bc0122c09aed2c:training-compute","entityId":"epoch:a31c72bc0122c09aed2c","metric":"training-compute","value":6.4172851e+21,"unit":"FLOP","observationPeriod":"2022-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 4 * 10^9 parameters * 267386880000 tokens [see dataset size notes] = 6.4172851e+21 FLOP"},{"id":"epoch:1eea970277a5941d7c68:parameters","entityId":"epoch:1eea970277a5941d7c68","metric":"parameters","value":357000000.0,"unit":"parameters","observationPeriod":"2022-03-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"357M\nTable 3\n\n\"Following Kasai et al. (2021), we choose similar hyperparameters to prior work (Baevski and Auli, 2019; Fan et al., 2020): 32 layers, 8 heads, 128 head dimensions, 1024 model dimensions, 4096 fully connected dimensions and\ndropout (Srivastava et al., 2014) and layer dropout rates of 0.2. \""},{"id":"epoch:1eea970277a5941d7c68:training-compute","entityId":"epoch:1eea970277a5941d7c68","metric":"training-compute","value":7.3e+18,"unit":"FLOP","observationPeriod":"2022-03-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 357000000 parameters * 103000000 tokens * 1 epoch [assumed for lower bound] = 2.20626e+17 FLOP\n\nalso to consider: \n 7.3 \u00d7 10^18 FLOP \nSOURCE: Impute based on Baevski and Auli 2019\n\n\"We generally follow the optimization method from Baevski and Auli (2019), with a slight modification for some hyperparameters including learning rate (we use 10\u22124), which shows better convergence. \"\n\nNOTES: Probably correct within an order of 2 or so"},{"id":"epoch:cdd96070786b49b27917:parameters","entityId":"epoch:cdd96070786b49b27917","metric":"parameters","value":256999999.99999997,"unit":"parameters","observationPeriod":"2022-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"257M\n(Table 1)\n\n18 layers, 16 heads, hidden size\n1024 batch size 128."},{"id":"epoch:cdd96070786b49b27917:training-compute","entityId":"epoch:cdd96070786b49b27917","metric":"training-compute","value":2.65e+19,"unit":"FLOP","observationPeriod":"2022-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 257000000 parameters * 384 tokens/batch \u00d7 128 [batch size] * 350000 steps = 2.6527334e+19 FLOP"},{"id":"epoch:065f162a533c15e5eb3e:parameters","entityId":"epoch:065f162a533c15e5eb3e","metric":"parameters","value":256999999.99999997,"unit":"parameters","observationPeriod":"2022-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"257M"},{"id":"epoch:065f162a533c15e5eb3e:training-compute","entityId":"epoch:065f162a533c15e5eb3e","metric":"training-compute","value":6.06e+18,"unit":"FLOP","observationPeriod":"2022-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"80k steps (Table 1)\n\"The input lengths are 150 for the base model and 384 for the large model.\"\n\"large model: 18 layers, 16 heads, hidden size 1024 batch size 128\"\n\n6 FLOP / token / parameter * 257000000 parameters * 384 tokens per sequence * 128 sequences per batch * 80000 steps = 6.06339072 \u00d7 10^18 FLOP "},{"id":"epoch:bf837f572e35e63ba6e1:parameters","entityId":"epoch:bf837f572e35e63ba6e1","metric":"parameters","value":151000000.0,"unit":"parameters","observationPeriod":"2022-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1 reports WikiText results, Table 2 arxiv. WikiText experiments use the larger model of the two."},{"id":"epoch:bf837f572e35e63ba6e1:training-compute","entityId":"epoch:bf837f572e35e63ba6e1","metric":"training-compute","value":1.74e+18,"unit":"FLOP","observationPeriod":"2022-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"base model: 16 layers, 10 heads, hidden size\n410, batch size 64, training steps 200k;\"\n\n\"The input lengths are 150 for the base model\nand 384 for the large model. \"\n\n6ND estimate:\n6*200000*64*150*151000000=1.73952e+18"},{"id":"epoch:d27a26a0a0b4d919a5d5:training-compute","entityId":"epoch:d27a26a0a0b4d919a5d5","metric":"training-compute","value":2.578e+24,"unit":"FLOP","observationPeriod":"2022-03-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://colab.research.google.com/drive/1QSxa8YCWjEBQU7mrXLhw6TP1VX5oqgdW#scrollTo=Gt6Z6oZ26clI"},{"id":"epoch:cee9ff95fb281a458fcb:parameters","entityId":"epoch:cee9ff95fb281a458fcb","metric":"parameters","value":1843000000.0,"unit":"parameters","observationPeriod":"2022-03-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"This is from the original ViT-G paper"},{"id":"epoch:cee9ff95fb281a458fcb:training-compute","entityId":"epoch:cee9ff95fb281a458fcb","metric":"training-compute","value":3.4e+21,"unit":"FLOP","observationPeriod":"2022-03-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"This is a fine-tuned version of ViT-G, which required 3.4e21 to train per PCD/Akronomicon.\n\nFine-tuning compute is likely minor in comparision:\n\"Models are fine-tuned at a batch size of 512 for either 10,000 or 20,000 steps (approximately 4 or 8 epochs)... all models are fine-tuned at 518 \u00d7 518 resolution\"\nAt 20k steps, we have (518^2) * 512 * 20k = 2.75e12 pixels seen in fine-tuning, compared to (224^2) * 32768 * 5M = 8.22e15 in pre-training."},{"id":"epoch:69861de1b89254a54c8f:parameters","entityId":"epoch:69861de1b89254a54c8f","metric":"parameters","value":6700000000.0,"unit":"parameters","observationPeriod":"2022-03-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:69861de1b89254a54c8f:training-compute","entityId":"epoch:69861de1b89254a54c8f","metric":"training-compute","value":1.28e+22,"unit":"FLOP","observationPeriod":"2022-03-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we outperform published numbers of the 6.7B GPT-3 model, with\ntuning cost only 7% of total pretraining cost.\"\n\nGPT-3 6.7B reported training compute is 1.2e+22 FLOP \n\n1.2e+22 FLOP * 1.07 = 1.284e+22 FLOP\n\n6 FLOP / parameter / token * 6660000000 parameters * 300000000000 tokens = 1.1988 \u00d7 10^22 FLOP\n\n"},{"id":"epoch:98d19503c689865e03cf:parameters","entityId":"epoch:98d19503c689865e03cf","metric":"parameters","value":612000000.0,"unit":"parameters","observationPeriod":"2022-03-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"612M from Table 5 - LSUN-cat row\nwe provide details for LSUN-cat with largest compute"},{"id":"epoch:98d19503c689865e03cf:training-compute","entityId":"epoch:98d19503c689865e03cf","metric":"training-compute","value":2.9113344e+18,"unit":"FLOP","observationPeriod":"2022-03-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use eight NVIDIA A100 GPUs to train RQ-Transformer of 1.4B parameters, and four GPUs to trainRQ-Transformers of other sizes. The training time is <9 days for LSUN-cat, LSUN-bedroom, <4.5 days for ImageNet, and CC-3M, and <1 day for LSUN-church and FFHQ. \"\nwe provide details for LSUN-cat with largest compute\n\nflops = (4) * (3.12e14) * (9*24 * 3600) * (0.3) = 2.9113344e+20\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)"},{"id":"epoch:183d5ecfc64912f5e214:parameters","entityId":"epoch:183d5ecfc64912f5e214","metric":"parameters","value":1388000000.0,"unit":"parameters","observationPeriod":"2022-03-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1388M from Table 5 - ImageNet rows"},{"id":"epoch:183d5ecfc64912f5e214:training-compute","entityId":"epoch:183d5ecfc64912f5e214","metric":"training-compute","value":1.05670656e+20,"unit":"FLOP","observationPeriod":"2022-03-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use eight NVIDIA A100 GPUs to train RQ-Transformer of 1.4B parameters, and four GPUs to train RQ-Transformers of other sizes. The training time is <9 days for LSUN-cat, LSUN-bedroom, <4.5 days for ImageNet, and CC-3M, and <1 day for LSUN-church and FFHQ. \"\n\nQuote probably involves a typo \u2013 I expect 8 A100s were used for 4.5 days to train the 3.8B model. 1.4B model probably used 4 GPUs.\n\nBest guess: (4) * (3.12e14) * (78.4 * 3600) * (0.3) = 1.057e20\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate) "},{"id":"epoch:2a351c15436bb988420e:parameters","entityId":"epoch:2a351c15436bb988420e","metric":"parameters","value":3822000000.0,"unit":"parameters","observationPeriod":"2022-03-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3822M from Table 5 - ImageNet rows"},{"id":"epoch:2a351c15436bb988420e:training-compute","entityId":"epoch:2a351c15436bb988420e","metric":"training-compute","value":2.9113344e+20,"unit":"FLOP","observationPeriod":"2022-03-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use eight NVIDIA A100 GPUs to train RQ-Transformer of 1.4B parameters, and four GPUs to train RQ-Transformers of other sizes. The training time is <9 days for LSUN-cat, LSUN-bedroom, <4.5 days for ImageNet, and CC-3M, and <1 day for LSUN-church and FFHQ. \"\n\nTaken literally, suggests 3.8B model used fewer GPUs than 1.4B. This seems likely to be a typo, so I assume details are meant to be given for largest model (3.8B).\n\n(8) * (3.12e14) * (4.5 * 24 * 3600) * (0.3) = 2.911e20\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)"},{"id":"epoch:6381a806bc6ad87251bd:parameters","entityId":"epoch:6381a806bc6ad87251bd","metric":"parameters","value":774000000.0,"unit":"parameters","observationPeriod":"2022-03-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6381a806bc6ad87251bd:training-compute","entityId":"epoch:6381a806bc6ad87251bd","metric":"training-compute","value":1.7901648e+22,"unit":"FLOP","observationPeriod":"2022-03-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"pretraining: \n6 FLOP/parameter/token * 774000000 parameters * 372000000000 tokens = 1.727568e+21 FLOP \n\nunknown number of epochs -> \"Likely\" confidence\n\nexpert iteration: \n312000000000000 FLOP/GPU/sec * 48000 GPU-hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.617408e+22 FLOP\n\n1.617408e+22 FLOP + 1.727568e+21 FLOP = 1.7901648e+22 FLOP"},{"id":"epoch:a301d4f0e43569ce4b3a:parameters","entityId":"epoch:a301d4f0e43569ce4b3a","metric":"parameters","value":3200000000.0,"unit":"parameters","observationPeriod":"2022-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Remarkably, on a multilingual benchmark with 7,482 translation directions, our 200-layer model with 3.2B parameters significantly outperforms the 48-layer state-of-the-art model with 12B parameters by 5 BLEU points, which indicates a promising scaling direction\"\n\nEDIT 05/05/2022: The 12B model was presented in an earlier paper. This paper presents a 3.2B model"},{"id":"epoch:c426b79b481bb7234308:parameters","entityId":"epoch:c426b79b481bb7234308","metric":"parameters","value":2700000000.0,"unit":"parameters","observationPeriod":"2022-02-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.7B for largest model"},{"id":"epoch:c426b79b481bb7234308:training-compute","entityId":"epoch:c426b79b481bb7234308","metric":"training-compute","value":1.1e+21,"unit":"FLOP","observationPeriod":"2022-02-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use GPT-NeoX toolkit 11 to\ntrain the model efficiently in parallel with 8 Nvidia RTX 8000 GPUs on a single machine. The wall\ntime used to train the largest 2.7B model is about 6 weeks\"\n\n8 * 130 TFLOP/s * 6 * 7 * 24 * 3600 * 0.3 (utilization) ~= 1.1e21"},{"id":"epoch:52dca517bccb5609e3b0:training-compute","entityId":"epoch:52dca517bccb5609e3b0","metric":"training-compute","value":3.4504704e+20,"unit":"FLOP","observationPeriod":"2022-02-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000 FLOP / GPU / sec [A100] * 64 GPUs * 16 hours * 3600 sec / hour * 0.3 [assumed utilization] = 3.4504704e+20 FLOP"},{"id":"epoch:2b182e56fc8efcd84711:parameters","entityId":"epoch:2b182e56fc8efcd84711","metric":"parameters","value":269000000000.0,"unit":"parameters","observationPeriod":"2022-02-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"269B. it's called ST-MoE-32B because it's equivalent to a 32B dense model."},{"id":"epoch:2b182e56fc8efcd84711:training-compute","entityId":"epoch:2b182e56fc8efcd84711","metric":"training-compute","value":2.9e+23,"unit":"FLOP","observationPeriod":"2022-02-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The paper claims \"scaling a sparse model to 269B parameters, with a computational cost comparable to a 32B dense encoder-decoder\". If this is true for training cost, then 6*32e9*1.5e12 = 2.9e23"},{"id":"epoch:fa47ad087eed8383ecae:parameters","entityId":"epoch:fa47ad087eed8383ecae","metric":"parameters","value":137000000000.0,"unit":"parameters","observationPeriod":"2022-02-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"LaMDA is a family of Transformer-based neural language models specialized for dialog, which have up to 137B parameters\""},{"id":"epoch:fa47ad087eed8383ecae:training-compute","entityId":"epoch:fa47ad087eed8383ecae","metric":"training-compute","value":3.55e+23,"unit":"FLOP","observationPeriod":"2022-02-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The total FLOPS is 56.5% * 123 TFLOPS/s * 1024 chips * 57.7 days\n= 3.55E+23\"\nFrom https://arxiv.org/pdf/2201.08239.pdf p.18\n"},{"id":"epoch:4024dee3f1dc8f18692b:parameters","entityId":"epoch:4024dee3f1dc8f18692b","metric":"parameters","value":16000000.0,"unit":"parameters","observationPeriod":"2022-02-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Altogether, it includes \u223c16M trainable parameters, making it substantially smaller than other protein language models\""},{"id":"epoch:4024dee3f1dc8f18692b:training-compute","entityId":"epoch:4024dee3f1dc8f18692b","metric":"training-compute","value":6.5e+19,"unit":"FLOP","observationPeriod":"2022-02-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Pretraining speed on a single GPU (Nvidia Quadro RTX 5000) was 280 protein records per second. We trained the model for 28\u2009days over \u223c670M records\"\n\n28 * 24 * 3600 * 89 TFLOP/s * 0.3 (assumed utilization) = 6.5e19\nhttps://www.wolframalpha.com/input?i=28+days+*+89+TFLOP%2Fs+*+0.3"},{"id":"epoch:3128319c95db6bbf0ba9:parameters","entityId":"epoch:3128319c95db6bbf0ba9","metric":"parameters","value":20000000000.0,"unit":"parameters","observationPeriod":"2022-02-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:3128319c95db6bbf0ba9:training-compute","entityId":"epoch:3128319c95db6bbf0ba9","metric":"training-compute","value":9.31627008e+22,"unit":"FLOP","observationPeriod":"2022-02-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Trained for 3 months on 96 A100s (according to correspondence with author). Let's say 0.4 utilization rate."},{"id":"epoch:2b3c52e5c476d5ca06a0:parameters","entityId":"epoch:2b3c52e5c476d5ca06a0","metric":"parameters","value":227000000.0,"unit":"parameters","observationPeriod":"2022-02-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"227M\nTable 1"},{"id":"epoch:b5e0e05c8766b3dec100:parameters","entityId":"epoch:b5e0e05c8766b3dec100","metric":"parameters","value":7500000000.0,"unit":"parameters","observationPeriod":"2022-02-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Retro provides a constant gain for models ranging from 150M to 7B parameters, and Retro can be improved at evaluation time by increasing the database size and the number of retrieved neighbours. \""},{"id":"epoch:b5e0e05c8766b3dec100:training-compute","entityId":"epoch:b5e0e05c8766b3dec100","metric":"training-compute","value":1.68e+22,"unit":"FLOP","observationPeriod":"2022-02-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"C=6ND = 6 * 7e9 * 400e9 = 1.7e22 "},{"id":"epoch:8931581341801d11fc40:parameters","entityId":"epoch:8931581341801d11fc40","metric":"parameters","value":41100000000.0,"unit":"parameters","observationPeriod":"2022-02-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"41.1B. Table 3"},{"id":"epoch:8931581341801d11fc40:training-compute","entityId":"epoch:8931581341801d11fc40","metric":"training-compute","value":2.38010000000001e+23,"unit":"FLOP","observationPeriod":"2022-02-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Using C=6ND, we have C = 6 FLOP/token/param * 41.1B params * 967B tokens = 2.38e23 FLOP.\n\nFigure 7 (a) shows a maximum training compute budget of approx 23000 TPU-days per model. This matches the operation-counting estimate at 44% utilization."},{"id":"epoch:697d23c9b75c0de1e801:training-compute","entityId":"epoch:697d23c9b75c0de1e801","metric":"training-compute","value":9.7e+18,"unit":"FLOP","observationPeriod":"2022-01-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Rounding up to the nearest day, if we were to re-perform DARK from nothing to having a trained DARK3 it would take 12 days when parallelized across ten V100 GPUS. Of that time, model training constitutes just over 3 days and only requires 1 GPU\"\n\n3 * 24 * 3600 * 125 teraFLOP/s * 0.3 (utilization) = 9.7e18"},{"id":"epoch:97015afc08b3a5e2c7a2:parameters","entityId":"epoch:97015afc08b3a5e2c7a2","metric":"parameters","value":175000000000.0,"unit":"parameters","observationPeriod":"2022-01-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train three model sizes (1.3B, 6B, and 175B parameters)\""},{"id":"epoch:97015afc08b3a5e2c7a2:training-compute","entityId":"epoch:97015afc08b3a5e2c7a2","metric":"training-compute","value":3.19181e+23,"unit":"FLOP","observationPeriod":"2022-01-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"training our 175B PPO-ptx model requires 60 petaflops/s-days, compared to 3,640 petaflops/s-days for GPT-3 (Brown et al., 2020)\"\n\n60/3640 = +1.65% to base model compute\n\nbase model was reported 3.14e+23 FLOP\n\n3.14e+23 * 1.0165 = 319181000000000000000000"},{"id":"epoch:ac5f744ef181c8cf0f07:parameters","entityId":"epoch:ac5f744ef181c8cf0f07","metric":"parameters","value":6000000000.0,"unit":"parameters","observationPeriod":"2022-01-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train three model sizes (1.3B, 6B, and 175B parameters)\""},{"id":"epoch:8b0fb16cf2c660ccd28f:parameters","entityId":"epoch:8b0fb16cf2c660ccd28f","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2022-01-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train three model sizes (1.3B, 6B, and 175B parameters)\""},{"id":"epoch:dd5b38c94b4b17959252:parameters","entityId":"epoch:dd5b38c94b4b17959252","metric":"parameters","value":350000000.0,"unit":"parameters","observationPeriod":"2022-01-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:be51ae923c756e0ab8c8:parameters","entityId":"epoch:be51ae923c756e0ab8c8","metric":"parameters","value":1900000000.0,"unit":"parameters","observationPeriod":"2022-01-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For instance, in a 1.9B parameter configuration similar to GPT-3 XL, Primer uses 1/3 of the training compute to achieve the same one-shot performance as Transformer\""},{"id":"epoch:be51ae923c756e0ab8c8:training-compute","entityId":"epoch:be51ae923c756e0ab8c8","metric":"training-compute","value":2.2049963e+22,"unit":"FLOP","observationPeriod":"2022-01-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Each model is trained using batches of \u223c2M tokens using 512 TPUv4 chips for \u223c140 hours (\u223c71.8K total accelerator hours or \u223c1M train steps).\"\n\n71800 h * 0.3 [assumed utilization] * 275e12 FLOP/s * 3600s/h = 2.13246e+22 FLOP\n\n6 FLOP / parameter / token * 1.9 * 10^9 parameters * 2 * 10^12 tokens = 2.28e+22 FLOP\n\nsqrt(2.13246e+22*2.28e+22) = 2.2049963e+22 FLOP"},{"id":"epoch:dd163314859b4b95f098:parameters","entityId":"epoch:dd163314859b4b95f098","metric":"parameters","value":124000000.0,"unit":"parameters","observationPeriod":"2022-01-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:de77cfb8c78d4c6de26a:parameters","entityId":"epoch:de77cfb8c78d4c6de26a","metric":"parameters","value":124000000.0,"unit":"parameters","observationPeriod":"2022-01-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"124M (Table 2)\n\"We set the hidden size to 512 and the number of layers to 16 for all models.\""},{"id":"epoch:2888d08a807d8f7978e4:parameters","entityId":"epoch:2888d08a807d8f7978e4","metric":"parameters","value":420000000.0,"unit":"parameters","observationPeriod":"2022-01-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For the protein encoder, we use the pre-trained ProtBert from Elnaggar et al. (2020).\""},{"id":"epoch:7d236d34ea4114066666:parameters","entityId":"epoch:7d236d34ea4114066666","metric":"parameters","value":355000000.0,"unit":"parameters","observationPeriod":"2022-01-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The hyperparameters were selected to be similar to those used\nin the RoBERTa paper (Liu et al., 2019).\"\n\nLiu et al., 2019 link: https://arxiv.org/pdf/1907.11692.pdf\n\"We begin by training RoBERTa following the BERTLARGE architecture (L = 24, H = 1024, A = 16, 355M parameters)\""},{"id":"epoch:a7f06fe4c67be675df86:parameters","entityId":"epoch:a7f06fe4c67be675df86","metric":"parameters","value":705134592.0,"unit":"parameters","observationPeriod":"2022-01-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 4: \"We experiment with two model sizes: data2vec Base and\ndata2vec Large, containing either L = 12 or L = 24 Trans-\nformer blocks with H = 768 or H = 1024 hidden dimen-\nsion (with 4 \u00d7 H feed-forward inner-dimension)\""},{"id":"epoch:c9f7656505e1b5e22c54:parameters","entityId":"epoch:c9f7656505e1b5e22c54","metric":"parameters","value":705134592.0,"unit":"parameters","observationPeriod":"2022-01-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 4: \"We experiment with two model sizes: data2vec Base and\ndata2vec Large, containing either L = 12 or L = 24 Trans-\nformer blocks with H = 768 or H = 1024 hidden dimen-\nsion (with 4 \u00d7 H feed-forward inner-dimension)\"\n"},{"id":"epoch:4772c5a9ae66bbaf943c:parameters","entityId":"epoch:4772c5a9ae66bbaf943c","metric":"parameters","value":705134592.0,"unit":"parameters","observationPeriod":"2022-01-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 4: \"We experiment with two model sizes: data2vec Base and\ndata2vec Large, containing either L = 12 or L = 24 Trans-\nformer blocks with H = 768 or H = 1024 hidden dimen-\nsion (with 4 \u00d7 H feed-forward inner-dimension)\"\n"},{"id":"epoch:0e527a48bb964b132bc6:parameters","entityId":"epoch:0e527a48bb964b132bc6","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2022-01-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6628a22fed5542419710:parameters","entityId":"epoch:6628a22fed5542419710","metric":"parameters","value":88000000.0,"unit":"parameters","observationPeriod":"2022-01-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from https://github.com/microsoft/Swin-Transformer Swin-B have 88M, \nfrom page 8 :  'Training our ResNet50 model takes \u223c 22 hours on 8 V100 GPUs. The large 21K Swin-B model trains in \u223c 24 hours on 32 GPUs.'"},{"id":"epoch:6628a22fed5542419710:training-compute","entityId":"epoch:6628a22fed5542419710","metric":"training-compute","value":2.34399744e+19,"unit":"FLOP","observationPeriod":"2022-01-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"28.26e12* 32 * 24*3600*0.3 =2.34e19 = peak flops * num gpus * num seconds * assumed utilization rate\nfor Swin-B model from page 8 :  'Training our ResNet50 model takes \u223c 22 hours on 8 V100 GPUs. The large 21K Swin-B model trains in \u223c 24 hours on 32 GPUs.'"},{"id":"epoch:a721b5ba289e95b23ef0:parameters","entityId":"epoch:a721b5ba289e95b23ef0","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2021-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"To explore the landscape of large-scale pre-training for bidirectional text-image generation, we pre-train a 10-billion parameter model on a large-scale dataset of 145 million high-quality Chinese image-text pairs.\""},{"id":"epoch:caedede42c67f75c5ba4:parameters","entityId":"epoch:caedede42c67f75c5ba4","metric":"parameters","value":231000.0,"unit":"parameters","observationPeriod":"2021-12-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"(3) standard convolutional neural network (CNN; with two convolutional layers with a window size of 7, connected through ReLU activations; dropout rate of 0.25; 231 k free parameters)\""},{"id":"epoch:cfb87857dcbf935f11c0:parameters","entityId":"epoch:cfb87857dcbf935f11c0","metric":"parameters","value":260000000000.0,"unit":"parameters","observationPeriod":"2021-12-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"[We] developed... distributed training technology, including fine-grained parallelism, heterogeneous hardware-aware training, and fault tolerance mechanism to train the 260B model on both Nvidia V100 GPU and Ascend 910 NPU clusters.\"\nSee also:\nhttps://twitter.com/BaiduResearch/status/1468633977242243078?t=6q4zuLNdTSc4GUBe9OM5Aw&s=19"},{"id":"epoch:cfb87857dcbf935f11c0:training-compute","entityId":"epoch:cfb87857dcbf935f11c0","metric":"training-compute","value":1.0421e+24,"unit":"FLOP","observationPeriod":"2021-12-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The paper suggests that ERNIE 3.0 Titan uses more compute than GPT-3. This is consistent with the 6ND approximation.\n\nC = 6ND = 6 (FLOP/param/token) * (260B params) * (668B tokens) = 1.0421*10^24 FLOP"},{"id":"epoch:56bd6acc890973b24fbc:parameters","entityId":"epoch:56bd6acc890973b24fbc","metric":"parameters","value":3500000000.0,"unit":"parameters","observationPeriod":"2021-12-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Samples from a 3.5 billion parameter text-conditional diffusion model using classifier-free guidance are favored by human evaluators to those from DALL-E, even when the latter uses expensive CLIP reranking\""},{"id":"epoch:56bd6acc890973b24fbc:training-compute","entityId":"epoch:56bd6acc890973b24fbc","metric":"training-compute","value":4.7e+22,"unit":"FLOP","observationPeriod":"2021-12-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Note that GLIDE was\ntrained with roughly the same training compute as DALL-E\nbut with a much smaller model (3.5 billion vs. 12 billion\nparameters)\""},{"id":"epoch:8e59bebd3ceda47fe92c:parameters","entityId":"epoch:8e59bebd3ceda47fe92c","metric":"parameters","value":1100000000000.0,"unit":"parameters","observationPeriod":"2021-12-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:8e59bebd3ceda47fe92c:training-compute","entityId":"epoch:8e59bebd3ceda47fe92c","metric":"training-compute","value":2.227e+22,"unit":"FLOP","observationPeriod":"2021-12-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Reported directly in paper. Authors calculate FLOPs analytically in appendix G"},{"id":"epoch:0115f8e7ba3b95c17311:parameters","entityId":"epoch:0115f8e7ba3b95c17311","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2021-12-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0115f8e7ba3b95c17311:training-compute","entityId":"epoch:0115f8e7ba3b95c17311","metric":"training-compute","value":3.267e+22,"unit":"FLOP","observationPeriod":"2021-12-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:2671f346833a68f5ada4:parameters","entityId":"epoch:2671f346833a68f5ada4","metric":"parameters","value":1450000000.0,"unit":"parameters","observationPeriod":"2021-12-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.45B"},{"id":"epoch:d793c0dd2f8c261f0ec6:parameters","entityId":"epoch:d793c0dd2f8c261f0ec6","metric":"parameters","value":7500000000.0,"unit":"parameters","observationPeriod":"2021-12-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our largest model with 7.5 billion parameters sets new state of the art\""},{"id":"epoch:d793c0dd2f8c261f0ec6:training-compute","entityId":"epoch:d793c0dd2f8c261f0ec6","metric":"training-compute","value":2.25e+22,"unit":"FLOP","observationPeriod":"2021-12-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The XGLM 7.5B model was trained on 256 A100 GPUs for about 3 weeks, at a speed of 311.6k words per second\"\n\n256 * 312 teraFLOP/s * 21 * 24 * 3600 * 0.3 utilization assumption ~= 4.3e22\n\nalso, it was trained for 500B tokens. Using Compute = 6ND, we have\n6 * 500B * 7.5B = 2.25e22\n\n311k tokens per second * 7.5B params * 6 is 1.35e16 FLOP/s. divide that by 312 teraFLOP/s, which is A100 peak compute, gets 43, suggesting low utilization (17%) of the 256-GPU cluster, or somewhat higher if there's more than one token per word. So I'll use the 6ND number."},{"id":"epoch:beb182e5bc7ed7c68761:parameters","entityId":"epoch:beb182e5bc7ed7c68761","metric":"parameters","value":564000000.0,"unit":"parameters","observationPeriod":"2021-12-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"XGLM-564M is a multilingual autoregressive language model (with 564 million parameters) trained on a balanced corpus of a diverse set of 30 languages totaling 500 billion sub-tokens."},{"id":"epoch:320c089f77f836fb4445:parameters","entityId":"epoch:320c089f77f836fb4445","metric":"parameters","value":345000000.0,"unit":"parameters","observationPeriod":"2021-12-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:320c089f77f836fb4445:training-compute","entityId":"epoch:320c089f77f836fb4445","metric":"training-compute","value":2.272000000071e+21,"unit":"FLOP","observationPeriod":"2021-12-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"base model: Gpt-2 355M SOURCE: Uses pretrained GPT2 with 345M param, so impute from \"Language Models are Unsupervised Multitask Learners\"\n\nsame \"Speculative\" confidence level"},{"id":"epoch:6329641b16d5efaec129:parameters","entityId":"epoch:6329641b16d5efaec129","metric":"parameters","value":110000000.0,"unit":"parameters","observationPeriod":"2021-12-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Based on BERT base, which had 110m params.\n\n\"We initialize the network with the publicly available BERT base uncased model.\""},{"id":"epoch:6329641b16d5efaec129:training-compute","entityId":"epoch:6329641b16d5efaec129","metric":"training-compute","value":1.57e+20,"unit":"FLOP","observationPeriod":"2021-12-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Pre-training:\n\"We use the random cropping data augmentation, with documents of 256 tokens... batch size of 2,048 and 500,000 steps\"\n256 * 2048 * 500k * 100M * 6 = 1.57e20\n\nFine-tuning looks unlikely to move final sum much beyond this."},{"id":"epoch:140234ae5841ac7979c9:parameters","entityId":"epoch:140234ae5841ac7979c9","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2021-12-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3B from section 4.1"},{"id":"epoch:2a92be43df61340268c1:parameters","entityId":"epoch:2a92be43df61340268c1","metric":"parameters","value":300000000000.0,"unit":"parameters","observationPeriod":"2021-12-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2a92be43df61340268c1:training-compute","entityId":"epoch:2a92be43df61340268c1","metric":"training-compute","value":1.6956e+24,"unit":"FLOP","observationPeriod":"2021-12-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"No indication of how images are processed. Supposing they used something like ViT-H/14, and training images were 512x512 (they state \"EXAONE shows remarkable performance such as [...] offering 1024x1024 sized image output\", but typically this size of image training would only be done during a relatively short, final stage of pre-training), there would be 37x37 = 1,369 patches per image\n1,369 * 250 million = around 342 billion image patch embeddings.\n\n300M parameters * (342 billion + 600 billion) image tokens * 6 = 1.6956e24"},{"id":"epoch:bd223dfb34de9762ac6e:parameters","entityId":"epoch:bd223dfb34de9762ac6e","metric":"parameters","value":1200000000000.0,"unit":"parameters","observationPeriod":"2021-12-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.2 trillion parameters"},{"id":"epoch:bd223dfb34de9762ac6e:training-compute","entityId":"epoch:bd223dfb34de9762ac6e","metric":"training-compute","value":3.6363112434e+23,"unit":"FLOP","observationPeriod":"2021-12-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The network activates 96.6 billion parameters per token and trained for 600B tokens.\n\n6 * 600B * 96.6B = 3.478e23\n\nDigitizing figure 4 (d) indicates 139.67 TPU-years of training. \n2.75e14 * 139.67 * 365.25 * 24 * 3600 * 0.3 = 3.636e23\n\nSince these are close, we will use the 6NC estimate and derive hardware utilization from the training time information.\n\nLater they say they measured 326W power usage per chip, which could maybe be used to estimate utilization."},{"id":"epoch:8e450db30c34904a5793:parameters","entityId":"epoch:8e450db30c34904a5793","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2021-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"ENGINE-XL has 48 layers and 1.5 billion parameters, on par with GPT2-1.5B and GROVER-Mega\""},{"id":"epoch:6c9d23f0486e472dd520:parameters","entityId":"epoch:6c9d23f0486e472dd520","metric":"parameters","value":124000000.0,"unit":"parameters","observationPeriod":"2021-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"ENGINE-Base has 12 layers and 124 million parameters, on par with GPT2-124M and ROVER-Base\""},{"id":"epoch:1b0eea1678c996b8021e:parameters","entityId":"epoch:1b0eea1678c996b8021e","metric":"parameters","value":355000000.0,"unit":"parameters","observationPeriod":"2021-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"ENGINE-Medium has 24 layers and 355 million parameters, on par with GPT2-355M and GROVER-Large\""},{"id":"epoch:515c0929331389cdec2e:parameters","entityId":"epoch:515c0929331389cdec2e","metric":"parameters","value":280000000000.0,"unit":"parameters","observationPeriod":"2021-12-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Language modelling provides a step towards intelligent communication systems by harnessing large repositories of written human knowledge to better predict and understand the world. In this paper, we present an analysis of Transformer-based language model performance across a wide range of model scales -- from models with tens of millions of parameters up to a 280 billion parameter model called Gopher. These models are evaluated on 152 diverse tasks, achieving state-of-the-art performance across the majority. Gains from scale are largest in areas such as reading comprehension, fact-checking, and the identification of toxic language, but logical and mathematical reasoning see less benefit. We provide a holistic analysis of the training dataset and model's behaviour, covering the intersection of model scale with bias and toxicity. Finally we discuss the application of language models to AI safety and the mitigation of downstream harms."},{"id":"epoch:515c0929331389cdec2e:training-compute","entityId":"epoch:515c0929331389cdec2e","metric":"training-compute","value":6.31e+23,"unit":"FLOP","observationPeriod":"2021-12-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table A26\n6.31E+08 Train PFLOPs"},{"id":"epoch:85d1de1dc89b931508bf:parameters","entityId":"epoch:85d1de1dc89b931508bf","metric":"parameters","value":7100000000.0,"unit":"parameters","observationPeriod":"2021-12-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.1B"},{"id":"epoch:61c407934c94c1183789:training-compute","entityId":"epoch:61c407934c94c1183789","metric":"training-compute","value":3.6679273004682866e+22,"unit":"FLOP","observationPeriod":"2021-12-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We trained a version of AlphaZero using its original settings in chess and Go, e.g. , using 800 MCTS simulations during training, with 3500 concurrent actors each on a single TPUv4, for a total of 800k training steps. SOG was trained using a similar amount of TPU resources.\""},{"id":"epoch:60885aa7a850a2676c57:parameters","entityId":"epoch:60885aa7a850a2676c57","metric":"parameters","value":70000000.0,"unit":"parameters","observationPeriod":"2021-12-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"'CTR-BERTuses about 70 million parameters which can be trained on 8 A100 GPUs in less than a day (<1000 USD).\n'"},{"id":"epoch:60885aa7a850a2676c57:training-compute","entityId":"epoch:60885aa7a850a2676c57","metric":"training-compute","value":6.469632e+19,"unit":"FLOP","observationPeriod":"2021-12-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"flops = (8) * (312 * 10**12) * (24 * 3600) * (0.3)\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)\n\n'CTR-BERT\nuses about 70 million parameters which can be trained on 8 A100 GPUs in less than a day (<1000\nUSD).\n'"},{"id":"epoch:a7f43517b886b3de85c1:parameters","entityId":"epoch:a7f43517b886b3de85c1","metric":"parameters","value":5400000000.0,"unit":"parameters","observationPeriod":"2021-12-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1 of the blogpost"},{"id":"epoch:87f3ff5cb16d17592f58:parameters","entityId":"epoch:87f3ff5cb16d17592f58","metric":"parameters","value":202999999.99999997,"unit":"parameters","observationPeriod":"2021-11-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"203M (Table 6)"},{"id":"epoch:87f3ff5cb16d17592f58:training-compute","entityId":"epoch:87f3ff5cb16d17592f58","metric":"training-compute","value":1.25454e+19,"unit":"FLOP","observationPeriod":"2021-11-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 203000000 parameters * 103000000 tokens * 100 epochs = 1.25454 \u00d7 10^19 FLOP"},{"id":"epoch:c9f3aeaaa1a89eb198ba:parameters","entityId":"epoch:c9f3aeaaa1a89eb198ba","metric":"parameters","value":68000000.0,"unit":"parameters","observationPeriod":"2021-11-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 6"},{"id":"epoch:c9f3aeaaa1a89eb198ba:training-compute","entityId":"epoch:c9f3aeaaa1a89eb198ba","metric":"training-compute","value":4.2024e+18,"unit":"FLOP","observationPeriod":"2021-11-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Using 6ND: \n\n6*68000000*103000000 tokens*100 epochs=4.2024e+18"},{"id":"epoch:85c1d937e040842a5313:parameters","entityId":"epoch:85c1d937e040842a5313","metric":"parameters","value":870000000.0,"unit":"parameters","observationPeriod":"2021-11-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 4.1"},{"id":"epoch:85c1d937e040842a5313:training-compute","entityId":"epoch:85c1d937e040842a5313","metric":"training-compute","value":7.24598784e+21,"unit":"FLOP","observationPeriod":"2021-11-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From AI Tracker:\n\"Compute cost: End of Sec 4.1: \"We pre-train on 64 A100 GPUs for two weeks\". \n\nHalf precision FLOPs of A100: 312000000000000\n\n64 gpus *312000000000000 FLOPs *0.3 utilization * 14 day* (24*60*60) seconds / day=7.245988e+21\n\n"},{"id":"epoch:d63b8b887bb4d8b349a4:parameters","entityId":"epoch:d63b8b887bb4d8b349a4","metric":"parameters","value":100000000000000.0,"unit":"parameters","observationPeriod":"2021-11-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"100 trillion"},{"id":"epoch:11771256e13f96a2501e:parameters","entityId":"epoch:11771256e13f96a2501e","metric":"parameters","value":893000000.0,"unit":"parameters","observationPeriod":"2021-11-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our Florence pretrained model has in total 893M parameters, including the language transformer with 256M parameters and the CoSwin-H transformer with 637M parameters.\""},{"id":"epoch:11771256e13f96a2501e:training-compute","entityId":"epoch:11771256e13f96a2501e","metric":"training-compute","value":4.831e+22,"unit":"FLOP","observationPeriod":"2021-11-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The model takes 10 days to train on 512 NVIDIA A100 GPUs with 40GB memory per GPU.\"\n512 * 312 teraFLOPS * 10 days * 35% utilization = 4.831e22 FLOP"},{"id":"epoch:ea50311c5d0059338afd:parameters","entityId":"epoch:ea50311c5d0059338afd","metric":"parameters","value":3070000000.0,"unit":"parameters","observationPeriod":"2021-11-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.4B image model + 670M text model"},{"id":"epoch:ea50311c5d0059338afd:training-compute","entityId":"epoch:ea50311c5d0059338afd","metric":"training-compute","value":4.12e+22,"unit":"FLOP","observationPeriod":"2021-11-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6.9k + 1k + 0.8k = 8.7k TPUv4 core-days for BASIC-L, per Table 8\n\nTwo cores per chip, and 275 teraflop/s per chip \n(https://cloud.google.com/tpu/docs/system-architecture-tpu-vm#tpu_v4)\n\n275 teraflops * 8700/2 * 24 * 3600 * 0.4 (assumed utilization) = 8.3e22"},{"id":"epoch:4566f07e9630d55539a8:parameters","entityId":"epoch:4566f07e9630d55539a8","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2021-11-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:4566f07e9630d55539a8:training-compute","entityId":"epoch:4566f07e9630d55539a8","metric":"training-compute","value":1.1e+21,"unit":"FLOP","observationPeriod":"2021-11-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"trained on \"<0.5k\" TPUv3 core-days per Table 2 (not trained on TPUs, this is a comparison with other papers)\n\nA core is 123/2 teraflops\n\n500 core-days\n= 500 * 123/2 trillion * 24 * 3600 * 0.4 utilization\n~= 1.1e21"},{"id":"epoch:60fe5f005aa474313243:parameters","entityId":"epoch:60fe5f005aa474313243","metric":"parameters","value":418000000.0,"unit":"parameters","observationPeriod":"2021-11-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from HF \"The DeBERTa V3 large model comes with 24 layers and a hidden size of 1024 . Its total parameter number is 418M since we use a vocabulary containing 128K tokens which introduce 131M parameters in the Embedding layer.  This model was trained using the 160GB data as DeBERTa V2.\""},{"id":"epoch:60fe5f005aa474313243:training-compute","entityId":"epoch:60fe5f005aa474313243","metric":"training-compute","value":1.07008e+20,"unit":"FLOP","observationPeriod":"2021-11-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 FLOP / parameter / token * 42666666667 tokens * 418000000 parameters = 1.07008e+20 FLOP"},{"id":"epoch:fb754077fa66617af9b4:parameters","entityId":"epoch:fb754077fa66617af9b4","metric":"parameters","value":650000000.0,"unit":"parameters","observationPeriod":"2021-11-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train ESM-1v, a 650M parameter transformer language model for prediction of variant effects\""},{"id":"epoch:fb754077fa66617af9b4:training-compute","entityId":"epoch:fb754077fa66617af9b4","metric":"training-compute","value":1.3500000000000007e+20,"unit":"FLOP","observationPeriod":"2021-11-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"ESM-1v models are pre-trained for 6 days on 64 V100 GPUs\" [F - Compute costs]\nAssuming  V100 PCIe single precision 14 TFLOPS and 0.3 utilization rate \n\nEstimate: (6*24*3600) s * 14e12 FLOP/s * 0.3 *64 = 1.4e20 FLOPs\n\nAlternative estimate based on Figure 7: \n10^(7.5) GPU-seconds * 14e12 FLOP/s * 0.3 = 1.3e20 FLOPs\n\nMean: 1.35e20 FLOP"},{"id":"epoch:3e2d274295b81b142837:parameters","entityId":"epoch:3e2d274295b81b142837","metric":"parameters","value":3005000000.0,"unit":"parameters","observationPeriod":"2021-11-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 7"},{"id":"epoch:817f16f4590f3813e1e8:training-compute","entityId":"epoch:817f16f4590f3813e1e8","metric":"training-compute","value":1.08e+19,"unit":"FLOP","observationPeriod":"2021-11-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training details here:\n\nhttps://docs.nvidia.com/bionemo-framework/latest/models/equidock.html\n\n32 A100s can do 30 epochs per hour on the DIPS dataset. Equidock was trained on 30 epochs on DIPS and 150 epochs on DB5.5. DIPS is about 100x bigger, so the large majority of compute was DIPS.\n\n32 A100-hours = 312 teraflops * 32 * 3600 * 0.3 \n~= 1.08e19"},{"id":"epoch:2a8aa1e72d21e2af5f83:parameters","entityId":"epoch:2a8aa1e72d21e2af5f83","metric":"parameters","value":18000000000.0,"unit":"parameters","observationPeriod":"2021-11-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2a8aa1e72d21e2af5f83:training-compute","entityId":"epoch:2a8aa1e72d21e2af5f83","metric":"training-compute","value":1.35e+18,"unit":"FLOP","observationPeriod":"2021-11-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"=18000000000 parameters *12500000 training tokens * 6 FLOP/token/parameter=1350000000000000000=1.35 \u00d7 10^18 FLOP\n\nunsure about training dataset size->speculative confidence level"},{"id":"epoch:7c23b0d05c4eabe594e0:parameters","entityId":"epoch:7c23b0d05c4eabe594e0","metric":"parameters","value":6166502400.0,"unit":"parameters","observationPeriod":"2021-11-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from https://huggingface.co/kakaobrain/kogpt\n\npossibly has 30B parameters based on this source: (would be a different version)\n\nhttps://post.naver.com/viewer/postView.naver?volumeNo=34605062&memberNo=35753905&vType=VERTICAL"},{"id":"epoch:19cc7b29825237669f4c:parameters","entityId":"epoch:19cc7b29825237669f4c","metric":"parameters","value":632000000.0,"unit":"parameters","observationPeriod":"2021-11-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Three models:\nViT-B (86M), ViT-L (304M), ViT-H (632M)"},{"id":"epoch:19cc7b29825237669f4c:training-compute","entityId":"epoch:19cc7b29825237669f4c","metric":"training-compute","value":4.6e+20,"unit":"FLOP","observationPeriod":"2021-11-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"128 TPU-v3 cores trained for 1600 epochs. Times are given for 800 epochs in Table 2; largest model (ViT-H) took 34.5 hrs for 800.\n128 TPU-v3 cores * 0.5 chips/core * 34.5 hours * 2 * 1.23E+14 FLOP/sec / chip * 3600 sec/hour  * 40% utilization = 7.84e20 FLOP\n\nNote that the operations counting method disagrees:\n2 \u00d7 632000000 connections \u00d7 3 \u00d7 1281167 training examples \u00d7 1600 of epochs  = 7.8e18 FLOP\n\nManual calculation with `calflops` package roughly agrees with hardware-time calculation: \n286.21 GFLOPS/observation * 1281167 observations * 1600 epochs = 5.86e20 FLOP\n\nSee reproduction here: https://colab.research.google.com/drive/1KCsmrfPzT9BgGO_YQthnz4oP3QRqbw5o?usp=sharing\n\nWeighting three estimates equally:\n(7.84e20 + 7.8e18 + 5.86e20)/3 = 4.6e20"},{"id":"epoch:2cfda97128a73ab49f98:parameters","entityId":"epoch:2cfda97128a73ab49f98","metric":"parameters","value":1600000000.0,"unit":"parameters","observationPeriod":"2021-11-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We examined the improvement in model size in detail by considering four model sizes: 0.35B, 0.7B, 1.1B, and 1.6B parameters\""},{"id":"epoch:2eb6b4d1f634e227ddec:parameters","entityId":"epoch:2eb6b4d1f634e227ddec","metric":"parameters","value":1500000.0,"unit":"parameters","observationPeriod":"2021-11-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:aee74d5245c133b106ed:parameters","entityId":"epoch:aee74d5245c133b106ed","metric":"parameters","value":132000000.0,"unit":"parameters","observationPeriod":"2021-11-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In all our experiments we use the GPT2-small configuration with 124M parameters, with 12 layers and 12 attention heads each for our base model.\"\n\n\"Our model has 132M parameters, a 6% increase, after the addition of the Entity-Gating layer and the entity representations.\""},{"id":"epoch:aee74d5245c133b106ed:training-compute","entityId":"epoch:aee74d5245c133b106ed","metric":"training-compute","value":1.1725747e+20,"unit":"FLOP","observationPeriod":"2021-11-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"GPT-2 124M compute ~ GPT-2 117M compute = 1.1699999999999999 \u00d7 10^20 FLOP\n\n1.17 * 10^20 FLOP + 2.57472e+17 FLOP = 1.1725747e+20 FLOP"},{"id":"epoch:d0b60d028c55ed1ae3c2:parameters","entityId":"epoch:d0b60d028c55ed1ae3c2","metric":"parameters","value":220000000.0,"unit":"parameters","observationPeriod":"2021-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We build CodeT5 based on Huggingface\u2019s T5 (Raffel et al., 2020) PyTorch implementation and employ two sizes of CodeT5-small (60M) and CodeT5-base (220M)\""},{"id":"epoch:d0b60d028c55ed1ae3c2:training-compute","entityId":"epoch:d0b60d028c55ed1ae3c2","metric":"training-compute","value":1.56e+21,"unit":"FLOP","observationPeriod":"2021-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We pre-train the model with the denoising objective for 100 epochs and bimodal dual training for further 50 epochs on a cluster of 16 NVIDIA A100 GPUs with 40G memory. The total training time for CodeT5-small and CodeT5- base is 5 and 12 days, respectively\"\n\n16 * 312 teraFLOP/s * 12 * 24 * 3600 * 0.3 (utilization assumption) = 1.56e21"},{"id":"epoch:89d4de2a1f92c5755865:training-compute","entityId":"epoch:89d4de2a1f92c5755865","metric":"training-compute","value":1.05e+19,"unit":"FLOP","observationPeriod":"2021-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"With this setting, each experiment takes roughly 100-200 GPU hours on a NVIDIA V100,\nfor more details we refer to the appendix.\"\n\n\"We conduct our experiments on an internal cluster with several nodes, each with up to 8 Quadro RTX\n6000 or NVIDIA V100 using PyTorch 1.7.1 and CUDA 11.0.\"\n\nIn appendix table 7, takes 10.1 seconds per 1k images on 8 Quadro RTX 6000s. Longest training run for Projected GAN appears to be in Figure 4 (left), at 14M images, though this is overtrained and the largest checkpoint used for evaluations was 10M.\n10M images * 10.1 s/1000 images * 8 * 3.26e13 FLOP/s * 0.4 = 1.05e19"},{"id":"epoch:64b54c8117efcba0bbc9:parameters","entityId":"epoch:64b54c8117efcba0bbc9","metric":"parameters","value":249000000.00000003,"unit":"parameters","observationPeriod":"2021-10-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"249M (Table 8)"},{"id":"epoch:64b54c8117efcba0bbc9:training-compute","entityId":"epoch:64b54c8117efcba0bbc9","metric":"training-compute","value":7.8328627e+19,"unit":"FLOP","observationPeriod":"2021-10-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 249000000 parameters * 8 GPUs * 8192 tokens/step/GPU * 800000 steps = 7.8328627e+19 FLOP\n \n\"our S4 model was trained with the simpler AdamW optimizer with a single cosine learning rate cycle with a maximum of 800000 steps.\nThe initial learning rate was set to 0.0005. We used 8 A100 GPUs with a batch size of 1 per gpu and context size 8192.\"\n\n"},{"id":"epoch:7da457eb136a649ef501:parameters","entityId":"epoch:7da457eb136a649ef501","metric":"parameters","value":15010300.0,"unit":"parameters","observationPeriod":"2021-10-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The Bayesian VAE architecture in EVE is comprised of a symmetric 3-layer encoder & decoder architecture (with 2,000-1,000-300 and 300-1,000-2,000 units respectively) and a latent space of dimension 50 [...] We use a single set of parameters for the encoder (\u03d5p) and learn a fully-factorized gaussian distribution over the weights of the decoder (\u03b8p)\"\nThey train a new VAE for each protein, and it doesn't seem like they trim the input sequence length, so the largest model will be the one trained for the largest input protein. Supplementary materials 1 gives statistics for each protein; the longest is 5202, which would indicate a network of size 15,010,300"},{"id":"epoch:8abed99d01857ad51632:training-compute","entityId":"epoch:8abed99d01857ad51632","metric":"training-compute","value":3.8e+19,"unit":"FLOP","observationPeriod":"2021-10-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"trained on a TPU v3-8 for 72 hours. That's 8 TPUv3 cores, or 4 TPUv3 chips (123 teraflop/s each)\n\nflops = (4) * (123 * 10**12) * (72 * 3600) * (0.3) = 3.8e19\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)\nfrom https://huggingface.co/dalle-mini/dalle-mini#dall%C2%B7e-mini-estimated-emissions"},{"id":"epoch:d5ea441072877bb1b78f:parameters","entityId":"epoch:d5ea441072877bb1b78f","metric":"parameters","value":250000000.0,"unit":"parameters","observationPeriod":"2021-10-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Following the RoBERTa-base setting, the hidden size, feed forward dimension,\nnumber of encoder layers, and attention heads of the base models are set as 768, 3072, 12, 12 re-\nspectively (denoted as MLM-base for MLM and PMLM-base for PMLM). A larger model named\nPMLM-large is pre-trained with the same setting except using 34 as the number of encoder layers.\""},{"id":"epoch:d5ea441072877bb1b78f:training-compute","entityId":"epoch:d5ea441072877bb1b78f","metric":"training-compute","value":3.81e+21,"unit":"FLOP","observationPeriod":"2021-10-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1176*60*60*125000000000000*24*0.3=3.81024e+21"},{"id":"epoch:c01a1f719e7bbaa0d627:parameters","entityId":"epoch:c01a1f719e7bbaa0d627","metric":"parameters","value":274000000.0,"unit":"parameters","observationPeriod":"2021-10-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"274M (table 1)\n\n\"We use the base version of deep Transformer language model with adaptive embeddings (Baevski & Auli, 2018) as our base LM. This model has 16 decoder layers. The dimensionality of word representations is 1,024, the number of multi-attention heads is 16, and the inner dimensionality of feedforward layers is 4,096.\""},{"id":"epoch:c01a1f719e7bbaa0d627:training-compute","entityId":"epoch:c01a1f719e7bbaa0d627","metric":"training-compute","value":5.2587456e+19,"unit":"FLOP","observationPeriod":"2021-10-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"base model compute: 4.47*10^19 FLOP \nfine-tune lower bound estimation: 1.69332e+17 FLOP -> 4.47*10^19 FLOP + 1.69332e+17 FLOP = 4.4869332e+19 FLOP total \nfine-tune upper bound estimation: 1.69332e+19 FLOP -> 4.47*10^19 FLOP + 1.69332e+19 FLOP = 6.16332e+19 FLOP total\n\ngeometric mean: sqrt(4.4869332e+19*6.16332e+19) = 5.2587456e+19 \n"},{"id":"epoch:c033d251743851866a4f:parameters","entityId":"epoch:c033d251743851866a4f","metric":"parameters","value":247000000.0,"unit":"parameters","observationPeriod":"2021-10-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d8a64d3084399cd08381:parameters","entityId":"epoch:d8a64d3084399cd08381","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2021-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d8a64d3084399cd08381:training-compute","entityId":"epoch:d8a64d3084399cd08381","metric":"training-compute","value":2.592e+20,"unit":"FLOP","observationPeriod":"2021-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"They report their compute directly.\n\nFrom section 8: \"About 62k GPU-hours on the Jean Zay HPC Cluster.\" Jean Zay uses both A100 and V100 GPUs, and maybe other stuff as well?\n\nNote they explicitly call out V100 in their Appendix A.\n\nhttps://www.hpcwire.com/2021/11/17/frances-jean-zay-supercomputer-gets-ai-boost-from-hpe-nvidia/"},{"id":"epoch:a5a0810561bf849fdd31:parameters","entityId":"epoch:a5a0810561bf849fdd31","metric":"parameters","value":1540000000.0,"unit":"parameters","observationPeriod":"2021-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"They don't actually say which GPT-2 version they use, but it appears to be XL (1.5B params), since Table 1 shows zero-shot accuracy of 18.34 on WikiText-2, which is what the 1.5B version got."},{"id":"epoch:a5a0810561bf849fdd31:training-compute","entityId":"epoch:a5a0810561bf849fdd31","metric":"training-compute","value":1.92e+21,"unit":"FLOP","observationPeriod":"2021-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"They fine-tuned GPT-2 (1.5B) for one epoch on WikiText-2. GPT-2 itself used 1.92e21 FLOP, and here we see 6 * 2M * 1.54B = 1.848e16 fine-tuning FLOP (which has almost no impact on total FLOP)."},{"id":"epoch:c73a3c075ce826961c52:parameters","entityId":"epoch:c73a3c075ce826961c52","metric":"parameters","value":90000000.0,"unit":"parameters","observationPeriod":"2021-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c73a3c075ce826961c52:training-compute","entityId":"epoch:c73a3c075ce826961c52","metric":"training-compute","value":8.8992e+18,"unit":"FLOP","observationPeriod":"2021-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(90M parameters)(103M words)(4/3 tokens/word)(120 epochs)*6"},{"id":"epoch:d0f35dc7a808154c46c4:parameters","entityId":"epoch:d0f35dc7a808154c46c4","metric":"parameters","value":40000000.0,"unit":"parameters","observationPeriod":"2021-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Small\" setup from https://arxiv.org/pdf/2102.11174 "},{"id":"epoch:d0f35dc7a808154c46c4:training-compute","entityId":"epoch:d0f35dc7a808154c46c4","metric":"training-compute","value":2.9664e+18,"unit":"FLOP","observationPeriod":"2021-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND formula:\n6*103000000*120*40000000=2.9664e+18"},{"id":"epoch:82f553cd1bf7b788c369:parameters","entityId":"epoch:82f553cd1bf7b788c369","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2021-10-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Unless specified otherwise, we use the XXL version which\nhas 11B parameters.\""},{"id":"epoch:82f553cd1bf7b788c369:training-compute","entityId":"epoch:82f553cd1bf7b788c369","metric":"training-compute","value":9.1819e+20,"unit":"FLOP","observationPeriod":"2021-10-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From Table 1 and section B.1, a single run uses 27 hours of a 512 core slice of a TPU-v3 pod. \n512 * 0.5 * 1.23e14 * 3600 * 27 * 0.3 = 9.18e20\n(cores) * (chip/core) * (FLOP/chip-sec) * (sec/hour) * (hours) * (utilization assumption)"},{"id":"epoch:3c57dfc352259cdd2009:parameters","entityId":"epoch:3c57dfc352259cdd2009","metric":"parameters","value":83000000.0,"unit":"parameters","observationPeriod":"2021-10-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The KnGPT2 model is compressed from the GPT-2Small Radford et al. (2019) model. GPT-2Small is 124 million parameters. Our baseline is DistilGPT2 which has about 82 million parameters so our KnGPT2 model is compressed to the same size (83 million parameters) for a fair comparison\""},{"id":"epoch:3c57dfc352259cdd2009:training-compute","entityId":"epoch:3c57dfc352259cdd2009","metric":"training-compute","value":7.9402496e+20,"unit":"FLOP","observationPeriod":"2021-10-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.936e+20 FLOP [base model, \"Speculative\" confidence\"] + 4.2496e+17 FLOP [fine-tune] = 7.9402496e+20 FLOP"},{"id":"epoch:6caacf9af13cbdb4e060:parameters","entityId":"epoch:6caacf9af13cbdb4e060","metric":"parameters","value":245730000000.0,"unit":"parameters","observationPeriod":"2021-10-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2: Parameters of Yuan models.\n\"Parameters (billion)\""},{"id":"epoch:6caacf9af13cbdb4e060:training-compute","entityId":"epoch:6caacf9af13cbdb4e060","metric":"training-compute","value":3.5380000000001e+23,"unit":"FLOP","observationPeriod":"2021-10-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 9: 4095 petaFLOPS-days which equals 3.538*10^23 FLOP\n\nhttps://www.wolframalpha.com/input?i=4095+petaFLOPS+*+1+day\n"},{"id":"epoch:75f189aeb9d468cef418:parameters","entityId":"epoch:75f189aeb9d468cef418","metric":"parameters","value":220000000.0,"unit":"parameters","observationPeriod":"2021-10-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"220M from Table 1 entry for TOME"},{"id":"epoch:75f189aeb9d468cef418:training-compute","entityId":"epoch:75f189aeb9d468cef418","metric":"training-compute","value":1.03809024e+21,"unit":"FLOP","observationPeriod":"2021-10-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"We have information about hardware \"All models are pre-trained on 128 TPUs using AdamW optimizer (Loshchilov& Hutter, 2019) with learning rate 1e-4 and batch size of 4096.\", but no information about training time and exact model of TPUs.\n\nThere is information about training 1.5M steps with batch size 4096 * 128 citation from appendix A:\n\"The Mention Encoder and BATCH-TOME are pre-trained\nfor 1 million steps with 50k warmup steps, and TOME is trained for 500k additional steps with 25k warmup steps after initialization from BATCH-TOME.\"\n\nIt seems like 6ND is appropriate, since all components are dense transformers. In this case, compute is:\n6 * 1.5M (steps) * 4096 (batch size) * 128 (seq len) * 220M (parameters) = 1.038e21"},{"id":"epoch:b2214e00967e177f9eae:parameters","entityId":"epoch:b2214e00967e177f9eae","metric":"parameters","value":530000000000.0,"unit":"parameters","observationPeriod":"2021-10-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b2214e00967e177f9eae:training-compute","entityId":"epoch:b2214e00967e177f9eae","metric":"training-compute","value":8.586e+23,"unit":"FLOP","observationPeriod":"2021-10-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.lesswrong.com/posts/bGuMrzhJdENCo8BxX/nvidia-and-microsoft-releases-530b-parameter-transformer?commentId=HSJSNspKp94tFcSCx\n\nsource: https://lair.lighton.ai/akronomicon/\n9938 PF-days * 3600 * 24 * 10^15  = 8.586432e+23\n\n6ND estimate: 6 * 530B * 270B = 8.586000e+23"},{"id":"epoch:f0c16afd16b27e65c967:parameters","entityId":"epoch:f0c16afd16b27e65c967","metric":"parameters","value":10000000000000.0,"unit":"parameters","observationPeriod":"2021-10-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We demonstrate a practice of pretraining unprecedented 10-trillion-parameter model, an order of magnitude larger than the state-of-the-art, on solely 512 GPUs within 10 days\""},{"id":"epoch:f0c16afd16b27e65c967:training-compute","entityId":"epoch:f0c16afd16b27e65c967","metric":"training-compute","value":5.53e+21,"unit":"FLOP","observationPeriod":"2021-10-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"512 GPUs in 10 days - using NVIDIA V100 GPUs\n\nUsing the NVIDIA V100 Specifications this works out to be: \n0.30 * 125E12 * 512 * 10 * 86400 = 1.66E22\n\n(Assuming 30% utilisation, and 125 TFLOPS)"},{"id":"epoch:addc8c68109e41e4938c:training-compute","entityId":"epoch:addc8c68109e41e4938c","metric":"training-compute","value":4.35e+21,"unit":"FLOP","observationPeriod":"2021-10-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section: 2.5. Training Regimen\n\"We train the model to convergence (approximately 10M samples, for 2 weeks) across 128 TPUv3 cores [...]. Then we [...] run two separate fine-tuning stages (one further day of training each)\"\n\nAssuming: FP16 and utilization 0.4\n\nCalculation: (14+2) days * 24 hours/day * 60 min/hour * 60 sec/min * (128 TPU cores/2 cores per chip) * 1.23e14 FLOP/s per chip * 0.4 utilization = 4.35e21 FLOPs"},{"id":"epoch:4f8d3833d653a6244ed1:parameters","entityId":"epoch:4f8d3833d653a6244ed1","metric":"parameters","value":2200000000.0,"unit":"parameters","observationPeriod":"2021-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.2B"},{"id":"epoch:4f8d3833d653a6244ed1:training-compute","entityId":"epoch:4f8d3833d653a6244ed1","metric":"training-compute","value":2.8983951e+22,"unit":"FLOP","observationPeriod":"2021-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"312000000000000 FLOP / GPU / sec [A100] * 256 GPUs * 336 hours * 3600 sec / hour * 0.3 [assumed utilization] = 2.8983951e+22 FLOP"},{"id":"epoch:ca296bb0707445511a76:parameters","entityId":"epoch:ca296bb0707445511a76","metric":"parameters","value":558000000.0,"unit":"parameters","observationPeriod":"2021-09-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"558M table 5"},{"id":"epoch:8db4ad012a486ed01d2a:parameters","entityId":"epoch:8db4ad012a486ed01d2a","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2021-09-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:8db4ad012a486ed01d2a:training-compute","entityId":"epoch:8db4ad012a486ed01d2a","metric":"training-compute","value":9.9e+21,"unit":"FLOP","observationPeriod":"2021-09-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In PLATO-XL, each model was trained for a total of 150B tokens, with\na batch size of 2M tokens.\"\n\n150B * 11B * 6 = 9.9e21"},{"id":"epoch:851525f91a090b541ac0:parameters","entityId":"epoch:851525f91a090b541ac0","metric":"parameters","value":3000000000000.0,"unit":"parameters","observationPeriod":"2021-09-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 1\n\nhttps://arxiv.org/abs/2104.05158"},{"id":"epoch:851525f91a090b541ac0:training-compute","entityId":"epoch:851525f91a090b541ac0","metric":"training-compute","value":1.1e+21,"unit":"FLOP","observationPeriod":"2021-09-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 1\n\nhttps://arxiv.org/abs/2104.05158"},{"id":"epoch:8354e838bd1dd537bdcf:parameters","entityId":"epoch:8354e838bd1dd537bdcf","metric":"parameters","value":45000000.0,"unit":"parameters","observationPeriod":"2021-09-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"MegaMolBART has eight layers, four attention heads, a hidden space dimension of 256, and contains 45M parameters\""},{"id":"epoch:8354e838bd1dd537bdcf:training-compute","entityId":"epoch:8354e838bd1dd537bdcf","metric":"training-compute","value":7.2e+20,"unit":"FLOP","observationPeriod":"2021-09-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"MegaMolBART was trained with data parallelism on 64 V100 32 GB GPUs (4 nodes x 16 GPUs) for 8 epochs (approximately 160k iterations or ~80 wall-clock hours) using a batch size of 32 molecules per GPU (micro batch)\"\n\nhttps://docs.nvidia.com/bionemo-framework/0.4.0/models/megamolbart.html\n\n64 * 130 teraflops * 80 * 3600 * 0.3 = 7.2e20"},{"id":"epoch:84d9d55c65d28b197888:parameters","entityId":"epoch:84d9d55c65d28b197888","metric":"parameters","value":82000000000.0,"unit":"parameters","observationPeriod":"2021-09-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We introduce a Korean in-context large-scale LM with 82B parameters, i.e., HyperCLOVA. This is the first discovery on near\n100B-scale non-English LM.\"\n\nAccording to media reports, HyperCLOVA has 204B parameters (i.e. a different version than in the paper)\nhttps://m.koreaherald.com/view.php?ud=20210525000824 "},{"id":"epoch:84d9d55c65d28b197888:training-compute","entityId":"epoch:84d9d55c65d28b197888","metric":"training-compute","value":1.476e+23,"unit":"FLOP","observationPeriod":"2021-09-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For experiments in Section 4, the model trained with 150B is used for fair comparison, because not all models are finished training at the same iteration. However, experiments in Section 5.2 use the model trained with 300B tokens, as HyperCLOVA Studio provided the 39B and 82B models trained with 300B tokens.\"\n\n82e9 connections * 2 FLOP/connection * 300e9 tokens * 3 backward pass = 1.476e23 FLOP\n\nCalculation using GPU time corroborates this:\n- \"Our model is based on megatron-LM (Shoeybi et al., 2019) and trained on the NVIDIA Superpod, which includes 128 strongly clustered DGX servers with 1,024 A100 GPUs.\"\n- \"It takes 13.4 days to train a model with 82B parameters with 150B tokens.\" Assume 300B tokens takes twice as long, 26.8 days.\n- Assume the default of 30% utilization rate for large language models.\n\n1024 A100 GPUs * 312e12 FLOP/second * 0.3 utilization * 26.8 days * 24 * 60 * 60 seconds/day = 2.219e+23 FLOP"},{"id":"epoch:e76fa04e66c09d5dd0a3:parameters","entityId":"epoch:e76fa04e66c09d5dd0a3","metric":"parameters","value":204000000000.0,"unit":"parameters","observationPeriod":"2021-09-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.navercorp.com/navercorp_/ir/announce/2023/NAVER_CEO%20letter%20to%20shareholders_Aug%202023_Eng.pdf"},{"id":"epoch:e76fa04e66c09d5dd0a3:training-compute","entityId":"epoch:e76fa04e66c09d5dd0a3","metric":"training-compute","value":2.0000000001e+23,"unit":"FLOP","observationPeriod":"2021-09-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Estimations for 82B model (marked as lower bound estimations)\n\n\"For experiments in Section 4, the model trained with 150B is used for fair comparison, because not all models are finished training at the same iteration. However, experiments in Section 5.2 use the model trained with 300B tokens, as HyperCLOVA Studio provided the 39B and 82B models trained with 300B tokens.\"\n\n82e9 connections * 2 FLOP/connection * 300e9 tokens * 3 backward pass = 1.476e23 FLOP\n\nCalculation using GPU time corroborates this:\n- \"Our model is based on megatron-LM (Shoeybi et al., 2019) and trained on the NVIDIA Superpod, which includes 128 strongly clustered DGX servers with 1,024 A100 GPUs.\"\n- \"It takes 13.4 days to train a model with 82B parameters with 150B tokens.\" Assume 300B tokens takes twice as long, 26.8 days.\n- Assume the default of 30% utilization rate for large language models.\n\n1024 A100 GPUs * 312e12 FLOP/second * 0.3 utilization * 26.8 days * 24 * 60 * 60 seconds/day = 2.219e+23 FLOP"},{"id":"epoch:1a3d8704df88a6902c38:parameters","entityId":"epoch:1a3d8704df88a6902c38","metric":"parameters","value":247000512.0,"unit":"parameters","observationPeriod":"2021-09-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The core language model is semi-parametric. There is a non-parametric component which involves learning a datastore which maps the training data to context vectors. This paper improves on the kNN-LM model primarily by reducing the size of the datastore to improve throughput. They also add a small 512 parameter neural net to adaptively adjust the interpolation hyperparameter.\n\nThey use the pre-trained model from https://arxiv.org/abs/1809.10853, which had 247M trainable parameters, so the model in this paper one has 247M + 512 parameters."},{"id":"epoch:1a3d8704df88a6902c38:training-compute","entityId":"epoch:1a3d8704df88a6902c38","metric":"training-compute","value":2.783e+19,"unit":"FLOP","observationPeriod":"2021-09-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The base model used was 247M parameters and trained for 286k updates with an effective batch size of 65,536 tokens.\n\n6 * 247M * 286k * 65536 = 2.778e19 FLOPs\n\nGenerating the datastore involves a single forward pass over all of the training examples, for another\n2 * 247M * 103M = 5.088e16 FLOPs\n\nThey also train the small retrieval adapter, but this converges in \"several minutes with a single GPU\" and so is unlikely to contribute significantly.\n\nThe only other computationally expensive step is the k-means datastore pruning: \"we perform 5000 separate k-means clustering passes only for the most frequent 5000 words due to high computational cost\". These are likely computed on CPUs; they have 32 CPU cores from 1.5GHz AMD EPYC 7282. Assuming that this means 32 processors, and not the 32 threads in a single processor, and assuming 4 cycles per multiplication, these CPUs could do 32 * 32 * (1.5e9 / 4) = 3.84e11 FLOP/s, so putting a dent in the 2.8e19 FLOPs from the base model would take a prohibitive amount of time.\n\nTotal: 2.783e19"},{"id":"epoch:66d22337766b6a0ab71e:parameters","entityId":"epoch:66d22337766b6a0ab71e","metric":"parameters","value":149697024.0,"unit":"parameters","observationPeriod":"2021-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Parameterization appears to be similar to a vanilla transformer. 6 layers, hidden dimension of 512, feed forward dimension of 1024, 8 attention heads. This would imply 20,447,232 parameters without embedding weights, and 512*vocab_size embedding weights (assuming tied embedding and unembedding projections)\n\nThey appear to use word-level tokenization: \"We evaluate unidirectional PermuteFormer on WikiText-103 (Merity et al., 2017). It is a language modeling dataset with about 103 million tokens,\" and I confirmed that word-level tokenization results in about 102M words across train-test-validation.\n\nIf this is the case, there are 267,735 unique words, so the embedding layer alone would be 137,080,320 parameters, for a total of 149,697,024."},{"id":"epoch:66d22337766b6a0ab71e:training-compute","entityId":"epoch:66d22337766b6a0ab71e","metric":"training-compute","value":2.775e+18,"unit":"FLOP","observationPeriod":"2021-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 * (30 * 103M) * 149,697,024 = 2.775e18\n\nThis seems a bit small relative to their statement: \"It takes about 10 days on 8 V100 GPUs to get all the figures in this paper\" which suggests about 2.7e20 FLOPs at 30% MFU.\n\nTable 2 indicates that Performer and PermuteFormer take 0.23x to 0.58x as long to train as a Transformer model. Figure 2 appears to be the most compute intensive figure, and would take about 4 * (2.775e18) + 1 * (2.775e18 / 0.365) = 1.9e19 FLOPs. "},{"id":"epoch:a2fceb709dcdf656bc96:parameters","entityId":"epoch:a2fceb709dcdf656bc96","metric":"parameters","value":5751892.0,"unit":"parameters","observationPeriod":"2021-09-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:27f4a5ad2c65d99b49e6:parameters","entityId":"epoch:27f4a5ad2c65d99b49e6","metric":"parameters","value":135000000000.0,"unit":"parameters","observationPeriod":"2021-09-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See paper title"},{"id":"epoch:ccdcb314c386c3128618:parameters","entityId":"epoch:ccdcb314c386c3128618","metric":"parameters","value":137000000000.0,"unit":"parameters","observationPeriod":"2021-09-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Abstract:\n\"We take a 137B parameter pretrained language model and instruction tune it on over 60 NLP datasets verbalized via natural language instruction templates. We evaluate this instruction-tuned model, which we call FLAN, on unseen task types.\"\n\nMany models seem to be using the same 137B base transformer model?"},{"id":"epoch:ccdcb314c386c3128618:training-compute","entityId":"epoch:ccdcb314c386c3128618","metric":"training-compute","value":2.047e+24,"unit":"FLOP","observationPeriod":"2021-09-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From section 2.4: Pretraining was done over 2.49T tokens.\n6 * 2.49T * 137B = 2.047e24 \nAlso, \"instruction tuning takes around 60 hours on a TPUv3 with 128 cores.\" 128 TPUv3 cores = 64 TPUv3 chips. Environmental considerations section claims this took less than 2% of total time\n1.23e14 * 64 * 60 * 3600 * 0.3 = 5.10e20"},{"id":"epoch:a5eb49dcc0654a5c4447:parameters","entityId":"epoch:a5eb49dcc0654a5c4447","metric":"parameters","value":124000000.0,"unit":"parameters","observationPeriod":"2021-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"assuming same as the base model GPT-2 small"},{"id":"epoch:a5eb49dcc0654a5c4447:training-compute","entityId":"epoch:a5eb49dcc0654a5c4447","metric":"training-compute","value":1.2e+22,"unit":"FLOP","observationPeriod":"2021-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.936 \u00d7 10^20 FLOP [base model compute estimation, \"Speculative\" confidence] + 1.488e+17 FLOP [fine-tune compute] =  7.937488e+20 FLOP"},{"id":"epoch:2be7e2bbebe4e75eca7c:parameters","entityId":"epoch:2be7e2bbebe4e75eca7c","metric":"parameters","value":16000000.0,"unit":"parameters","observationPeriod":"2021-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f9a70bb0e0604ee2e868:parameters","entityId":"epoch:f9a70bb0e0604ee2e868","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2021-08-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f9a70bb0e0604ee2e868:training-compute","entityId":"epoch:f9a70bb0e0604ee2e868","metric":"training-compute","value":8.1e+20,"unit":"FLOP","observationPeriod":"2021-08-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From figure 5, 6000 GPU hours (Nvidia V100) \n\n6000*  125 teraflop/s * 3600 * 0.3 = 8.1e20"},{"id":"epoch:743ab3cdb2ee7d5c54ce:parameters","entityId":"epoch:743ab3cdb2ee7d5c54ce","metric":"parameters","value":10700000000.0,"unit":"parameters","observationPeriod":"2021-08-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 2.1:\n\" ...XLM-RXXL (L= 48, H = 4096, A = 32, 10.7B params)\""},{"id":"epoch:743ab3cdb2ee7d5c54ce:training-compute","entityId":"epoch:743ab3cdb2ee7d5c54ce","metric":"training-compute","value":3.366e+22,"unit":"FLOP","observationPeriod":"2021-08-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Trained for 500k steps at a batch size of 2048 with sequence length of 512 = 524,288,000,000 tokens seen.\n6 * 10700000000 * 524,288,000,000 = 3.366e22"},{"id":"epoch:dfbcd134d88f116da911:parameters","entityId":"epoch:dfbcd134d88f116da911","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2021-08-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We have trained multiple largescale models on the PFAM[7] dataset, the largest with 3 billion parameters\""},{"id":"epoch:dfbcd134d88f116da911:training-compute","entityId":"epoch:dfbcd134d88f116da911","metric":"training-compute","value":1.6e+22,"unit":"FLOP","observationPeriod":"2021-08-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We pretrained two large models on a 480 GPUs (TeslaV100-32GB) cluster for about three weeks\"\n\n21 * 24* 3600 * 480 * 125 teraFLOP/s * 0.3 (utilization) * 0.5 (two models) = 1.6e22"},{"id":"epoch:b4683206ac90f1484cb9:parameters","entityId":"epoch:b4683206ac90f1484cb9","metric":"parameters","value":110000000.0,"unit":"parameters","observationPeriod":"2021-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We used the same model architecture as the BERT base, which consists of 12 Transformer layers with 768 hidden units and 12 attention heads in each layer, and the same parameter setting across all the four DNABERT models during pre-training\"\n\nKnown to have 110 million parameters as reported in: https://arxiv.org/pdf/1810.04805v2.pdf\n\"We primarily report results on two model sizes: BERTBASE (L=12, H=768, A=12, Total Parameters=110M) [...]\""},{"id":"epoch:b4683206ac90f1484cb9:training-compute","entityId":"epoch:b4683206ac90f1484cb9","metric":"training-compute","value":1.07e+20,"unit":"FLOP","observationPeriod":"2021-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Since the pre-training of DNABERT model is resource-intensive (about 25\u2009days on 8 NVIDIA 2080Ti GPUs)\"\n\nAssuming FP16 and 30% utilization\n\nCalculation = (25 * 24 *3600) s * 2.7e13 FLOP/s per GPU * 8 GPUs * 0.3 utilization = 1.4e20 FLOP\n\nAlternatively:\n\"DNABERT takes a sequence with a max length of 512 as input... We pre-trained DNABERT for 120k steps with a batch size of 2000\"\n6 * 512 * 2000 * 120k * 110M = 8.11e19\n\nGeometric mean: 1.07e20"},{"id":"epoch:a9e98b70a6b89103035e:parameters","entityId":"epoch:a9e98b70a6b89103035e","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2021-08-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a9e98b70a6b89103035e:training-compute","entityId":"epoch:a9e98b70a6b89103035e","metric":"training-compute","value":2.3811e+21,"unit":"FLOP","observationPeriod":"2021-08-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Experiments replicating GPT-2 models (117M and 1.5B) show that <..> our method reduces the required number of training tokens and wall clock time by up to 2.2x and 3.7x, respectively.\"\n\nGPT-2 estimated training compute: 1.9200000000009998e+21 FLOP (Speculative confidence) -> \n1.9200000000009998e+21 / 3.7 = 5.1891892e+20 FLOP\n\n\" All of the experiments are performed on 128 NVIDIA V100 GPUs (32GB memory). There are 16 nodes and 8 GPUs per node\"\n155 hours (Table 2) \n\n125000000000000 FLOP/sec * 128 GPUs * 155 hours * 3600 sec / hour * 0.3 = 2.6784e+21 FLOP\n\nbatch size 4000\nsequence length 1000\n58.8K steps\n\n6 FLOP / parameter / token * 1.5 * 10^9 parameters * 4000 sequences per batch * 1000 token per sequence * 58800 steps = 2.1168e+21 FLOP\n\ngeometric mean of hardware and operation counting method (they are more accurate than comparison with GPT-2): \nsqrt(2.6784e+21*2.1168e+21) = 2.3811e+21 FLOP"},{"id":"epoch:4d022a4e5ac97cc67688:parameters","entityId":"epoch:4d022a4e5ac97cc67688","metric":"parameters","value":117000000.0,"unit":"parameters","observationPeriod":"2021-08-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"batch size 512, 300K total training steps (157B tokens)"},{"id":"epoch:4d022a4e5ac97cc67688:training-compute","entityId":"epoch:4d022a4e5ac97cc67688","metric":"training-compute","value":1.10214e+20,"unit":"FLOP","observationPeriod":"2021-08-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND formula: \n6*157000000000*117000000=1.10214e+20"},{"id":"epoch:56e982341ba13a392306:parameters","entityId":"epoch:56e982341ba13a392306","metric":"parameters","value":178000000000.0,"unit":"parameters","observationPeriod":"2021-08-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Jurassic-1 models come in two sizes, where the Jumbo version, at 178B parameters, is the largest and most sophisticated language model ever released for general use by developers.\""},{"id":"epoch:56e982341ba13a392306:training-compute","entityId":"epoch:56e982341ba13a392306","metric":"training-compute","value":3.7e+23,"unit":"FLOP","observationPeriod":"2021-08-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"see here https://docs.google.com/document/d/1B8x6XYcmB1u6Tmq3VcbAtj5bzhDaj2TcIPyK6Wpupx4/edit\n\n6 * 178B * 300B = 3.204000e+23"},{"id":"epoch:f32494dccafb2050bffa:parameters","entityId":"epoch:f32494dccafb2050bffa","metric":"parameters","value":3200000000.0,"unit":"parameters","observationPeriod":"2021-08-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\u517132\u4ebf\u53c2\u6570 translated as A total of 3.2 billion parameters "},{"id":"epoch:f32494dccafb2050bffa:training-compute","entityId":"epoch:f32494dccafb2050bffa","metric":"training-compute","value":8.016e+20,"unit":"FLOP","observationPeriod":"2021-08-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4.175e10 * 3.2e9 * 6 = 8.016e20 FLOP"},{"id":"epoch:0d5f2a39c435c79a8e9c:parameters","entityId":"epoch:0d5f2a39c435c79a8e9c","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2021-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1B for XXL model"},{"id":"epoch:98c4bdc81e92e61b2a04:parameters","entityId":"epoch:98c4bdc81e92e61b2a04","metric":"parameters","value":99100000.0,"unit":"parameters","observationPeriod":"2021-08-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"99.1M, table 3"},{"id":"epoch:98c4bdc81e92e61b2a04:training-compute","entityId":"epoch:98c4bdc81e92e61b2a04","metric":"training-compute","value":6.34275e+20,"unit":"FLOP","observationPeriod":"2021-08-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train the models for a total of 300 epochs with 5 epochs warmup on COCO train2017 [17]. We use stochastic gradient descent (SGD) for training ... The batch size is 128 by default to typical 8-GPU devices ... input size is evenly drawn from 448 to 832 with 32 strides\"\n\nTraining is done on 300 epochs of the 2.5 million image-label pairs in COCO train2017.\n\nTable 3 indicates 281.9 GFLOP per forward pass on a 640x640 image. The mean image width/height is 640, though using this to estimate training FLOPs is probably a slight underestimate as FLOPs will scale roughly linearly in the number of pixels (which scale at width^2).\n\nIgnoring this slight issue: 281.9e9 * 2.5M * 300 * 3 = 6.34e20"},{"id":"epoch:0834d70926a110c977fa:parameters","entityId":"epoch:0834d70926a110c977fa","metric":"parameters","value":40000000.0,"unit":"parameters","observationPeriod":"2021-08-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"40M (caption to Table 2)"},{"id":"epoch:0834d70926a110c977fa:training-compute","entityId":"epoch:0834d70926a110c977fa","metric":"training-compute","value":2.472e+16,"unit":"FLOP","observationPeriod":"2021-08-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 40 * 10^6 parameters * 103*10^6 tokens * 1 epoch [assumed, not given] = 2.472e+16 FLOP\n\n________\nestimation from the Algorithmic progress paper: 4.3*10^17 FLOP"},{"id":"epoch:44257df680180ad7b835:parameters","entityId":"epoch:44257df680180ad7b835","metric":"parameters","value":5000000.0,"unit":"parameters","observationPeriod":"2021-08-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Agent parameter counts were all 5 \u2212 6 million parameters, excluding parameters in auxiliary modules\""},{"id":"epoch:512f70100fb382c241a1:parameters","entityId":"epoch:512f70100fb382c241a1","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2021-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From abstract:\n\" Our final SElf-supERvised (SEER) model, a RegNetY with 1.3B parameters...\""},{"id":"epoch:512f70100fb382c241a1:training-compute","entityId":"epoch:512f70100fb382c241a1","metric":"training-compute","value":1.8e+22,"unit":"FLOP","observationPeriod":"2021-07-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Numbers from section 3.2, they specifically mention using mixed precision training.\n6125 ms / batch * 114890 batches = 8.14 days (they round to 8 in the text)\n\n512 GPUs * 8.14 days * 24h/day * 3600s/h * 125 TFLOP/s * 0.4 (assumed utilization) = 1.800e22\n\n\"on 512 V100 32GB NVIDIA GPUs. Training this model on 1 billion images requires 114, 890 training iterations for a batch size of 8, 704 images, summing to 8 days of training over 512 GPUs.\""},{"id":"epoch:ec03c5b5f407a9e69471:parameters","entityId":"epoch:ec03c5b5f407a9e69471","metric":"parameters","value":3472816.0,"unit":"parameters","observationPeriod":"2021-07-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"estimate described here: https://docs.google.com/document/d/1S9xZyCeITDOs-P1W_-liNW0WgVN-OLsSudVrPXMaLqw/edit?usp=sharing"},{"id":"epoch:ec03c5b5f407a9e69471:training-compute","entityId":"epoch:ec03c5b5f407a9e69471","metric":"training-compute","value":2.412e+22,"unit":"FLOP","observationPeriod":"2021-07-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"[Final calculation]\n(8 TPUs) * (1.23e14 FLOP/TPU-s) * (0.1 utilization) / (50k steps/s) = 1.968e9 FLOP/step\n\n(32 agents) * (383B steps/agent) * (1.968e9 FLOP/step) = 2.412e22 FLOPs\n\n==========================\nNOTES BELOW\n\n6.1: Each agent is trained using 8 TPUv3s and consumes approximately 50,000 agent steps (observations) per second.\nMultiple agents interacting probably mean a fairly low utilization rate, so let\u2019s assume 0.10\n8 * 1.23e14 * 0.1 / 50k = 1.968e9 FLOPs per step\n\nThe paper doesn\u2019t say exactly how many agents they train in each population. The original PBT paper uses 32 agents for one task (in general it uses between 10 and 80), so as a guesstimate let\u2019s go with that.\n\nFigure 16: They train over 5 generations. Summing the number of steps, it looks like there were roughly 383B steps\n32 * 383B * 1.968e9 = 2.412e22\n\nFinal estimate:\n2.412e22\n\nI do a confidence interval analysis here and find a 90% CI of 6.9e21 to 1.3e23, so we can call this estimate \"likely\" (within 1 OOM): https://colab.research.google.com/drive/1wGSTQxBExY6Fa0-d7msVumf5-KnsWLe6?usp=sharing"},{"id":"epoch:6e55f27849b551578a1a:parameters","entityId":"epoch:6e55f27849b551578a1a","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2021-07-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From abstract:\n\"Using a 1B parameter model, HuBERT shows up to 19% and 13% relative WER reduction on the more challenging dev-other and test-other evaluation subsets\""},{"id":"epoch:6e55f27849b551578a1a:training-compute","entityId":"epoch:6e55f27849b551578a1a","metric":"training-compute","value":5.54e+21,"unit":"FLOP","observationPeriod":"2021-07-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"GPU NOT SPECIFIED - for the sake of argument I assume something on the order of 1 TFLOP/s\n\nNumbers from Section IV part C\n0.1 * (960h * 32GPUs + 60000h * 256 GPUs) * 3600s/h * 1 TFLOP/s/GPU"},{"id":"epoch:2d856e6e6841d3ec7319:parameters","entityId":"epoch:2d856e6e6841d3ec7319","metric":"parameters","value":34300000.0,"unit":"parameters","observationPeriod":"2021-07-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6290efc9bd438bd48718:parameters","entityId":"epoch:6290efc9bd438bd48718","metric":"parameters","value":12000000000.0,"unit":"parameters","observationPeriod":"2021-07-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"With just a single sample, a 12B parameter Codex solves 28.8% of these problems, and a 300M parameter Codex solves 13.2% of these problems\""},{"id":"epoch:6290efc9bd438bd48718:training-compute","entityId":"epoch:6290efc9bd438bd48718","metric":"training-compute","value":7.344e+22,"unit":"FLOP","observationPeriod":"2021-07-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The original training of GPT-3-12B consumed hundreds of petaflop/sdays of compute, while fine-tuning it to create Codex-12B\nconsumed a similar amount of compute.\"\n1 PFLOP/s-day = 8.64e19 FLOPs.\n\"Hundreds\" is likely between 200 and 900, geometric mean = 425.\n2 * 425 * 8.64e19 = 7.344e22\n"},{"id":"epoch:d84bda0c90183e2afb86:parameters","entityId":"epoch:d84bda0c90183e2afb86","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2021-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We trained the model with 10 billion parameters on a 4TB corpus consisting of plain texts and a large-scale knowledge graph.\""},{"id":"epoch:d84bda0c90183e2afb86:training-compute","entityId":"epoch:d84bda0c90183e2afb86","metric":"training-compute","value":2.25e+22,"unit":"FLOP","observationPeriod":"2021-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 3.3.3: \n\"\"The model is trained for a total of 375 billion tokens\"\n\nTotal compute approximated as 6*N*D"},{"id":"epoch:a5619895277a09af5b75:parameters","entityId":"epoch:a5619895277a09af5b75","metric":"parameters","value":1900000.0,"unit":"parameters","observationPeriod":"2021-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.9M parameters"},{"id":"epoch:a5619895277a09af5b75:training-compute","entityId":"epoch:a5619895277a09af5b75","metric":"training-compute","value":7.34832e+17,"unit":"FLOP","observationPeriod":"2021-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"11340000000000 FLOP / GPU / sec [fp32 assumed] * 60 hours * 3600 sec / hour * 1 GPU * 0.3 [assumed utilization] = 7.34832e+17 FLOP"},{"id":"epoch:7ade6ffc79fd1e29c47c:parameters","entityId":"epoch:7ade6ffc79fd1e29c47c","metric":"parameters","value":98000000.0,"unit":"parameters","observationPeriod":"2021-06-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"98M (Table 1)"},{"id":"epoch:7ade6ffc79fd1e29c47c:training-compute","entityId":"epoch:7ade6ffc79fd1e29c47c","metric":"training-compute","value":2.9e+19,"unit":"FLOP","observationPeriod":"2021-06-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"26900000000000 FLOP / sec [assumed precision: FP16] *4 GPUs * 250 hours * 3600 sec / hour * 0.3 [assumed utilization] = 2.9052e+19 FLOP\n\n6 FLOP / parameter / token * 98 * 10^6 parameters * 103*10^6 tokens * 23 epochs = 1.392972e+18 FLOP\n\n________________\nNB reporting hardware estimation as it was used in the Algorithmic progress paper"},{"id":"epoch:37481dae21574ea44232:parameters","entityId":"epoch:37481dae21574ea44232","metric":"parameters","value":247000000.00000003,"unit":"parameters","observationPeriod":"2021-06-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:37481dae21574ea44232:training-compute","entityId":"epoch:37481dae21574ea44232","metric":"training-compute","value":8.58e+19,"unit":"FLOP","observationPeriod":"2021-06-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train Transformer model for 50k steps and Adaptive Input Transformer for 286k steps\"\n\nassuming 1 step took 1 second:\n\n125000000000000 FLOP / sec [assumed precision: bf16] * 286000 seconds * 8 GPUs * 0.3 [assumed utilization] = 8.58e+19 FLOP"},{"id":"epoch:5ea191bb5477a0d5189a:parameters","entityId":"epoch:5ea191bb5477a0d5189a","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2021-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we pretrain two models: an encoder-decoder bilingual model with 11 billion parameters (CPM2) and its corresponding MoE version with\n198 billion parameters. \""},{"id":"epoch:5ea191bb5477a0d5189a:training-compute","entityId":"epoch:5ea191bb5477a0d5189a","metric":"training-compute","value":3.600000000001e+22,"unit":"FLOP","observationPeriod":"2021-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"epoch count not stated\nC = 6 FLOP * 11B * 464B = "},{"id":"epoch:90a2bdd34335acc87ee8:parameters","entityId":"epoch:90a2bdd34335acc87ee8","metric":"parameters","value":12427904.0,"unit":"parameters","observationPeriod":"2021-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Three components, a sequence encoder, fold encoder, and sequence decoder. (1) and (3) have similar architectures, a sequence embedder + transformer block (the encoder uses a transformer encoder, the decoder uses a decoder block). \n\nThe fold encoder adds 6 residual conv blocks and has a 3d positional encoder rather than sequence embedder. Each residual block has two 3D-convolutional layers (3\u00d73\u00d73) and batch normalization layers.\n\nEach transformer block has 4 layers and d = 256 latent dimensions.\n\nCalculations here: https://docs.google.com/document/d/1luTCTQLmfaBfmnjbsBpWuG51x_IwuiN7mqNoOXtoi8Y/edit?usp=sharing"},{"id":"epoch:90a2bdd34335acc87ee8:training-compute","entityId":"epoch:90a2bdd34335acc87ee8","metric":"training-compute","value":1.4e+17,"unit":"FLOP","observationPeriod":"2021-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train our model on 2 Tesla K80 GPUs, with batch size 128. In every training stage we train up to 200 epochs with an early stopping strategy based on the validation loss\"\nSee calculations here: https://docs.google.com/document/d/1luTCTQLmfaBfmnjbsBpWuG51x_IwuiN7mqNoOXtoi8Y/edit?usp=sharing\n\nBlock 1:\nCNN 1: 2*20*20*20*3*3*3*4*8=13824000\nCNN 2: 2*20*20*20*3*3*3*8*8=27648000\nBlock 2:\nCNN 1: 2*20*20*20*3*3*3*8*16=55296000\nCNN 2: 2*20*20*20*3*3*3*16*16=110592000\nBlock 3:\nCNN 1: 2*20*20*20*3*3*3*16*32=221184000\nCNN 2: 2*20*20*20*3*3*3*32*32=442368000\nBlock 4:\nCNN 1: 2*10*10*10*3*3*3*32*64=110592000\nCNN 2: 2*10*10*10*3*3*3*64*64=221184000\nBlock 5:\nCNN 1: 2*10*10*10*3*3*3*64*128=442368000\nCNN 2: 2*10*10*10*3*3*3*128*128=884736000\nBlock 6:\nCNN 1: 2*10*10*10*3*3*3*128*256=1769472000\nCNN 2: 2*10*10*10*3*3*3*256*256=3538944000\nCNN forward FLOP: 13824000+27648000+55296000+110592000+221184000+442368000+110592000+221184000+442368000+884736000+1769472000+3538944000=7838208000\nTraining FLOP: 45995*200*(3299020800+4074393600+7838208000)=1.4e17\n\n\nNote I make several assumptions about the CNN architecture which could change this value."},{"id":"epoch:ba9ad11cbfd3c19ad8cb:parameters","entityId":"epoch:ba9ad11cbfd3c19ad8cb","metric":"parameters","value":208000000.0,"unit":"parameters","observationPeriod":"2021-06-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"208M for XL version (Table 7, page 7)"},{"id":"epoch:ba9ad11cbfd3c19ad8cb:training-compute","entityId":"epoch:ba9ad11cbfd3c19ad8cb","metric":"training-compute","value":9.56e+19,"unit":"FLOP","observationPeriod":"2021-06-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 7, page 7: 45 hours on 32 TPUv3 cores.\n\n\"Each v3 TPU chip contains two TensorCores.\"\nTPU performance per chip = 123e12 FLOP/s\n32 cores = 16 chips\n\n123e12 FLOP/s per chip * (32 cores / 2 cores per chip) * 45 hours * 3600 seconds/hour * 0.30 utilization = 9.56e19 FLOP\n\nhttps://www.wolframalpha.com/input?i=123+terahertz+*+16+*+45+hours+*+0.3"},{"id":"epoch:64fcb33f785c97ca3cec:parameters","entityId":"epoch:64fcb33f785c97ca3cec","metric":"parameters","value":2230000.0,"unit":"parameters","observationPeriod":"2021-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:64fcb33f785c97ca3cec:training-compute","entityId":"epoch:64fcb33f785c97ca3cec","metric":"training-compute","value":1.70208e+21,"unit":"FLOP","observationPeriod":"2021-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"125000000000000 FLOP / GPU / sec [V100] * 8 GPUs * 1576 hours [see training time notes] * 3600 sec / hour * 0.3 [assumed utilization] = 1.70208e+21 FLOP"},{"id":"epoch:a57665a7177289132dc8:parameters","entityId":"epoch:a57665a7177289132dc8","metric":"parameters","value":1580000.0,"unit":"parameters","observationPeriod":"2021-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a57665a7177289132dc8:training-compute","entityId":"epoch:a57665a7177289132dc8","metric":"training-compute","value":2.42784e+21,"unit":"FLOP","observationPeriod":"2021-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"125000000000000 FLOP / GPU / sec [V100] * 8 GPUs * 2248 hours [see training time notes] * 3600 sec / hour * 0.3 [assumed utilization] = 2.42784e+21 FLOP"},{"id":"epoch:16402d0b97d2a93b19b3:parameters","entityId":"epoch:16402d0b97d2a93b19b3","metric":"parameters","value":820000000.0,"unit":"parameters","observationPeriod":"2021-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From author communication\n\n480M (image tower) + 340 M (text tower)"},{"id":"epoch:16402d0b97d2a93b19b3:training-compute","entityId":"epoch:16402d0b97d2a93b19b3","metric":"training-compute","value":2.598670000001e+22,"unit":"FLOP","observationPeriod":"2021-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From author communication\n14.82K TPUv3 core-days\nPrecision: bfloat16\n\nEstimation\nTPUv3 at float16: 123 TFLOPS/chip\n\n123*10^12 TFLOPS/chip * (1 chip / 2 cores) * 14820 TPU core-days * 86400 s/day * 33% utilization = 2.599*10^22 FLOP\nhttps://www.wolframalpha.com/input?i=14820+days+*+123+teraFLOPS+%2F+2+*+0.33"},{"id":"epoch:9fda6ae97e965c8bd26d:parameters","entityId":"epoch:9fda6ae97e965c8bd26d","metric":"parameters","value":256000000.0,"unit":"parameters","observationPeriod":"2021-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Appendix B: \n\" Our CIFAR10 model has 35.7 million parameters, and our LSUN and\nCelebA-HQ models have 114 million parameters. We also trained a larger variant of the LSUN Bedroom model with approximately 256 million parameters by increasing filter count.\""},{"id":"epoch:9fda6ae97e965c8bd26d:training-compute","entityId":"epoch:9fda6ae97e965c8bd26d","metric":"training-compute","value":7.840125000000001e+19,"unit":"FLOP","observationPeriod":"2021-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Numbers in Appendix B\n\"Our CelebA-HQ/LSUN (2562) models train at 2.2 steps per second at batch size 64, [...] The larger LSUN Bedroom model was trained for 1.15M steps.\"\n10.6h for the CIFAR model (batch size 128, 21 step/s)\n2.2 step/s for the LSUN model, 1.15M steps so 702.8 hours\n\n1 step takes 1/2.2 =0.4545 seconds\n1.15M steps * 0.4545 seconds = 522675 seconds = 145 hours\n\nThis is for TPUv3-8's, which seems to mean 8 cores (standard chip is 125 teraflop/s for 2 cores) -> 4 chips\nhttps://cloud.google.com/tpu/docs/regions-zones\n\n1.25E14 FLOP/s * (4 chips) * 522675 seconds * 0.3 = 78401250000000010000"},{"id":"epoch:168b6ed88944f850d26a:parameters","entityId":"epoch:168b6ed88944f850d26a","metric":"parameters","value":44600000.0,"unit":"parameters","observationPeriod":"2021-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"44.6M"},{"id":"epoch:168b6ed88944f850d26a:training-compute","entityId":"epoch:168b6ed88944f850d26a","metric":"training-compute","value":1.1000000000000001e+18,"unit":"FLOP","observationPeriod":"2021-06-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 44600000 parameters * 103000000 tokens * 40 epochs [assumed -> \"Speculative\" confidence] = 1.102512e+18 FLOP"},{"id":"epoch:f9bb1dfc5767d9125a62:parameters","entityId":"epoch:f9bb1dfc5767d9125a62","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2021-06-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"...we scale up DeBERTa by training a larger version that consists of 48 Transform layers with 1.5 billion parameters\"\n\nOther versions are smaller and use a smaller pre-training dataset. These are distinguished in the paper (e.g. DeBERTa1.5B is the version of DeBERTa with 1.5 billion parameters)."},{"id":"epoch:f9bb1dfc5767d9125a62:training-compute","entityId":"epoch:f9bb1dfc5767d9125a62","metric":"training-compute","value":2.588e+22,"unit":"FLOP","observationPeriod":"2021-06-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 8: 16 DGX-2 nodes (x16 V100s each) for 30 days\n16 * 16 * 1.3e14 * 30 * 24 * 3600 * 0.3 = 2.588e22"},{"id":"epoch:d9f825732dd4a7990627:parameters","entityId":"epoch:d9f825732dd4a7990627","metric":"parameters","value":2440000000.0,"unit":"parameters","observationPeriod":"2021-06-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d9f825732dd4a7990627:training-compute","entityId":"epoch:d9f825732dd4a7990627","metric":"training-compute","value":4.27e+22,"unit":"FLOP","observationPeriod":"2021-06-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"20.1K TPU-v3 core-days\n\nTPUs have two cores per chip, and a chip is 123 teraflop/s \nhttps://cloud.google.com/tpu/docs/system-architecture-tpu-vm#tpu_v3\n\n123 teraflop/s * 20100/2 * 24 * 3600 * 0.4 (utilization assumption for non-language models) = 4.27e22"},{"id":"epoch:9f10e2317b9f372f94d7:parameters","entityId":"epoch:9f10e2317b9f372f94d7","metric":"parameters","value":440000000.0,"unit":"parameters","observationPeriod":"2021-06-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:9f10e2317b9f372f94d7:training-compute","entityId":"epoch:9f10e2317b9f372f94d7","metric":"training-compute","value":1.91e+21,"unit":"FLOP","observationPeriod":"2021-06-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We run all of our experiments on a machine with 96 CPUs, 1.3TB physical memory, and 16 A100 GPUs. We use PyTorch (Paszke et al., 2019) to implement our proposed model. With this hardware setup, our experiments on NQ and TriviaQA took approximately 25 hours to complete, while experiments on WebQ took roughly 8 hours to complete. Before supervised training, we also perform a one-time unsupervised MSS pre-training for 82,000 steps that took roughly 1 week.\"\n\n1 week + 25 hours * 16 A100s\n= ~193 * 16 A100-hours\n= 193 * 16 * 3600 * 312 trillion * 0.3 = 1.04e21\n\nAdditionally, the model uses BERT, ICT, and T5 models. These required:\n- BERT: 6 * 110M parameters * (1M * 256 * 256) inputs = 4.33e19 FLOP\n- ICT: 6 * 220M parameters * (100k * 4096 * 256) inputs = 1.38e20 FLOP\n- T5: 6 * 220M parameters * (1M * 2048 * 256) inputs = 6.92e20 FLOP\n\nTotal: 1.04e21 + 4.33e19 + 1.38e20 + 6.92e20 = 1.91e21"},{"id":"epoch:d1734bebc72c8059dace:parameters","entityId":"epoch:d1734bebc72c8059dace","metric":"parameters","value":1843000000.0,"unit":"parameters","observationPeriod":"2021-06-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2 of paper"},{"id":"epoch:d1734bebc72c8059dace:training-compute","entityId":"epoch:d1734bebc72c8059dace","metric":"training-compute","value":5.85e+22,"unit":"FLOP","observationPeriod":"2021-06-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Digitizing Figure 9 indicates training used 27,200 TPUv3 core-days. TPUv3 is 123 teraflop/s per chip, 2 cores per chip.\n\n1.23e14 * (1/2) * 27,200 * 24 * 3600 * 0.4 = 5.78e22\n\nAlternatively, Table 2 indicates 965.3e9 FLOPs per forward pass on a 224^2 image. Table 4 indicates 5 million steps at a (normalized) batch size of 4096, and total flops including backward pass would be 3x the FLOPs from forward passes alone, so we get:\n4096 * 5e6 * 965.3e9 * 3 = 5.93e22\n\n(Note that actual batch size appears to have been 32,768)\n\nGeometric mean: sqrt(5.78e22*5.93e22) = 5.85e22\n\nHowever note that this leaderboard claims ViT-G/14 took 34 PF-days, or 2.94e21 FLOPs: https://web.archive.org/web/20211218185755/https://lair.lighton.ai/akronomicon/"},{"id":"epoch:78161b493c3e73e44c37:parameters","entityId":"epoch:78161b493c3e73e44c37","metric":"parameters","value":2040000.0,"unit":"parameters","observationPeriod":"2021-06-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"- RNN with 1 hidden layer \n- 200 hidden units\n- Vocabulary size = 10,000\n\nThe RNN layer will have:\n\n- Input weight matrix U: (vocab size x hidden size) = (10,000 x 200) = 2,000,000 parameters\n\n- Recurrent weight matrix W: (hidden size x hidden size) = (200 x 200) = 40,000 parameters\n\n- Bias vector b: (hidden size) = (200) = 200 parameters \n\nTotal parameters:\n\nU = 2,000,000\nW = 40,000 \nb = 200\n\nTotal parameters = U + W + b\n            = 2,000,000 + 40,000 + 200\n            = 2,040,200\n\nTherefore, the total number of parameters with 200 hidden units and 10,000 vocab size is 2,040,200.\n\n_______\nfrom the Algorithmic progress paper spreadsheet"},{"id":"epoch:78161b493c3e73e44c37:training-compute","entityId":"epoch:78161b493c3e73e44c37","metric":"training-compute","value":223363710000000.0,"unit":"FLOP","observationPeriod":"2021-06-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 2,040,200 parameters  * 912344 tokens * 20 epochs = 2.2336371e+14 FLOP"},{"id":"epoch:827f28caa8f5cd963bc5:parameters","entityId":"epoch:827f28caa8f5cd963bc5","metric":"parameters","value":13600000.0,"unit":"parameters","observationPeriod":"2021-06-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Vocabulary size: 33000\nHidden size: 200\nParameters: 2*33000*200+200*200=13600000"},{"id":"epoch:827f28caa8f5cd963bc5:training-compute","entityId":"epoch:827f28caa8f5cd963bc5","metric":"training-compute","value":6528000000000000.0,"unit":"FLOP","observationPeriod":"2021-06-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Forward FLOP: 2*13600000=27200000\nTraining compute: 27200000*3*2000000*40=6.528e+15"},{"id":"epoch:98c51579a43d72ce66d4:parameters","entityId":"epoch:98c51579a43d72ce66d4","metric":"parameters","value":760000000.0,"unit":"parameters","observationPeriod":"2021-06-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:98c51579a43d72ce66d4:training-compute","entityId":"epoch:98c51579a43d72ce66d4","metric":"training-compute","value":4.9540697e+21,"unit":"FLOP","observationPeriod":"2021-06-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"base model compute (speculative confidence) 4.9536e+21 FLOP + fine-tune compute 4.6968e+17 FLOP = 4.9540697e+21 FLOP\n\n________\nestimation from the Arithmetic progress paper: 1.6E+21 FLOP"},{"id":"epoch:76d8f8a5bf6ae0cf9190:parameters","entityId":"epoch:76d8f8a5bf6ae0cf9190","metric":"parameters","value":1750000000000.0,"unit":"parameters","observationPeriod":"2021-05-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"It's been trained on 1.75 trillion parameters\"\n\nMoE architecture, \"tens of thousands of experts\"\nhttps://keg.cs.tsinghua.edu.cn/jietang/publications/wudao-3.0-meta-en.pdf"},{"id":"epoch:76d8f8a5bf6ae0cf9190:training-compute","entityId":"epoch:76d8f8a5bf6ae0cf9190","metric":"training-compute","value":1.54350000000001e+24,"unit":"FLOP","observationPeriod":"2021-05-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"It's a mixture-of-experts model, so all 1.75 trillion params were most likely not trained on each token.\n\n\"The parameter scale of Enlightenment 2.0 reached a record-breaking 1.75 trillion. According to reports, the new generation FastMoE technology is the key to the realization of the \"Trillion Model\" cornerstone of Enlightenment 2.0.\"\n\nSpeculatively assuming 3% of parameters are active per forward pass and the model was trained for one epoch: \n6 FLOP / token / parameter * 1.75 * 10^12 parameters * 0.03  * 4.9 * 10^12 tokens [see dataset size notes] = 1.54e+24 FLOP"},{"id":"epoch:71afb06857f9ae332d05:parameters","entityId":"epoch:71afb06857f9ae332d05","metric":"parameters","value":246930000.0,"unit":"parameters","observationPeriod":"2021-05-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:e88efb28c847f8e58298:parameters","entityId":"epoch:e88efb28c847f8e58298","metric":"parameters","value":12900000000.0,"unit":"parameters","observationPeriod":"2021-05-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12.9B, from Table 1"},{"id":"epoch:e88efb28c847f8e58298:training-compute","entityId":"epoch:e88efb28c847f8e58298","metric":"training-compute","value":8.1e+22,"unit":"FLOP","observationPeriod":"2021-05-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Like mT5, we set our sequence length to 1024 (bytes rather than tokens), and train for 1 million steps over batches of 2^20 tokens.\"\n\n12.9 billion * 1 million * 2^20 * 6 = ~8.1e22"},{"id":"epoch:cb298a0db2af0564eb05:parameters","entityId":"epoch:cb298a0db2af0564eb05","metric":"parameters","value":90100000.0,"unit":"parameters","observationPeriod":"2021-05-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table A2, NesT-B is the largest size."},{"id":"epoch:cb298a0db2af0564eb05:training-compute","entityId":"epoch:cb298a0db2af0564eb05","metric":"training-compute","value":2.40576e+19,"unit":"FLOP","observationPeriod":"2021-05-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"17.9 GFLOPS per forward pass\n300 epochs\n1.28M training examples\n3.5 f_to_b pass ratio\n(From Imagenet paper-data, Besiroglu et al., forthcoming) \n\n17.9e9 FLOP *300 epoch *1.28M images *3.5 forward-backward-ratio = 24057600000000000000"},{"id":"epoch:7cb2381e889efb8ac85b:parameters","entityId":"epoch:7cb2381e889efb8ac85b","metric":"parameters","value":4000000000.0,"unit":"parameters","observationPeriod":"2021-05-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We propose CogView, a 4-billion-parameter Transformer with VQ-VAE tokenizer to advance this problem.\""},{"id":"epoch:7cb2381e889efb8ac85b:training-compute","entityId":"epoch:7cb2381e889efb8ac85b","metric":"training-compute","value":2.68e+22,"unit":"FLOP","observationPeriod":"2021-05-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://lair.lighton.ai/akronomicon/\narchived: https://github.com/lightonai/akronomicon/tree/main/akrodb"},{"id":"epoch:0b910aba578a7140abf7:parameters","entityId":"epoch:0b910aba578a7140abf7","metric":"parameters","value":340000000.0,"unit":"parameters","observationPeriod":"2021-05-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0b910aba578a7140abf7:training-compute","entityId":"epoch:0b910aba578a7140abf7","metric":"training-compute","value":2.8e+20,"unit":"FLOP","observationPeriod":"2021-05-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Fine-tuning was done using a single Nvidia V100 GPU for a few minutes -> 1.0E+15 to 5.0E+15 (2 to 10 min)\n\nFoundation model is BeRT with 2.8e+20 FLOP.\n\nSo total compute is 2.8e+20."},{"id":"epoch:4c4a986bac6dd1196a1d:parameters","entityId":"epoch:4c4a986bac6dd1196a1d","metric":"parameters","value":17000000.0,"unit":"parameters","observationPeriod":"2021-05-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"17M from \"This is possibly due to the fact that the untrained Med-BERT is an over-parameterized model (around 17 million parameters) with a huge\nnumber of configurations, so it might overfit to the training data\""},{"id":"epoch:4c4a986bac6dd1196a1d:training-compute","entityId":"epoch:4c4a986bac6dd1196a1d","metric":"training-compute","value":9.47e+18,"unit":"FLOP","observationPeriod":"2021-05-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"flops = (1) * (3.13e13) * (24*7 * 3600) * (0.5) = 9.47e18\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)\nI assume higher utilization rate, because only 1 GPU is used.\nCitation from the text:\n\"We used a single Nvidia Tesla V100GPU of 32 GB graphics memory capacity, and we trained the model for a week for more than 45 million steps, for which each step consists of 32 patients (batch size).\" - page 11\n\nNote that public code appears not to make use of the tensor core speed up, thus I use 3.13e13 FLOP/sec"},{"id":"epoch:5b6d204521b2c63747db:parameters","entityId":"epoch:5b6d204521b2c63747db","metric":"parameters","value":559000000.0,"unit":"parameters","observationPeriod":"2021-05-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Largest model is denoted ImageNet 512, has 559M parameters"},{"id":"epoch:5b6d204521b2c63747db:training-compute","entityId":"epoch:5b6d204521b2c63747db","metric":"training-compute","value":6.2e+21,"unit":"FLOP","observationPeriod":"2021-05-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Largest run with their architecture improvements is the ImageNet 512 variant. Table 7 suggests utilization is around 30% for largest models (though we only see 256 x 256 and 128 -> 512)\n\nTable 10: ImageNet 512 variant took 1914 V100-days of training\n125e12 FLOP/sec * 1914 days * 24 h/day * 3600 sec/h * 0.3 = 6.2e21"},{"id":"epoch:fbbcc0907359ae7f4333:parameters","entityId":"epoch:fbbcc0907359ae7f4333","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2021-05-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://lair.lighton.ai/akronomicon/\n\narchived: https://github.com/lightonai/akronomicon/tree/main/akrodb"},{"id":"epoch:fbbcc0907359ae7f4333:training-compute","entityId":"epoch:fbbcc0907359ae7f4333","metric":"training-compute","value":7.37e+22,"unit":"FLOP","observationPeriod":"2021-05-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.37e22 from:\nsource: https://lair.lighton.ai/akronomicon/\narchived: https://github.com/lightonai/akronomicon/blob/main/akrodb/Technical%20University%20of%20Munich/ProtT5-XXL.json\n\n3.7E+22 from Table 9 https://www.biorxiv.org/content/10.1101/2023.07.05.547496v1\n\nManual calculation: forward passes on 512 * (4096*920k + 2048*343k) = 2.3T tokens\nBackward passes on 15% of those, 2.3T * 0.15 = 343B tokens.\nTotal FLOPs: (2 * 11B * 2.3T) + (4 * 11B * 343B) = 6.57e22"},{"id":"epoch:3d96e66b848c0bca3847:parameters","entityId":"epoch:3d96e66b848c0bca3847","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2021-05-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:3d96e66b848c0bca3847:training-compute","entityId":"epoch:3d96e66b848c0bca3847","metric":"training-compute","value":3.7e+22,"unit":"FLOP","observationPeriod":"2021-05-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"FLOP = 11B*2*(920k*512*4096) +  11B*4*(920k*512*4096), 920k steps using seq length 512 batch size 4096, "},{"id":"epoch:158a09c0e3cc91a59677:parameters","entityId":"epoch:158a09c0e3cc91a59677","metric":"parameters","value":420000000.0,"unit":"parameters","observationPeriod":"2021-05-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:158a09c0e3cc91a59677:training-compute","entityId":"epoch:158a09c0e3cc91a59677","metric":"training-compute","value":3.9e+22,"unit":"FLOP","observationPeriod":"2021-05-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"FLOP = 420M * 6 * (800k*512*32k + 200k*2048*6k) \n1M steps total split into two phases, (1) 800k steps, seq length 512 (batch size 32k) and (2) 200k steps, seq length 2048 (batch size 6k)\nsingle TPU Pod V3-1024 (64 nodes and 1024 TPUs) info from paper and https://huggingface.co/Rostlab/prot_bert_bfd"},{"id":"epoch:e8c089371e8b79eb4d12:parameters","entityId":"epoch:e8c089371e8b79eb4d12","metric":"parameters","value":420000000.0,"unit":"parameters","observationPeriod":"2021-05-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:e8c089371e8b79eb4d12:training-compute","entityId":"epoch:e8c089371e8b79eb4d12","metric":"training-compute","value":7.27e+21,"unit":"FLOP","observationPeriod":"2021-05-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2, two stages with different sequence lengths\nStage 1: 300k steps, sequence length 512, batch size 15360\nStage 2: 100k steps, sequence length 2048, batch size 2560\n\n6ND formula\nStage 1: 6*300000*15360*512*420000000=5.9454259e+21\nStage 2: 6*100000*2560*2048*420000000=1.3212058e+21\nTotal: 7.2666317e+21"},{"id":"epoch:6632ad7998464f9b2076:parameters","entityId":"epoch:6632ad7998464f9b2076","metric":"parameters","value":3000000000.0,"unit":"parameters","observationPeriod":"2021-05-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:6632ad7998464f9b2076:training-compute","entityId":"epoch:6632ad7998464f9b2076","metric":"training-compute","value":1.8704498688e+22,"unit":"FLOP","observationPeriod":"2021-05-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"991K steps, 2048 batch, 512 sequence length (Table 2)\n\nTotal tokens: 991K*2048*512=1039138816000\n\nFLOP: 6*1039138816000*3000000000=18704498688000000000000"},{"id":"epoch:bbdc91309f932ea318e2:parameters","entityId":"epoch:bbdc91309f932ea318e2","metric":"parameters","value":246000000.0,"unit":"parameters","observationPeriod":"2021-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"246M\n\"Transformer-XL is a language model with 246 million parameters.\""},{"id":"epoch:bbdc91309f932ea318e2:training-compute","entityId":"epoch:bbdc91309f932ea318e2","metric":"training-compute","value":3.7848651e+20,"unit":"FLOP","observationPeriod":"2021-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"base model training compute: 3.7832771e+20 FLOP\n "},{"id":"epoch:d2023742a2781a500efa:parameters","entityId":"epoch:d2023742a2781a500efa","metric":"parameters","value":6053381344.0,"unit":"parameters","observationPeriod":"2021-05-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: model details table in GitHub"},{"id":"epoch:d2023742a2781a500efa:training-compute","entityId":"epoch:d2023742a2781a500efa","metric":"training-compute","value":1.5e+22,"unit":"FLOP","observationPeriod":"2021-05-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: zero shot evaluation table in GitHub"},{"id":"epoch:8b6f439d861352bae3b0:parameters","entityId":"epoch:8b6f439d861352bae3b0","metric":"parameters","value":85000000.0,"unit":"parameters","observationPeriod":"2021-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"85M, table 1"},{"id":"epoch:8b6f439d861352bae3b0:training-compute","entityId":"epoch:8b6f439d861352bae3b0","metric":"training-compute","value":2.1e+20,"unit":"FLOP","observationPeriod":"2021-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Overall, training DINO with Vision Transformers\nachieves 76.1 top-1 accuracy using two 8-GPU servers for 3\ndays\"\n\nGPU is V100\n\n16 * 125 teraflops * 3 days * 0.4 utilization\n= 2.1e20\n\nHowever, this isn't the best result in the paper (which is 80.1% with ViT-B/8). 76.1% is the result from ViT-B/16 per Table 2, which may be 5x cheaper than ViT-B/8 based on Table 1?\n\nupd:\n \"Table 8: Time and memory requirements. We show total running\ntime and peak memory per GPU (\u201cmem.\u201d) when running ViT-S/16\nDINO models on two 8-GPU machines.\"\n\n2*8*125 teraflops*72.6h*3600*0.4=2.09088e+20"},{"id":"epoch:62a9505c2474b5a42438:parameters","entityId":"epoch:62a9505c2474b5a42438","metric":"parameters","value":207000000000.0,"unit":"parameters","observationPeriod":"2021-04-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:62a9505c2474b5a42438:training-compute","entityId":"epoch:62a9505c2474b5a42438","metric":"training-compute","value":5.112e+22,"unit":"FLOP","observationPeriod":"2021-04-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 8 digitized with https://automeris.io/wpd/?v=5_2, training the 200B model runs for 42.6B tokens. 6 * 200B * 42.6B = 5.112e22\n\nElsewhere, the akronomicon leaderboard indicates 538 PF-days, or 5.037e22 FLOPs.\nhttps://web.archive.org/web/20220211115721/https://lair.lighton.ai/akronomicon/\n\nNote however, that this seems to contradict Table 4, which appears to indicate that PanGu-\u03b1 200B saw 317.569B tokens. If this were the case, training compute would be ~ 6 * 200B * 317.569B = 3.811e23"},{"id":"epoch:3e10d0e4d59f9d7fb950:parameters","entityId":"epoch:3e10d0e4d59f9d7fb950","metric":"parameters","value":247000000.0,"unit":"parameters","observationPeriod":"2021-04-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"They base their architecture off of \"Transformer (Adaptive Input Embeddings) WT103\", a 16 layer transformer with 16 heads per MHA layer, d_model=1024 and d_ff=4096."},{"id":"epoch:3e10d0e4d59f9d7fb950:training-compute","entityId":"epoch:3e10d0e4d59f9d7fb950","metric":"training-compute","value":3.02e+19,"unit":"FLOP","observationPeriod":"2021-04-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The authors say they train their base model following Baevski & Auli (2019). That paper trained for 286k steps in batches of 65,536 tokens.\nThis suggests training the base model took 6 * 247M * 286k * 65536 = 2.778E19 FLOPs. \n\nHowever, DiffQ introduces some additional computational overhead during training: \"Using DiffQ usually increase the training time by some amount. On the language modeling task, the time per batch went from 115ms to 125ms.\"\n\nThis suggests 125/115 - 1 = 8.7% overhead, relative to the direct FLOPs used to do vanilla forward and backward passes on the model weights.\n\n2.778E19 * 125 / 115 = 3.020e19"},{"id":"epoch:c31eed6cf14a2e05589f:parameters","entityId":"epoch:c31eed6cf14a2e05589f","metric":"parameters","value":27000000000.0,"unit":"parameters","observationPeriod":"2021-04-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c31eed6cf14a2e05589f:training-compute","entityId":"epoch:c31eed6cf14a2e05589f","metric":"training-compute","value":3.5997696e+22,"unit":"FLOP","observationPeriod":"2021-04-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"128 Nvidia A100 for 35 days"},{"id":"epoch:b8c49813cc63c688860b:parameters","entityId":"epoch:b8c49813cc63c688860b","metric":"parameters","value":12000000000000.0,"unit":"parameters","observationPeriod":"2021-04-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"They instantiated a 12T-parameter model to show that their hardware setup can train it despite the huge memory requirements."},{"id":"epoch:cc1f1a2edb527bd66b4c:parameters","entityId":"epoch:cc1f1a2edb527bd66b4c","metric":"parameters","value":1000000000000.0,"unit":"parameters","observationPeriod":"2021-04-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"[NOTE: They didn't train the model fully end-to-end, probably just to obtain enough information to gauge the ability to do model parallelisation]\n\n\"Our approach allows us to perform training iterations on a model with 1 trillion parameters at 502 petaFLOP/s on 3072 GPUs with achieved per-GPU throughput of 52% of theoretical peak.\""},{"id":"epoch:27a000f39674c0c967e6:parameters","entityId":"epoch:27a000f39674c0c967e6","metric":"parameters","value":148000000.0,"unit":"parameters","observationPeriod":"2021-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"148M, Table 6"},{"id":"epoch:27a000f39674c0c967e6:training-compute","entityId":"epoch:27a000f39674c0c967e6","metric":"training-compute","value":1.8877734e+18,"unit":"FLOP","observationPeriod":"2021-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 148000000 parameters * 10240 tokens per batch * 200000 steps = 1.818624e+18 FLOP\n\n11340000000000 FLOP / sec [assumed fp32] * 4 GPUs * 40 hours * 3600 sec/ hour * 0.3 [assumed precision] = 1.959552e+18 FLOP\n\ngeometric mean: \nsqrt(1.818624e+18*1.959552e+18) = 1.8877734e+18 "},{"id":"epoch:4848f958e4fc4f08de0d:parameters","entityId":"epoch:4848f958e4fc4f08de0d","metric":"parameters","value":97600000.0,"unit":"parameters","observationPeriod":"2021-04-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b1940c8ca389006dd27b:parameters","entityId":"epoch:b1940c8ca389006dd27b","metric":"parameters","value":97600000.0,"unit":"parameters","observationPeriod":"2021-04-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1e3a37b017c1b03fbb6c:parameters","entityId":"epoch:1e3a37b017c1b03fbb6c","metric":"parameters","value":450000000.0,"unit":"parameters","observationPeriod":"2021-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"450M\nsource: https://arxiv.org/pdf/2203.12644\ntable 3"},{"id":"epoch:1e3a37b017c1b03fbb6c:training-compute","entityId":"epoch:1e3a37b017c1b03fbb6c","metric":"training-compute","value":2.749544e+19,"unit":"FLOP","observationPeriod":"2021-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 450000000 parameters * 74000 tokens per batch * 286000 steps = 5.71428e+19 FLOP\n\n125000000000000 [assumed bf16] * 98 GPU-hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.323e+19 FLOP\n\ngeometric mean:\nsqrt(5.71428e+19*1.323e+19) = 2.749544e+19"},{"id":"epoch:4dc2183deefc49bdbf9f:parameters","entityId":"epoch:4dc2183deefc49bdbf9f","metric":"parameters","value":668893184.0,"unit":"parameters","observationPeriod":"2021-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Vocabulary size unclear, ~260k if they directly use WT 103\nAssumed tied embeddings\n\nRemaining parameters: \n32 layers\n8 attention heads (H)\n128 head dimensions (N, D)\n1024 model (embedding) dimensions (W, M)\n4096 feedforward dimensions\n\"We partition the training data into non-overlapping\nblocks of 512 contiguous tokens\"\n\nMLP layer: 2*1024*4096=8388608\nAtt layer: \nH\u00d7(W\u00d7(2\u00d7D + N) + N\u00d7M) \n8*(1024*(2*128+128)+128*1024)=4194304\nTotal non embedding: 32*(8388608+4194304)=402653184\nEmbedding: 260000*1024=266240000\nTotal: 402653184+266240000=668893184"},{"id":"epoch:4dc2183deefc49bdbf9f:training-compute","entityId":"epoch:4dc2183deefc49bdbf9f","metric":"training-compute","value":1.3517515512289972e+19,"unit":"FLOP","observationPeriod":"2021-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 4, \"We found that we could\nspeed up training by reducing the warm-up steps,\ntotal update steps, maximum and minimum rates,\nand batch size to 8K steps, 142K steps, 5 \u22c5 10\u22126,\n0.5, and 25K tokens \"\n\n142K updates, 25k token batches\nTraining compute: \n6*668893184*142000*25000=1.4247425e+19\n\n\nGPU hour estimate: \n95*60*60*125000000000000*0.3=1.2825e+19\n\nGeometric mean: 13517515512289972224\n\nEpochs: 142000*25000/103000000=34.4660194175"},{"id":"epoch:a796f1c423a78f9a6ea5:parameters","entityId":"epoch:a796f1c423a78f9a6ea5","metric":"parameters","value":668893184.0,"unit":"parameters","observationPeriod":"2021-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Vocabulary size unclear, ~260k if they directly use WT 103\nAssumed tied embeddings\n\nRemaining parameters: \n32 layers\n8 attention heads (H)\n128 head dimensions (N, D)\n1024 model (embedding) dimensions (W, M)\n4096 feedforward dimensions\n\"We partition the training data into non-overlapping\nblocks of 512 contiguous tokens\"\n\nMLP layer: 2*1024*4096=8388608\nAtt layer: \nH\u00d7(W\u00d7(2\u00d7D + N) + N\u00d7M) \n8*(1024*(2*128+128)+128*1024)=4194304\nTotal non embedding: 32*(8388608+4194304)=402653184\nEmbedding: 260000*1024=266240000\nTotal: 402653184+266240000=668893184"},{"id":"epoch:a796f1c423a78f9a6ea5:training-compute","entityId":"epoch:a796f1c423a78f9a6ea5","metric":"training-compute","value":1.372929105052406e+19,"unit":"FLOP","observationPeriod":"2021-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 4, \"We found that we could\nspeed up training by reducing the warm-up steps,\ntotal update steps, maximum and minimum rates,\nand batch size to 8K steps, 142K steps, 5 \u22c5 10\u22126,\n0.5, and 25K tokens \"\n\n142K updates, 25k token batches\nTraining compute: \n6*668893184*142000*25000=1.4247425e+19\n\n\nGPU hour estimate: \n98*60*60*125000000000000*0.3=1.323e+19\n\nGeometric mean: 13729291050524059648\n\nEpochs: 142000*25000/103000000=34.4660194175"},{"id":"epoch:2d3f400aaaf56e2110a6:parameters","entityId":"epoch:2d3f400aaaf56e2110a6","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2021-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2a52c455267d043a0d99:parameters","entityId":"epoch:2a52c455267d043a0d99","metric":"parameters","value":2700000000.0,"unit":"parameters","observationPeriod":"2021-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://www.eleuther.ai/projects/gpt-neo/\n\nNote: Directory of LLMs (https://docs.google.com/spreadsheets/d/1gc6yse74XCwBx028HV_cvdxwXkmXejVjkO-Mz2uwE0k/edit#gid=0) gives a somewhat lower estimate (2e9)"},{"id":"epoch:2a52c455267d043a0d99:training-compute","entityId":"epoch:2a52c455267d043a0d99","metric":"training-compute","value":7.9e+21,"unit":"FLOP","observationPeriod":"2021-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://www.aitracker.org/\n\n6 FLOP / token / parameter * 2.7 * 10^9 parameters * 420000000000 tokens [see dataset size notes] = 6.804e+21 FLOP"},{"id":"epoch:f3dc3c61ffa026cf6ffa:parameters","entityId":"epoch:f3dc3c61ffa026cf6ffa","metric":"parameters","value":2700000000.0,"unit":"parameters","observationPeriod":"2021-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:bed48b025311241e5713:parameters","entityId":"epoch:bed48b025311241e5713","metric":"parameters","value":2700000000.0,"unit":"parameters","observationPeriod":"2021-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:46a96ce681cc2da45790:parameters","entityId":"epoch:46a96ce681cc2da45790","metric":"parameters","value":125000000.0,"unit":"parameters","observationPeriod":"2021-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:658f7eab29e56ebb6c6a:parameters","entityId":"epoch:658f7eab29e56ebb6c6a","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2021-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5329d4bc65c7bdd308d8:parameters","entityId":"epoch:5329d4bc65c7bdd308d8","metric":"parameters","value":125000000.0,"unit":"parameters","observationPeriod":"2021-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:81ff4716787db4b7a2f1:parameters","entityId":"epoch:81ff4716787db4b7a2f1","metric":"parameters","value":1300000000.0,"unit":"parameters","observationPeriod":"2021-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:95e66bc08aa56a4c22a4:parameters","entityId":"epoch:95e66bc08aa56a4c22a4","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2021-03-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:95e66bc08aa56a4c22a4:training-compute","entityId":"epoch:95e66bc08aa56a4c22a4","metric":"training-compute","value":4.9397553e+21,"unit":"FLOP","observationPeriod":"2021-03-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 10*10^9 parameters * 1024 samples per batch * 512 tokens per sample * 200000 steps [reported for smaller models] = 6.291456e+21 FLOP\n\n31330000000000 [assumed fp16 precision] * 64 GPUs * 1791 hours [see training time notes] * 3600 sec / hour * 0.3 [assumed utilization] = 3.8784635e+21 FLOP\n\nsqrt(6.291456e+21*3.8784635e+21) = 4.9397553e+21 \n__________ \nestimation from the Algorithmic progress paper (I am not sure where the dataset size estimation comes from):\n\n6 FLOP / token / parameter * (570649396491.227-334000000000+395050630177) tokens * 10*10^9 parameter = 3.7902002e+22 FLOP "},{"id":"epoch:13c40d0997fcc2ab4ffb:parameters","entityId":"epoch:13c40d0997fcc2ab4ffb","metric":"parameters","value":2000000000.0,"unit":"parameters","observationPeriod":"2021-03-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e58acb5b84161ccbd298:parameters","entityId":"epoch:e58acb5b84161ccbd298","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2021-03-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1855b46a88aad5a8a743:parameters","entityId":"epoch:1855b46a88aad5a8a743","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2021-03-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:15899072b33d0453dea9:parameters","entityId":"epoch:15899072b33d0453dea9","metric":"parameters","value":125000000.0,"unit":"parameters","observationPeriod":"2021-03-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"125M"},{"id":"epoch:15899072b33d0453dea9:training-compute","entityId":"epoch:15899072b33d0453dea9","metric":"training-compute","value":1.8144e+21,"unit":"FLOP","observationPeriod":"2021-03-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"125000000000000 FLOP / GPU / sec [V100 reported, bf16 assumed] * 420 hours [see training time notes] * 32 GPUs * 3600 sec / hour * 0.3 [assumed utilization] = 1.8144e+21 FLOP"},{"id":"epoch:ae11a9712d69e2372e03:parameters","entityId":"epoch:ae11a9712d69e2372e03","metric":"parameters","value":192000000.0,"unit":"parameters","observationPeriod":"2021-03-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 7 appendix B"},{"id":"epoch:ae11a9712d69e2372e03:training-compute","entityId":"epoch:ae11a9712d69e2372e03","metric":"training-compute","value":1.763328e+22,"unit":"FLOP","observationPeriod":"2021-03-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(350) * (128000000000) * (1312 * 10**5) * 3 = 17633280000000000000000\n(epochs) * (inference FLOP) * (dataset size) * (constant to account for backpropagation)\nfrom 4.2 \"Our training method closely matches that of EfficientNet, where we train for 350 epochs, but with a few small differences\"350 epochs from description of Table 8 in appendix C"},{"id":"epoch:0c650a93c696d51c87b0:parameters","entityId":"epoch:0c650a93c696d51c87b0","metric":"parameters","value":136000000.0,"unit":"parameters","observationPeriod":"2021-03-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12 encoder layers, 12 attention heads, 768 hidden size, and 512 maximum input sequence length for a total of 136M\nparameters. "},{"id":"epoch:0c650a93c696d51c87b0:training-compute","entityId":"epoch:0c650a93c696d51c87b0","metric":"training-compute","value":2.5587079e+20,"unit":"FLOP","observationPeriod":"2021-03-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 262144000000 tokens * 136000000 parameters = 2.139095e+20 FLOP\n\nA TPUv3-8 has 8 cores. TPUv3 has 2 cores per chip. So 4 chips.\n\n123000000000000 FLOP / chip / sec * 4 chips * 576 hours [see training time notes] * 3600 sec / hour * 0.3 [assumed utilization] = 3.0606336e+20 FLOP\n\nsqrt(2.139095e+20*3.0606336e+20) = 2.5587079e+20 FLOP"},{"id":"epoch:addbb32780d207d965f9:parameters","entityId":"epoch:addbb32780d207d965f9","metric":"parameters","value":1002700000000.0,"unit":"parameters","observationPeriod":"2021-03-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 5. Note model is sparse MoE with 960 experts; not all parameters are activated on the forward pass."},{"id":"epoch:addbb32780d207d965f9:training-compute","entityId":"epoch:addbb32780d207d965f9","metric":"training-compute","value":5.5e+21,"unit":"FLOP","observationPeriod":"2021-03-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Estimate taken from https://www.governance.ai/research-paper/recent-trends-chinas-llm-landscape"},{"id":"epoch:8d734df6039bb55e33ff:parameters","entityId":"epoch:8d734df6039bb55e33ff","metric":"parameters","value":9431810048.0,"unit":"parameters","observationPeriod":"2021-03-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The largest model is a transformer with 9.4B parameters (Table 2)"},{"id":"epoch:8d734df6039bb55e33ff:training-compute","entityId":"epoch:8d734df6039bb55e33ff","metric":"training-compute","value":1.449e+22,"unit":"FLOP","observationPeriod":"2021-03-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Both our 2.7B and 9.4B parameter models were trained with batches of approximately 500k label BPE tokens per batch [...] The 9.4B parameter model was trained [...] for a total of 200k SGD steps.\"\n\nAlso note that the full dataset contains 56.8B label BPE tokens and 88.8B context tokens, so for each batch of 500k label tokens, there are likely 500k * 88.8B / 56.8B = 780k context tokens.\n\n6 * 9.4318B * 200k * (500k + 780k) = 1.449e22"},{"id":"epoch:282a1f7130ec84ec76a7:parameters","entityId":"epoch:282a1f7130ec84ec76a7","metric":"parameters","value":736000000.0,"unit":"parameters","observationPeriod":"2021-03-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"parameters (Table 3):  738e6"},{"id":"epoch:282a1f7130ec84ec76a7:training-compute","entityId":"epoch:282a1f7130ec84ec76a7","metric":"training-compute","value":4.428e+21,"unit":"FLOP","observationPeriod":"2021-03-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 738 * 10^6 parameters * 1000 * 10^9 tokens = 4.428e+21 FLOP"},{"id":"epoch:29c61267d53c46c9999f:parameters","entityId":"epoch:29c61267d53c46c9999f","metric":"parameters","value":60000000.0,"unit":"parameters","observationPeriod":"2021-03-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The final architecture of the network comprised 45 layers with over 60 million trainable parameters\""},{"id":"epoch:29c61267d53c46c9999f:training-compute","entityId":"epoch:29c61267d53c46c9999f","metric":"training-compute","value":4.3e+18,"unit":"FLOP","observationPeriod":"2021-03-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.5 million steps, batch size 64, 210 hours on NVIDIA Tesla P100 system"},{"id":"epoch:bcd714428b729d36ecaf:parameters","entityId":"epoch:bcd714428b729d36ecaf","metric":"parameters","value":242000000.0,"unit":"parameters","observationPeriod":"2021-03-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"242m"},{"id":"epoch:bcd714428b729d36ecaf:training-compute","entityId":"epoch:bcd714428b729d36ecaf","metric":"training-compute","value":7.14e+18,"unit":"FLOP","observationPeriod":"2021-03-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 242000000 parameters * 512 tokens per sample * 64 samples per batch * 150000 steps = 7.1368704e+18 FLOP"},{"id":"epoch:6076832c40500d5cfb96:parameters","entityId":"epoch:6076832c40500d5cfb96","metric":"parameters","value":11300000000.0,"unit":"parameters","observationPeriod":"2021-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Wu Dao \u2014 Wen Hui has reached 11.3 billion parameters, and through simple fine-tuning can generate poetry, make videos, draw pictures, retrieve text, perform complex reasoning, etc.\"\n\nhttps://medium.com/syncedreview/chinas-gpt-3-baai-introduces-superscale-intelligence-model-wu-dao-1-0-98a573fc4d70"},{"id":"epoch:6076832c40500d5cfb96:training-compute","entityId":"epoch:6076832c40500d5cfb96","metric":"training-compute","value":1.161216e+20,"unit":"FLOP","observationPeriod":"2021-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"64 Nvidia V100 GPUs for 2.5 days\n\n64 GPUs * 2.8e13 FLOP/s /GPU * 2.5*24*60*60s* 0.3 [utilization rate]\n\n"},{"id":"epoch:7cb4cf8154d7dfd794af:parameters","entityId":"epoch:7cb4cf8154d7dfd794af","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2021-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Currently, the model has 1 billion parameters and is trained on 50 million graphic pairs collected from open sources.\"\n\nhttps://medium.com/syncedreview/chinas-gpt-3-baai-introduces-superscale-intelligence-model-wu-dao-1-0-98a573fc4d70"},{"id":"epoch:7cb4cf8154d7dfd794af:training-compute","entityId":"epoch:7cb4cf8154d7dfd794af","metric":"training-compute","value":7.1995392e+21,"unit":"FLOP","observationPeriod":"2021-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"128 Nvidia A100 GPUs for 7 days\n\n128 GPUs * 3.1e14 FLOP/s /GPU * 7*24*60*60s* 0.3 [utilization rate]\n\n"},{"id":"epoch:c7222f7e86beb544ab56:parameters","entityId":"epoch:c7222f7e86beb544ab56","metric":"parameters","value":480000000.0,"unit":"parameters","observationPeriod":"2021-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 4\n 480M"},{"id":"epoch:c7222f7e86beb544ab56:training-compute","entityId":"epoch:c7222f7e86beb544ab56","metric":"training-compute","value":4.79e+22,"unit":"FLOP","observationPeriod":"2021-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From communication with author:\n\n22671 TPU days on specific hardware.\n\nWhich hardware did you use and in which configuration?\n2048 cores of TPU v3.\n\nPrecision: Mixed. bfloat16 for activations, float32 for weights and optimizer slots.\n\n2048 TPUv3 cores means 1024 TPUv3 chips, and the spec is 123e12 FLOP/second per chip with bfloat16 precision (Source: https://cloud.google.com/tpu/docs/system-architecture-tpu-vm)\n\nSo the compute estimate is:\n1024 chips * 123e12 FLOP/second * 0.4 utilization * 11 days * 24 * 60 * 60 = 4.788191232e+22 FLOP"},{"id":"epoch:12655724cabfa32e998d:parameters","entityId":"epoch:12655724cabfa32e998d","metric":"parameters","value":10000000000.0,"unit":"parameters","observationPeriod":"2021-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We scale the\nmodel size up to 10 billion and 100 billion parameters, and build\nthe largest pretrained model in Chinese.\""},{"id":"epoch:694ecf1e0c3440162b24:parameters","entityId":"epoch:694ecf1e0c3440162b24","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2021-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We scale the\nmodel size up to 10 billion and 100 billion parameters, and build\nthe largest pretrained model in Chinese.\""},{"id":"epoch:98a667db34e87bc27a95:parameters","entityId":"epoch:98a667db34e87bc27a95","metric":"parameters","value":234000000.0,"unit":"parameters","observationPeriod":"2021-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 5"},{"id":"epoch:98a667db34e87bc27a95:training-compute","entityId":"epoch:98a667db34e87bc27a95","metric":"training-compute","value":2.1173704e+19,"unit":"FLOP","observationPeriod":"2021-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 234000000 parameters * 1024 tokens per sample * 8*8 samples per batch * 400000 steps = 3.6805018e+19 FLOP\n\n31330000000000 FLOP / sec * 360 GPU-hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.2181104e+19 FLOP\n\nsqrt(3.6805018e+19*1.2181104e+19) = 2.1173704e+19 \n\n"},{"id":"epoch:87c6f0ecdd760c42435f:parameters","entityId":"epoch:87c6f0ecdd760c42435f","metric":"parameters","value":148000000.0,"unit":"parameters","observationPeriod":"2021-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:87c6f0ecdd760c42435f:training-compute","entityId":"epoch:87c6f0ecdd760c42435f","metric":"training-compute","value":2.592e+19,"unit":"FLOP","observationPeriod":"2021-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8 V100 GPU days\n\n8*24*60*60*125000000000000*0.3=2.592e+19\n"},{"id":"epoch:c483510cb078e08fdb06:parameters","entityId":"epoch:c483510cb078e08fdb06","metric":"parameters","value":225000000.0,"unit":"parameters","observationPeriod":"2021-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c483510cb078e08fdb06:training-compute","entityId":"epoch:c483510cb078e08fdb06","metric":"training-compute","value":3.564e+19,"unit":"FLOP","observationPeriod":"2021-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"11 V100 GPU days\n\n11*24*60*60*125000000000000*0.3=3.564e+19"},{"id":"epoch:3a684fc6d4f41c9736bc:parameters","entityId":"epoch:3a684fc6d4f41c9736bc","metric":"parameters","value":90000000.0,"unit":"parameters","observationPeriod":"2021-02-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"90M (medium)"},{"id":"epoch:3a684fc6d4f41c9736bc:training-compute","entityId":"epoch:3a684fc6d4f41c9736bc","metric":"training-compute","value":3.89e+18,"unit":"FLOP","observationPeriod":"2021-02-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 90000000 parameters * 103000000 tokens * 70 epochs = 3.8934e+18 FLOP"},{"id":"epoch:c0c1c5a1e05738039866:parameters","entityId":"epoch:c0c1c5a1e05738039866","metric":"parameters","value":40000000.0,"unit":"parameters","observationPeriod":"2021-02-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c0c1c5a1e05738039866:training-compute","entityId":"epoch:c0c1c5a1e05738039866","metric":"training-compute","value":1.7304e+18,"unit":"FLOP","observationPeriod":"2021-02-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For the small and medium configurations, we use batch sizes\nof 96 and 56 sequences, respectively, and train for about\n120 and 70 epochs.\"\n\n\nTraining compute: 6*40000000*70*103000000=1.7304e+18"},{"id":"epoch:cdd7dd3d535cb26f1e1b:parameters","entityId":"epoch:cdd7dd3d535cb26f1e1b","metric":"parameters","value":1683456.0,"unit":"parameters","observationPeriod":"2021-02-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See figure 7"},{"id":"epoch:244820cc6e17de13c4a4:parameters","entityId":"epoch:244820cc6e17de13c4a4","metric":"parameters","value":100000000.0,"unit":"parameters","observationPeriod":"2021-02-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train an MSA Transformer model with 100M parameters...\" "},{"id":"epoch:244820cc6e17de13c4a4:training-compute","entityId":"epoch:244820cc6e17de13c4a4","metric":"training-compute","value":5.49e+21,"unit":"FLOP","observationPeriod":"2021-02-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Based on: https://docs.google.com/spreadsheets/d/1enan21dFx03TkwufHgOwTVNBtuYlqNY9uurjIK6YS-8/edit#gid=0\n\nNumber of steps 4.5e5, batch size (tokens) 6.1e7, parameters 1e8\n\nCalculation = 4e8 FLOP/bp * 4.5e5 bp + 2e8 FLOP/fp * 2.75e13 fp\n\nBatch size: 512\nSeq length: 100 * 1192 tokens\nAll models are trained on 32 V100 GPUs for 100k updates. The four models with best contact precision are then further trained to 150k updates. Finally, the best model at 150k updates is trained to 450k updates.\n\n450k * 512 * 100 * 1192 * 100M * 6 = 1.65e22"},{"id":"epoch:b6fdaf376c590faf045f:parameters","entityId":"epoch:b6fdaf376c590faf045f","metric":"parameters","value":2676376.0,"unit":"parameters","observationPeriod":"2021-02-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:b6fdaf376c590faf045f:training-compute","entityId":"epoch:b6fdaf376c590faf045f","metric":"training-compute","value":5.6845152e+18,"unit":"FLOP","observationPeriod":"2021-02-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"31330000000000 FLOP / GPU/ sec [Tesla V100] * 1 GPU * 168 hours * 3600 sec / hour * 0.3 [assumed utilization] = 5.6845152e+18 FLOP"},{"id":"epoch:81ee2f02421de8c3b1a7:parameters","entityId":"epoch:81ee2f02421de8c3b1a7","metric":"parameters","value":25200000.0,"unit":"parameters","observationPeriod":"2021-01-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"25.2M (Table 2)"},{"id":"epoch:81ee2f02421de8c3b1a7:training-compute","entityId":"epoch:81ee2f02421de8c3b1a7","metric":"training-compute","value":1.178e+16,"unit":"FLOP","observationPeriod":"2021-01-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 10:\n3.1e16 FLOP * 0.38 = 1.178e+16 FLOP\n\n6 FLOP/token/parameter * 25200000 parameters * 912344 tokens * 100 epochs = 1.3794641e+16 FLOP"},{"id":"epoch:d982d8570226766ab226:parameters","entityId":"epoch:d982d8570226766ab226","metric":"parameters","value":25200000.0,"unit":"parameters","observationPeriod":"2021-01-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:7a8f437a7b8ae9c126dc:parameters","entityId":"epoch:7a8f437a7b8ae9c126dc","metric":"parameters","value":7600000.0,"unit":"parameters","observationPeriod":"2021-01-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:83d92804a1526ab121fd:parameters","entityId":"epoch:83d92804a1526ab121fd","metric":"parameters","value":15600000.0,"unit":"parameters","observationPeriod":"2021-01-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:10367c26598160f4dd8d:parameters","entityId":"epoch:10367c26598160f4dd8d","metric":"parameters","value":86000000.0,"unit":"parameters","observationPeriod":"2021-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(DeiT-B)"},{"id":"epoch:10367c26598160f4dd8d:training-compute","entityId":"epoch:10367c26598160f4dd8d","metric":"training-compute","value":7.884e+19,"unit":"FLOP","observationPeriod":"2021-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2*86000000 parameters*3*1280000 training examples*300 epochs=1.98144e+17 FLOPs\n\ncompute [FLOP] = training time [s] \u00d7 # of GPUs/TPUs \u00d7 peak FLOP/s \u00d7 utilization rate\n\n(53h+20h)*3600*8*125000000000000 peak FLOP/s*0.3=7.884e+19\n\n"},{"id":"epoch:2ec4ad64a099389e608e:parameters","entityId":"epoch:2ec4ad64a099389e608e","metric":"parameters","value":1571000000000.0,"unit":"parameters","observationPeriod":"2021-01-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Combining expert, model and data parallelism, we design two large Switch Transformer models, one with 395 billion and 1.6 trillion parameters\"\nTable 9 gives more precise count of 1571B parameters"},{"id":"epoch:2ec4ad64a099389e608e:training-compute","entityId":"epoch:2ec4ad64a099389e608e","metric":"training-compute","value":8.22e+22,"unit":"FLOP","observationPeriod":"2021-01-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 4\nhttps://arxiv.org/ftp/arxiv/papers/2104/2104.10350.pdf"},{"id":"epoch:24f2af8a2d3cea1c1e42:parameters","entityId":"epoch:24f2af8a2d3cea1c1e42","metric":"parameters","value":2600000000.0,"unit":"parameters","observationPeriod":"2021-01-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"It has 2.6 billion parameters and is capable of performing cognitive activities such as memorization, comprehension, retrieval, numerical calculation, multi-language, etc.\"\n\nhttps://medium.com/syncedreview/chinas-gpt-3-baai-introduces-superscale-intelligence-model-wu-dao-1-0-98a573fc4d70"},{"id":"epoch:24f2af8a2d3cea1c1e42:training-compute","entityId":"epoch:24f2af8a2d3cea1c1e42","metric":"training-compute","value":6.5028096e+20,"unit":"FLOP","observationPeriod":"2021-01-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"64 Nvidia V100 GPUs for two weeks\n\n64 GPUs * 2.8e13 FLOP/s /GPU * 14*24*60*60s * 0.3 [utilization rate]\n\n"},{"id":"epoch:902fa5737c5b275519da:parameters","entityId":"epoch:902fa5737c5b275519da","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2021-01-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"... we study the utility of large models, with the parameter count ranging from 600M to 8B...\""},{"id":"epoch:7db1c6603c4bf627ffba:training-compute","entityId":"epoch:7db1c6603c4bf627ffba","metric":"training-compute","value":5.94e+19,"unit":"FLOP","observationPeriod":"2021-01-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"125000000000000 FLOP / GPU / sec * 8 GPUs * 55 hours * 3600 sec / hour * 0.3 [assumed utilization] = 5.94e+19 FLOP\n\n[Table 6]"},{"id":"epoch:a318e029589a76ccd332:training-compute","entityId":"epoch:a318e029589a76ccd332","metric":"training-compute","value":5.184e+20,"unit":"FLOP","observationPeriod":"2021-01-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"125000000000000 FLOP / GPU / sec * 24 GPUs * 160 hours * 3600 sec / hour * 0.3 [assumed utilization] = 5.184e+20 FLOP\n\n[Table 6]"},{"id":"epoch:9cca06037f6c9ad0df8d:training-compute","entityId":"epoch:9cca06037f6c9ad0df8d","metric":"training-compute","value":3.0456e+20,"unit":"FLOP","observationPeriod":"2021-01-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"125000000000000 FLOP / GPU / sec * 24 GPUs * 94 hours * 3600 sec / hour * 0.3 [assumed utilization] = 3.0456e+20 FLOP\n\n[Table 6]"},{"id":"epoch:1ce6889dabab09cbbf01:parameters","entityId":"epoch:1ce6889dabab09cbbf01","metric":"parameters","value":12000000000.0,"unit":"parameters","observationPeriod":"2021-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"DALL\u00b7E is a 12-billion parameter version of GPT-3 trained to generate images from text descriptions"},{"id":"epoch:1ce6889dabab09cbbf01:training-compute","entityId":"epoch:1ce6889dabab09cbbf01","metric":"training-compute","value":4.7e+22,"unit":"FLOP","observationPeriod":"2021-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://lair.lighton.ai/akronomicon/\n\narchived: https://github.com/lightonai/akronomicon/tree/main/akrodb\n\nVAE training \"on 64 16 GB NVIDIA V100 GPUs, with a per-GPU batch size of 8, resulting in a total batch size of 512. It is trained for a total of 3,000,000 updates.\"\n\nTransformer training: \"We trained the model using 1024, 16 GB NVIDIA V100 GPUs and a total batch size of 1024, for a total of 430,000 updates.\"; \"We concatenate up to 256 BPE-encoded text tokens with the 32 \u00d7 32 = 1024 image tokens\"\nTotal tokens: 430000 steps * 1024 batch size * 1280 sequence length = 563609600000\n\nTransformer FLOP: 6 * 12B parameter * 563609600000 tokens = 4.057989e+22\n\nEstimating the VAE at ~15% seems reasonable"},{"id":"epoch:36e3217471ef0dbaca1e:parameters","entityId":"epoch:36e3217471ef0dbaca1e","metric":"parameters","value":370000000.0,"unit":"parameters","observationPeriod":"2021-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Image encoder\nVision Transformer\nTable 1 in https://arxiv.org/pdf/2010.11929.pdf\nAuthors fine-tuned ViT L/14 at additional 336px resolution, hence the @336 (See ViT)\n307M params\n\nText encoder\n~Transformer (from paper)\n63M params"},{"id":"epoch:36e3217471ef0dbaca1e:training-compute","entityId":"epoch:36e3217471ef0dbaca1e","metric":"training-compute","value":1.05e+22,"unit":"FLOP","observationPeriod":"2021-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://docs.google.com/document/d/156miAJkFN9DDX06C3s03UDsretCtymCKiGDddLBCgQE/edit?usp=sharing\n\n"},{"id":"epoch:7d6bdb1ac05d3828b865:parameters","entityId":"epoch:7d6bdb1ac05d3828b865","metric":"parameters","value":88600000.0,"unit":"parameters","observationPeriod":"2021-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Image encoder\n~ResNet-50 (from paper)\n25.6M params\n\nText encoder\n~Transformer (from paper)\n63M params"},{"id":"epoch:e2e9bda85e5602c914b5:parameters","entityId":"epoch:e2e9bda85e5602c914b5","metric":"parameters","value":35000000.0,"unit":"parameters","observationPeriod":"2021-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:107acd094686bdce01ff:parameters","entityId":"epoch:107acd094686bdce01ff","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2021-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use the same model size as specified by Dai et al. (2019).\""},{"id":"epoch:107acd094686bdce01ff:training-compute","entityId":"epoch:107acd094686bdce01ff","metric":"training-compute","value":5.7829941034035304e+16,"unit":"FLOP","observationPeriod":"2021-01-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train every configuration from scratch for 160,000 steps, using\na batch size of 16 and a segment length of 70\"\n\nTotal tokens: 160000*16*70=179200000\n\nEpochs: 179200000/929000=192\n\nTraining flop: 6*24000000*179200000=2.58048e+16\n\nGPU hour estimate: 40*60*4*45000000000000*0.3=1.296e+17\n\nGeometric mean: 57829941034035304"},{"id":"epoch:ac97cf55027d547127d4:parameters","entityId":"epoch:ac97cf55027d547127d4","metric":"parameters","value":122000000.0,"unit":"parameters","observationPeriod":"2021-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"122M (Table 5)\n"},{"id":"epoch:ac97cf55027d547127d4:training-compute","entityId":"epoch:ac97cf55027d547127d4","metric":"training-compute","value":5.1450348e+18,"unit":"FLOP","observationPeriod":"2021-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/token/parameter * 122000000 parameters * 8 GPUs * 3072 tokens per batch per GPU * 286000 steps = 5.1450348e+18 FLOP"},{"id":"epoch:e51e84bb9ae76ea361ed:parameters","entityId":"epoch:e51e84bb9ae76ea361ed","metric":"parameters","value":83000000.0,"unit":"parameters","observationPeriod":"2021-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e51e84bb9ae76ea361ed:training-compute","entityId":"epoch:e51e84bb9ae76ea361ed","metric":"training-compute","value":3.5e+18,"unit":"FLOP","observationPeriod":"2021-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Warming up the learn-\ning rate to 1.0 for 16K iterations, and then an-\nnealing for 270K iterations using a cosine anneal-\ning schedule\"\n\"When training our lan-\nguage models, we use 8 GPUs with 3072 tokens\nper GPU and an update frequency of 3\"\n\n286k iterations \nEffective batch size: 8*3072=24576\nTotal training tokens: 286000*24576=7028736000\n\nTraining FLOP:\n6*7028736000*83000000=3.5003105e+18"},{"id":"epoch:4cbdea9cf5e765c9c577:parameters","entityId":"epoch:4cbdea9cf5e765c9c577","metric":"parameters","value":96000000.0,"unit":"parameters","observationPeriod":"2021-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:4cbdea9cf5e765c9c577:training-compute","entityId":"epoch:4cbdea9cf5e765c9c577","metric":"training-compute","value":4.05e+18,"unit":"FLOP","observationPeriod":"2021-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Warming up the learn-\ning rate to 1.0 for 16K iterations, and then an-\nnealing for 270K iterations using a cosine anneal-\ning schedule\"\n\"When training our lan-\nguage models, we use 8 GPUs with 3072 tokens\nper GPU and an update frequency of 3\"\n\n286k iterations \nEffective batch size: 8*3072=24576\nTotal training tokens: 286000*24576=7028736000\n\nTraining FLOP:\n6*7028736000*96000000=4.0485519e+18"},{"id":"epoch:7c3314c136fbb58b5acb:parameters","entityId":"epoch:7c3314c136fbb58b5acb","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2020-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://lair.lighton.ai/akronomicon/\n\narchived: https://github.com/lightonai/akronomicon/tree/main/akrodb"},{"id":"epoch:7c3314c136fbb58b5acb:training-compute","entityId":"epoch:7c3314c136fbb58b5acb","metric":"training-compute","value":2e+21,"unit":"FLOP","observationPeriod":"2020-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://github.com/lightonai/akronomicon/blob/10adaca9c74afa7d11f196947e410d248f25abe9/akrodb/American%20University%20of%20Beirut/AraGPT2-Mega.json\n\nAkronomicon uses units of petaflop/s-days. 20 petaflop/s-days ~= 2e21 FLOP.\n\nOur own validation of this estimate is below.\n\nFor the Mega model: 9 days on a TPUv3-128, bfloat16 precision  (from author communication)\n\nA TPUv3-128 has 128 cores (you can infer this from footnote 9 on p.4 of the paper - 128 * 16GB = 2TB). TPUv3 has 2 cores per chip. So 64 chips.\n\nTPUv3 FLOP/s: 1.23E+14\n\nUtilization: use default value of 30% for Language domain (https://epoch.ai/blog/estimating-training-compute)\n\n64 chips * 30% * 1.23E+14 FLOP/s * 9 days * 24h/day * 3600s/h\n~= 2e21 FLOP\n\nnum of examples (seq len = 1024): 9.7M\nbatch size: 256\nnumber of steps: 800k\n\n6 FLOP / token / parameter * 1.46 * 10^9 parameters * 800000 steps * 256 sequences per step * 1024 tokens per step = 1.8371052e+21 FLOP\n"},{"id":"epoch:17c6bb142c34161dfb60:parameters","entityId":"epoch:17c6bb142c34161dfb60","metric":"parameters","value":247000000.0,"unit":"parameters","observationPeriod":"2020-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"247M (Table 5)"},{"id":"epoch:17c6bb142c34161dfb60:training-compute","entityId":"epoch:17c6bb142c34161dfb60","metric":"training-compute","value":3.04e+18,"unit":"FLOP","observationPeriod":"2020-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 247000000 parameters * 103000000 tokens * 205 epochs = 3.129243e+19 FLOP \n\n________\nearlier here was a mistake (1 OOM) from when the calculation for the Algorithmic progress paper was done\nprevious estimation: 3.04e+18 FLOP"},{"id":"epoch:a93e30a618c403e9261f:parameters","entityId":"epoch:a93e30a618c403e9261f","metric":"parameters","value":247000000.00000003,"unit":"parameters","observationPeriod":"2020-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a93e30a618c403e9261f:training-compute","entityId":"epoch:a93e30a618c403e9261f","metric":"training-compute","value":3.0302798e+19,"unit":"FLOP","observationPeriod":"2020-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 247000000 parameters * 416000 steps * 128 sequences per batch * 384 tokens per sequence = 3.0302798e+19 FLOP"},{"id":"epoch:58d3423d61d1e57cc461:parameters","entityId":"epoch:58d3423d61d1e57cc461","metric":"parameters","value":151000000.0,"unit":"parameters","observationPeriod":"2020-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:58d3423d61d1e57cc461:training-compute","entityId":"epoch:58d3423d61d1e57cc461","metric":"training-compute","value":1.7395e+18,"unit":"FLOP","observationPeriod":"2020-12-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"200k steps with batch size 64 and sequence length 150\nTotal tokens: 200000*64*150=1920000000\n\nEpochs: 1920000000/103000000=18.4\n\nTraining compute:\n6*151000000*1920000000=1.73952e+18"},{"id":"epoch:0190fbb316332f5486ca:parameters","entityId":"epoch:0190fbb316332f5486ca","metric":"parameters","value":45000000.0,"unit":"parameters","observationPeriod":"2020-12-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"45M\nTable 2"},{"id":"epoch:0190fbb316332f5486ca:training-compute","entityId":"epoch:0190fbb316332f5486ca","metric":"training-compute","value":5.4e+17,"unit":"FLOP","observationPeriod":"2020-12-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 45000000 parameters * 2000000 tokens * 1000 epochs = 5.4e+17 FLOP"},{"id":"epoch:fdd4b0ca2bcb129bc682:parameters","entityId":"epoch:fdd4b0ca2bcb129bc682","metric":"parameters","value":45000000.0,"unit":"parameters","observationPeriod":"2020-12-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:fdd4b0ca2bcb129bc682:training-compute","entityId":"epoch:fdd4b0ca2bcb129bc682","metric":"training-compute","value":5.4e+17,"unit":"FLOP","observationPeriod":"2020-12-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 45000000 parameters * 2000000 tokens * 1000 epochs = 5.4e+17 FLOP"},{"id":"epoch:2e2340febba6e90bf7fd:parameters","entityId":"epoch:2e2340febba6e90bf7fd","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2020-12-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c33e8a9c0d3ad3a806a9:parameters","entityId":"epoch:c33e8a9c0d3ad3a806a9","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2020-12-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5e9c2b40a5058cf5e415:training-compute","entityId":"epoch:5e9c2b40a5058cf5e415","metric":"training-compute","value":2.09952e+18,"unit":"FLOP","observationPeriod":"2020-12-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" flops = (8) * (1215 * 10**10) * (20 * 3600) * 3 // 10 = 2099520000000000000\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)\n\nmodel of GPU from appendix B (Titan Xp)\nnumber of GPUs from table in appendix A\nflops from https://www.techpowerup.com/gpu-specs/titan-xp.c2948"},{"id":"epoch:4f8bdd4451902411af4d:parameters","entityId":"epoch:4f8bdd4451902411af4d","metric":"parameters","value":652400000.0,"unit":"parameters","observationPeriod":"2020-12-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 9"},{"id":"epoch:4f8bdd4451902411af4d:training-compute","entityId":"epoch:4f8bdd4451902411af4d","metric":"training-compute","value":5.1e+21,"unit":"FLOP","observationPeriod":"2020-12-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Information: \n128 NVIDIA V100 GPUs [Pre-training details]\n8.5 hours on 64 GPUs per epoch, 56 epochs of UR50/S [Appendix B, ESM-1b Hyperparameter optimization, Experimental set-up]\n128 NVIDIA V100 GPU, assuming  V100 PCIe half precision 130 TFLOPS and 0.3 utilization rate\n\nEstimate: (8.5*56*3600) s * 1.3e14 FLOP/s * 0.3 *64 = 4.277e21 FLOP\n\n6NC method:\nUR50/S has 27.1M sequences, which are capped at 1024 amino acids. \n27.1M * 1024 * 56 * 652.4M * 6 = 6.08e21 FLOP\n\nGeometric mean: 5.1e21"},{"id":"epoch:0a81cd87652b74a35ecd:training-compute","entityId":"epoch:0a81cd87652b74a35ecd","metric":"training-compute","value":1.2276e+19,"unit":"FLOP","observationPeriod":"2020-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware: 4x NVIDIA Tesla P100 SXM2 GPUs (18.7 TFLOP/s per GPU in FP16)\nThe architecture is based on Bert-base, with 110M parameters\n\n6*9300000000*2*110000000=1.2276e+19"},{"id":"epoch:c23bf006ec9381a26fb4:parameters","entityId":"epoch:c23bf006ec9381a26fb4","metric":"parameters","value":2600000000.0,"unit":"parameters","observationPeriod":"2020-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"To the best of our knowledge, CPM, with 2.6 billion parameters and 100GB Chinese training data, is the largest Chinese pre-trained language mode\""},{"id":"epoch:c23bf006ec9381a26fb4:training-compute","entityId":"epoch:c23bf006ec9381a26fb4","metric":"training-compute","value":2.6052e+20,"unit":"FLOP","observationPeriod":"2020-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://lair.lighton.ai/akronomicon/\n\narchived: https://github.com/lightonai/akronomicon/tree/main/akrodb\n\n6*2600000000 parameter *16700000000 tokens=2.605200e+20"},{"id":"epoch:fd143bbc8e7a04f74ea5:training-compute","entityId":"epoch:fd143bbc8e7a04f74ea5","metric":"training-compute","value":4.9999999999999993e+20,"unit":"FLOP","observationPeriod":"2020-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: 8x NVIDIA Tesla V100 GPUs (130 TFLOP/s each)\n\n2. Training duration: 2 weeks (1.2e+6 seconds) - directly provided\n\n3. Utilization rate: 40%\n\n4. Final calculation:\n   8 GPUs \u00d7 1.30e+14 FLOP/s \u00d7 1.2e+6 seconds \u00d7 0.4 utilization = 5.0e+20 FLOPs"},{"id":"epoch:05e1a83492f6c2fe76e5:parameters","entityId":"epoch:05e1a83492f6c2fe76e5","metric":"parameters","value":93000000.0,"unit":"parameters","observationPeriod":"2020-11-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://arxiv.org/abs/2207.05477 reimplements AlphaFold 2 in a more efficient way, and states there are 93M parameters in the original version (Table 1)"},{"id":"epoch:05e1a83492f6c2fe76e5:training-compute","entityId":"epoch:05e1a83492f6c2fe76e5","metric":"training-compute","value":2.99e+21,"unit":"FLOP","observationPeriod":"2020-11-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"123 teraFLOPS / TPU v3 chip * 128 cores * (1 chip / 2 cores) * 11 days * 40% utilization = 2.99e21 FLOP\nhttps://www.wolframalpha.com/input?i=123+teraFLOPS+*+128+*+11+days+*+0.4\n\n\"Training regimen\" section: \n\"We train the model on Tensor Processing Unit (TPU) v3 with a batch size of 1 per TPU core, hence the model uses 128 TPU v3 cores. [...] The initial training stage takes approximately 1 week, and the fine-tuning stage takes approximately 4 additional days.\""},{"id":"epoch:c6f1b943cfc3ed27eb9d:parameters","entityId":"epoch:c6f1b943cfc3ed27eb9d","metric":"parameters","value":125000000.0,"unit":"parameters","observationPeriod":"2020-11-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c6f1b943cfc3ed27eb9d:training-compute","entityId":"epoch:c6f1b943cfc3ed27eb9d","metric":"training-compute","value":1.66e+21,"unit":"FLOP","observationPeriod":"2020-11-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From author communication\n\n\"About 128 GPU-days using Nvidia V100 (16GB). \"\n\nprecision: float16\n\nV100 GPU for float16: 28000000000000 (2.8E+13)\n\n0.4 * 28TFLOP/s * 128 GPU-days * 24h/day * 3600s/h\n= 1.24E+20\n\n\"and use the released roberta.base parameters for\ninitialization, which is a common practice to save\npre-training time\"\n\nRoberta base FLOP: 1.536e+21\nTotal:1.660000e+21\n"},{"id":"epoch:0e0c466c45078951e9b4:parameters","entityId":"epoch:0e0c466c45078951e9b4","metric":"parameters","value":37000000.0,"unit":"parameters","observationPeriod":"2020-11-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"all WT2 models have roughly 37M parameters\""},{"id":"epoch:0e0c466c45078951e9b4:training-compute","entityId":"epoch:0e0c466c45078951e9b4","metric":"training-compute","value":7.104e+17,"unit":"FLOP","observationPeriod":"2020-11-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 37000000 parameters * 2000000 tokens * 1600 epochs = 7.104e+17 FLOP"},{"id":"epoch:9dba0d79a4a573e5b5c8:parameters","entityId":"epoch:9dba0d79a4a573e5b5c8","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2020-11-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:3430d49d181fdcd4ed05:training-compute","entityId":"epoch:3430d49d181fdcd4ed05","metric":"training-compute","value":1.799999999999987e+18,"unit":"FLOP","observationPeriod":"2020-11-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1. Hardware setup: \n576 CPUs (288 nodes \u00d7 2 CPUs/node), Intel Xeon E5-2683 v4\nFLOP/s per CPU: 2.688 \u00d7 10\u00b9\u00b9 FLOP/s\nTotal system: 1.548288 \u00d7 10\u00b9\u2074 FLOP/s\n\n2. Training duration: \nDirectly provided: 8 hours = 28,800 seconds\n\n3. Utilization rate: \n40% (0.4)\n\n4. Final calculation:\n1.548288 \u00d7 10\u00b9\u2074 FLOP/s \u00d7 28,800s \u00d7 0.4 = 1.78 \u00d7 10\u00b9\u2078 FLOPs"},{"id":"epoch:2f899a0787f8aa2512c6:parameters","entityId":"epoch:2f899a0787f8aa2512c6","metric":"parameters","value":1700000.0,"unit":"parameters","observationPeriod":"2020-11-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b4ee923e8ec426817135:parameters","entityId":"epoch:b4ee923e8ec426817135","metric":"parameters","value":795000000.0,"unit":"parameters","observationPeriod":"2020-10-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From author communication\n\nWe trained different model sizes (from 24M to 795M), and they're summarized in Table 1 of the paper (https://arxiv.org/pdf/2006.10029.pdf)."},{"id":"epoch:09d38eb1e7694bd6938a:parameters","entityId":"epoch:09d38eb1e7694bd6938a","metric":"parameters","value":125000000.0,"unit":"parameters","observationPeriod":"2020-10-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our implementation of RoBERTa uses 12 attention heads and 6 layers, resulting in 72 distinct attention mechanisms\" -> base model is RoBERTa Base"},{"id":"epoch:09d38eb1e7694bd6938a:training-compute","entityId":"epoch:09d38eb1e7694bd6938a","metric":"training-compute","value":8.4654366e+18,"unit":"FLOP","observationPeriod":"2020-10-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 125 * 10^6 parameters * 64 sequences per batch [assumption] * 512 tokens per sequence [upper bound] * 450000 steps = 1.10592e+19 FLOP\n\n125000000000000 FLOP / GPU / sec [bf16 assumed] * 1 GPU * 48 hours * 3600 sec / hour * 0.3 [assumed utilization] = 6.48e+18 FLOP\n\nsqrt(1.10592e+19* 6.48e+18) = 8.4654366e+18 FLOP"},{"id":"epoch:e470cda56c8437ff6355:parameters","entityId":"epoch:e470cda56c8437ff6355","metric":"parameters","value":632000000.0,"unit":"parameters","observationPeriod":"2020-10-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1 https://arxiv.org/pdf/2010.11929.pdf"},{"id":"epoch:e470cda56c8437ff6355:training-compute","entityId":"epoch:e470cda56c8437ff6355","metric":"training-compute","value":4.262e+21,"unit":"FLOP","observationPeriod":"2020-10-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 6: 4.262e21 FLOPs\n\nAgrees with Table 2 (2.5k TPUv3-core days), if MFU is around 0.32. \n\n2500 * 24 * 3600 * (0.5 * 1.23e14) * 0.32 = 4.25e21"},{"id":"epoch:3524bbd4bc0f355d9e96:parameters","entityId":"epoch:3524bbd4bc0f355d9e96","metric":"parameters","value":317000000.0,"unit":"parameters","observationPeriod":"2020-10-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 5.1:\n\"We consider two model sizes: BASE (95m parameters) and LARGE (317m parameters)\n"},{"id":"epoch:3524bbd4bc0f355d9e96:training-compute","entityId":"epoch:3524bbd4bc0f355d9e96","metric":"training-compute","value":3.87072e+21,"unit":"FLOP","observationPeriod":"2020-10-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From surveying the authors:\n\nWe trained the base model on 64 V100 GPUs for 400k updates. This takes about 3 days to complete. The large model is trained on 128 V100 GPUs for 1 million updates, and this takes about 7 days to complete.\n\nV100 GPU peak: 125TFLOP/s (https://www.nvidia.com/en-gb/data-center/tesla-v100/)\nAssume 40% utilization based on default for non-Language domain (https://epoch.ai/blog/estimating-training-compute)\n\n128 GPUs * 40% * 125TFLOP/s * 7 days * 24h/day * 3600s/h\n~= 3.870720e+21"},{"id":"epoch:b87ea54bb1337002f6d4:parameters","entityId":"epoch:b87ea54bb1337002f6d4","metric":"parameters","value":86000000.0,"unit":"parameters","observationPeriod":"2020-10-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1 https://arxiv.org/pdf/2010.11929.pdf"},{"id":"epoch:af841ccbcc5cea16810e:parameters","entityId":"epoch:af841ccbcc5cea16810e","metric":"parameters","value":335000000.0,"unit":"parameters","observationPeriod":"2020-10-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"335M from Table 5"},{"id":"epoch:af841ccbcc5cea16810e:training-compute","entityId":"epoch:af841ccbcc5cea16810e","metric":"training-compute","value":2.2444646e+21,"unit":"FLOP","observationPeriod":"2020-10-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"flops = (64) * (123* 10**12) * (11 * 24 * 3600) * (0.3) = 2.24e21\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)\n\n'large models were trained on pods of 16 TPUs v3 (128 cores).' - from section 4.1it was trained for 11 days from Table 2"},{"id":"epoch:fa292e8a6b283a7c98a2:parameters","entityId":"epoch:fa292e8a6b283a7c98a2","metric":"parameters","value":335000000.0,"unit":"parameters","observationPeriod":"2020-10-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"335M from Table 5"},{"id":"epoch:fa292e8a6b283a7c98a2:training-compute","entityId":"epoch:fa292e8a6b283a7c98a2","metric":"training-compute","value":1.42829568e+21,"unit":"FLOP","observationPeriod":"2020-10-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"flops = (64) * (123* 10**12) * (7 * 24 * 3600) * (0.3) = 1.4e21\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)\n\n'large models were trained on pods of 16 TPUs v3 (128 cores).' - from section 4.1 it was trained for 7 days from Table 2\n\nAgrees with 6CN:\nTokens seen: 512 (seq len) * 1024 (batch size) * 1 million (steps) = 5.24e11\nFLOPs: 6 * 335M * 5.24e11 = 1.05e21"},{"id":"epoch:8d4434e242583fac82c6:parameters","entityId":"epoch:8d4434e242583fac82c6","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2020-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1B for XXL model"},{"id":"epoch:8d4434e242583fac82c6:training-compute","entityId":"epoch:8d4434e242583fac82c6","metric":"training-compute","value":7.6e+21,"unit":"FLOP","observationPeriod":"2020-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train with global batch size 2048 on 256/512 Google TPU V3 cores for 3-4 days for the XL/XXL models respectively...\nWe fine-tune the pre-trained checkpoints (400k steps) with global batch\nsize 1024/512 on 256/512 Google TPU v3 cores for 1-3 days for the XL/XXL models\"\n\nTPU v3 chips are 123 teraflop/s. 2 chips per core\n\n512 cores * 7 days * 24 * 3600 * 123 tflops * (1 chip/2 cores) * 0.4 (assumed utilization) = 7.6e21"},{"id":"epoch:1dbf13ef8506490addd3:parameters","entityId":"epoch:1dbf13ef8506490addd3","metric":"parameters","value":13000000000.0,"unit":"parameters","observationPeriod":"2020-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13 billion"},{"id":"epoch:1dbf13ef8506490addd3:training-compute","entityId":"epoch:1dbf13ef8506490addd3","metric":"training-compute","value":8.2e+22,"unit":"FLOP","observationPeriod":"2020-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We pre-train our mT5 model variants for 1 million steps on batches of 1024 length-1024 input sequences, corresponding to roughly 1 trillion input tokens total.\"\n\n1 million steps * 1024 batchsize * 1024 length * 13 billion params * 6 = 8.2e22\n\nIgnores fine-tuning compute; this is likely a small fraction of pre-training compute."},{"id":"epoch:0a979b78061925d9c526:parameters","entityId":"epoch:0a979b78061925d9c526","metric":"parameters","value":67000000.0,"unit":"parameters","observationPeriod":"2020-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"67M"},{"id":"epoch:0a979b78061925d9c526:training-compute","entityId":"epoch:0a979b78061925d9c526","metric":"training-compute","value":3.9798e+18,"unit":"FLOP","observationPeriod":"2020-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 67000000 parameters * 3300000000 tokens * 3 epochs = 3.9798e18 FLOP"},{"id":"epoch:a6e8578c740d742cff41:parameters","entityId":"epoch:a6e8578c740d742cff41","metric":"parameters","value":76200000.0,"unit":"parameters","observationPeriod":"2020-10-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a6e8578c740d742cff41:training-compute","entityId":"epoch:a6e8578c740d742cff41","metric":"training-compute","value":1.2e+19,"unit":"FLOP","observationPeriod":"2020-10-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32400000000000 FLOP / second / GPU * 4 GPUs * 96 hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.3436928e+19 FLOP\n\ntime could be given for all language modeling not just for this model -> likely to be an upper bound"},{"id":"epoch:24d1196e048287d8108e:parameters","entityId":"epoch:24d1196e048287d8108e","metric":"parameters","value":483000000.0,"unit":"parameters","observationPeriod":"2020-10-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The total number of parameters is approximately 483 M, consisting of 355 M in RoBERTa and 128 M in our entity embeddings\""},{"id":"epoch:24d1196e048287d8108e:training-compute","entityId":"epoch:24d1196e048287d8108e","metric":"training-compute","value":1.8144e+22,"unit":"FLOP","observationPeriod":"2020-10-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Uses RoBERTa Large as a base model, which used 1.66e22 FLOPs in training.\n\nLUKE's additional training was:\n(16) * (1.25e14) * (30 * 24 * 3600) * (0.3) = 1.5552e21\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)\n\nfrom appendix A: \"Werun the pretraining on NVIDIA\u2019s PyTorch Docker\ncontainer 19.02 hosted on a server with two Intel Xeon Platinum 8168 CPUs and 16 NVIDIA Tesla V100 GPUs. The training takes approximately 30 days.\"\n\nAssuming 16 bit tensor core computations, 1.25e14 FLOP/s per V100\n\nTotal: 1.65888e22 + 1.5552e21 = 1.8144e22"},{"id":"epoch:ee6e396ae7567b829ad9:parameters","entityId":"epoch:ee6e396ae7567b829ad9","metric":"parameters","value":33000000.0,"unit":"parameters","observationPeriod":"2020-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:989aeafe91294d1c5d3d:parameters","entityId":"epoch:989aeafe91294d1c5d3d","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2020-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:44ee7fcb014ad7e8a2ff:parameters","entityId":"epoch:44ee7fcb014ad7e8a2ff","metric":"parameters","value":28500000.0,"unit":"parameters","observationPeriod":"2020-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2aa85a81d99edfb5c741:parameters","entityId":"epoch:2aa85a81d99edfb5c741","metric":"parameters","value":13300000.0,"unit":"parameters","observationPeriod":"2020-09-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:eac015e7c61f73a4b136:parameters","entityId":"epoch:eac015e7c61f73a4b136","metric":"parameters","value":247000000.00000003,"unit":"parameters","observationPeriod":"2020-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:bfe80d36926c62a1755c:parameters","entityId":"epoch:bfe80d36926c62a1755c","metric":"parameters","value":44000000.0,"unit":"parameters","observationPeriod":"2020-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In total, our model has approximately 44M trainable parameters.\""},{"id":"epoch:bfe80d36926c62a1755c:training-compute","entityId":"epoch:bfe80d36926c62a1755c","metric":"training-compute","value":9.72e+18,"unit":"FLOP","observationPeriod":"2020-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we pre-train PRoBERTa on 4 NVIDIA V100 GPUs in 18 hours\"\n4 * 125 tFLOP/s * 18 * 3600 * 0.3 (assumed utilization) = 9.72e18"},{"id":"epoch:3bb512ef16c8f6b37ed6:parameters","entityId":"epoch:3bb512ef16c8f6b37ed6","metric":"parameters","value":669200000.0,"unit":"parameters","observationPeriod":"2020-08-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 1"},{"id":"epoch:3bb512ef16c8f6b37ed6:training-compute","entityId":"epoch:3bb512ef16c8f6b37ed6","metric":"training-compute","value":4.4e+20,"unit":"FLOP","observationPeriod":"2020-08-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Information: \n128 NVIDIA V100 GPUs [Pre-training details]\n840k steps [See Table S2: Hyperparameters]\n131,072 tokens per batch [\"We trained with 131,072 tokens per batch (128 gpus x 1024 tokens).\" - Pre-training details]\n\nEstimate:  840e3 updates * 3 * 131072 tokens/update * 2 * 669.2e6 parameters = 4.4e20 FLOP"},{"id":"epoch:3e10c4f0a699ff50254b:parameters","entityId":"epoch:3e10c4f0a699ff50254b","metric":"parameters","value":669200000.0,"unit":"parameters","observationPeriod":"2020-08-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 1"},{"id":"epoch:3e10c4f0a699ff50254b:training-compute","entityId":"epoch:3e10c4f0a699ff50254b","metric":"training-compute","value":4.8e+20,"unit":"FLOP","observationPeriod":"2020-08-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Information: \n128 NVIDIA V100 GPUs [Pre-training details]\n906k steps [See Table S2: Hyperparameters]\n131,072 tokens per batch [\"We trained with 131,072 tokens per batch (128 gpus x 1024 tokens).\" - Pre-training details]\n\nEstimate:  906e3 updates * 3 * 131072 tokens/update * 2 * 669.2e6 parameters = 4.8e20 FLOP"},{"id":"epoch:cc508b5e69c5870a16d0:parameters","entityId":"epoch:cc508b5e69c5870a16d0","metric":"parameters","value":669200000.0,"unit":"parameters","observationPeriod":"2020-08-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 1"},{"id":"epoch:cc508b5e69c5870a16d0:training-compute","entityId":"epoch:cc508b5e69c5870a16d0","metric":"training-compute","value":1.4e+20,"unit":"FLOP","observationPeriod":"2020-08-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Information: \n128 NVIDIA V100 GPUs [Pre-training details]\n275k steps [See Table S2: Hyperparameters]\n131,072 tokens per batch [\"We trained with 131,072 tokens per batch (128 gpus x 1024 tokens).\" - Pre-training details]\n\nEstimate:  275e3 updates * 3 * 131072 tokens/update * 2 * 669.2e6 parameters = 1.4e20 FLOP"},{"id":"epoch:e14ff3911af1827481b2:parameters","entityId":"epoch:e14ff3911af1827481b2","metric":"parameters","value":85100000.0,"unit":"parameters","observationPeriod":"2020-08-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 1"},{"id":"epoch:e14ff3911af1827481b2:training-compute","entityId":"epoch:e14ff3911af1827481b2","metric":"training-compute","value":5.6e+19,"unit":"FLOP","observationPeriod":"2020-08-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Information: \n128 NVIDIA V100 GPUs [Pre-training details]\n840k steps [See Table S2: Hyperparameters]\n131,072 tokens per batch [\"We trained with 131,072 tokens per batch (128 gpus x 1024 tokens).\" - Pre-training details]\n\nEstimate:  840e3 updates * 3 * 131072 tokens/update * 2 * 85.1e6 parameters = 5.6e19 FLOP"},{"id":"epoch:40522bc5b3e2ebbfec63:parameters","entityId":"epoch:40522bc5b3e2ebbfec63","metric":"parameters","value":42600000.0,"unit":"parameters","observationPeriod":"2020-08-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 1"},{"id":"epoch:40522bc5b3e2ebbfec63:training-compute","entityId":"epoch:40522bc5b3e2ebbfec63","metric":"training-compute","value":2.8e+19,"unit":"FLOP","observationPeriod":"2020-08-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Information: \n128 NVIDIA V100 GPUs [Pre-training details]\n840k steps [See Table S2: Hyperparameters]\n131,072 tokens per batch [\"We trained with 131,072 tokens per batch (128 gpus x 1024 tokens).\" - Pre-training details]\n\nEstimate:  840e3 updates * 3 * 131072 tokens/update * 2 * 42.6e6 parameters = 2.8e19 FLOP"},{"id":"epoch:cb4268cdc7d270d984b7:parameters","entityId":"epoch:cb4268cdc7d270d984b7","metric":"parameters","value":124000000.0,"unit":"parameters","observationPeriod":"2020-08-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:cb4268cdc7d270d984b7:training-compute","entityId":"epoch:cb4268cdc7d270d984b7","metric":"training-compute","value":7.9375326e+20,"unit":"FLOP","observationPeriod":"2020-08-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"base model compute (speculative confidence): 7.936e+20 FLOP \nfine-tune compute: 1.53264e+17 FLOP \n\n7.936e+20 FLOP + 1.53264e+17 FLOP  = 7.9375326e+20 FLOP"},{"id":"epoch:bb4179727732d2dc7c4b:parameters","entityId":"epoch:bb4179727732d2dc7c4b","metric":"parameters","value":340000000.0,"unit":"parameters","observationPeriod":"2020-08-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train a base model ERNIEGENBASE (L=12, H=768, A=12, Total Parameters=110M)1\nand a large model ERNIE-GENLARGE (L=24, H=1024,\nA=16, Total Parameters=340M) with parameters initialized\nby BERTBASE and BERTLARGE respectively\""},{"id":"epoch:bb4179727732d2dc7c4b:training-compute","entityId":"epoch:bb4179727732d2dc7c4b","metric":"training-compute","value":2e+20,"unit":"FLOP","observationPeriod":"2020-08-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"430GB text for 1 epoch\n\napprox 430 * 200 million words = 86B words, or 100B tokens per https://docs.google.com/document/d/1G3vvQkn4x_W71MKg0GmHVtzfd9m0y3_Ofcoew0v902Q/edit#heading=h.ieihc08p8dn0\n\n6 * 340 million params * 100 billion tokens ~= 2e20"},{"id":"epoch:1dd12c50c40d0d205818:parameters","entityId":"epoch:1dd12c50c40d0d205818","metric":"parameters","value":99000000.0,"unit":"parameters","observationPeriod":"2020-08-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"99M (Table 4b)"},{"id":"epoch:1dd12c50c40d0d205818:training-compute","entityId":"epoch:1dd12c50c40d0d205818","metric":"training-compute","value":3.8016e+18,"unit":"FLOP","observationPeriod":"2020-08-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 99 * 10^6 parameters * 100000 steps * 64000 tokens per batch = 3.8016e+18 FLOP\n\n31330000000000 FLOP / second / GPU * 8 GPUs * 30 hours [assumed based on smaller models reported training time] * 3600 sec / hour * 0.3 [assumed utilization] = 8.120736e+18 FLOP\n\nOperation counting method uses less assumptions"},{"id":"epoch:c3f6059297f011e8e142:parameters","entityId":"epoch:c3f6059297f011e8e142","metric":"parameters","value":610000000.0,"unit":"parameters","observationPeriod":"2020-08-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"610M\nfrom https://github.com/facebookresearch/fairseq/tree/main/examples/mbart"},{"id":"epoch:c3f6059297f011e8e142:training-compute","entityId":"epoch:c3f6059297f011e8e142","metric":"training-compute","value":1.45152e+22,"unit":"FLOP","observationPeriod":"2020-08-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"flops = (256) * (125000000000000) * (2.5 * 7 * 24 * 3600) * (0.3) = 1.45152e+22\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)\n\n\n\"mBART trained for 2.5 weeks on 256 Nvidia V100 GPUs\"\nV100 have peak flop 28.26 TFLOPS  from https://www.techpowerup.com/gpu-specs/tesla-v100-pcie-16-gb.c2957"},{"id":"epoch:d98db577a57ce012e639:parameters","entityId":"epoch:d98db577a57ce012e639","metric":"parameters","value":77000000.0,"unit":"parameters","observationPeriod":"2020-07-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"EfficientDet-D7 achieves stateof-the-art 55.1 AP on COCO test-dev with 77M parameters and 410B FLOPs\""},{"id":"epoch:ee4ad85fdb1450d90805:parameters","entityId":"epoch:ee4ad85fdb1450d90805","metric":"parameters","value":86500000.0,"unit":"parameters","observationPeriod":"2020-07-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e04956aa34d6b689777a:parameters","entityId":"epoch:e04956aa34d6b689777a","metric":"parameters","value":1000000000000.0,"unit":"parameters","observationPeriod":"2020-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 1\n\nhttps://arxiv.org/abs/2104.05158"},{"id":"epoch:e04956aa34d6b689777a:training-compute","entityId":"epoch:e04956aa34d6b689777a","metric":"training-compute","value":3e+20,"unit":"FLOP","observationPeriod":"2020-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 1\n\nhttps://arxiv.org/abs/2104.05158"},{"id":"epoch:11c7ff68e4e1fb8e105b:parameters","entityId":"epoch:11c7ff68e4e1fb8e105b","metric":"parameters","value":2300000000.0,"unit":"parameters","observationPeriod":"2020-06-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our best quality dense single Transformer model (2.3B parameters) achieving \u2206BLEU of 6.1, was trained with GPipe [15] on 2048 TPU v3 cores for 6 weeks or total of 235.5 TPU v3 core-years.\""},{"id":"epoch:11c7ff68e4e1fb8e105b:training-compute","entityId":"epoch:11c7ff68e4e1fb8e105b","metric":"training-compute","value":4.765e+22,"unit":"FLOP","observationPeriod":"2020-06-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Trained for a total of 235.5 TPU v3 core-years.\nHardware estimate: 235.5 * 365.25 * 24 * 3600 * (1.23e14 / 2) * 0.3 = 1.371e23\n\nFootnote 10 indicates 300k steps and 4M tokens/step -> 1.2T tokens\nArithmetic estimate: 6 * 2.3B * 1.2T = 1.656e22 FLOPs\n\nGeometric mean: sqrt(1.371e23 * 1.656e22) = 4.765e22"},{"id":"epoch:0950ef59dda702f47070:parameters","entityId":"epoch:0950ef59dda702f47070","metric":"parameters","value":600000000000.0,"unit":"parameters","observationPeriod":"2020-06-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The 600B parameters model that achieved the best translation quality was trained with 2048 TPU v3 cores for 4 days, a total cost of 22 TPU v3 core-years.\""},{"id":"epoch:0950ef59dda702f47070:training-compute","entityId":"epoch:0950ef59dda702f47070","metric":"training-compute","value":1.33e+22,"unit":"FLOP","observationPeriod":"2020-06-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://arxiv.org/ftp/arxiv/papers/2104/2104.10350.pdf\nTable 4\n\n123000000000000 FLOP / chip / sec * 96360 chip-hours [see training time notes] * 3600 sec / hour * 0.3 [assumed utilization] = 1.2800462e+22 FLOP"},{"id":"epoch:e904d1ab3d25dbce7cdc:parameters","entityId":"epoch:e904d1ab3d25dbce7cdc","metric":"parameters","value":6660000000.0,"unit":"parameters","observationPeriod":"2020-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6.7B\n"},{"id":"epoch:e904d1ab3d25dbce7cdc:training-compute","entityId":"epoch:e904d1ab3d25dbce7cdc","metric":"training-compute","value":1.2e+22,"unit":"FLOP","observationPeriod":"2020-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table D.1\nhttps://arxiv.org/abs/2005.14165"},{"id":"epoch:6684f356882a0980747d:parameters","entityId":"epoch:6684f356882a0980747d","metric":"parameters","value":1320000000.0,"unit":"parameters","observationPeriod":"2020-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" 1.3B"},{"id":"epoch:6684f356882a0980747d:training-compute","entityId":"epoch:6684f356882a0980747d","metric":"training-compute","value":2.38e+21,"unit":"FLOP","observationPeriod":"2020-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table D.1\nhttps://arxiv.org/abs/2005.14165"},{"id":"epoch:9690d1f47fe040daf9fe:parameters","entityId":"epoch:9690d1f47fe040daf9fe","metric":"parameters","value":125000000.0,"unit":"parameters","observationPeriod":"2020-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"125M"},{"id":"epoch:9690d1f47fe040daf9fe:training-compute","entityId":"epoch:9690d1f47fe040daf9fe","metric":"training-compute","value":2.25e+20,"unit":"FLOP","observationPeriod":"2020-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table D.1\nhttps://arxiv.org/abs/2005.14165"},{"id":"epoch:6ba49112fdbdd03cf602:parameters","entityId":"epoch:6ba49112fdbdd03cf602","metric":"parameters","value":356000000.0,"unit":"parameters","observationPeriod":"2020-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"350M"},{"id":"epoch:6ba49112fdbdd03cf602:training-compute","entityId":"epoch:6ba49112fdbdd03cf602","metric":"training-compute","value":6.41e+20,"unit":"FLOP","observationPeriod":"2020-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table D.1\nhttps://arxiv.org/abs/2005.14165"},{"id":"epoch:a08435f5c951ae46f3ae:parameters","entityId":"epoch:a08435f5c951ae46f3ae","metric":"parameters","value":760000000.0,"unit":"parameters","observationPeriod":"2020-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"760M"},{"id":"epoch:a08435f5c951ae46f3ae:training-compute","entityId":"epoch:a08435f5c951ae46f3ae","metric":"training-compute","value":1.37e+21,"unit":"FLOP","observationPeriod":"2020-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table D.1\nhttps://arxiv.org/abs/2005.14165"},{"id":"epoch:a88c26838d31cd926259:parameters","entityId":"epoch:a88c26838d31cd926259","metric":"parameters","value":2650000000.0,"unit":"parameters","observationPeriod":"2020-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.7B\n"},{"id":"epoch:a88c26838d31cd926259:training-compute","entityId":"epoch:a88c26838d31cd926259","metric":"training-compute","value":4.77e+21,"unit":"FLOP","observationPeriod":"2020-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table D.1\nhttps://arxiv.org/abs/2005.14165"},{"id":"epoch:1ae789f2543baed88459:parameters","entityId":"epoch:1ae789f2543baed88459","metric":"parameters","value":12850000000.0,"unit":"parameters","observationPeriod":"2020-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"13B"},{"id":"epoch:1ae789f2543baed88459:training-compute","entityId":"epoch:1ae789f2543baed88459","metric":"training-compute","value":2.31e+22,"unit":"FLOP","observationPeriod":"2020-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table D.1\nhttps://arxiv.org/abs/2005.14165"},{"id":"epoch:75bc94ec2ab97dd825d1:parameters","entityId":"epoch:75bc94ec2ab97dd825d1","metric":"parameters","value":6801000000.0,"unit":"parameters","observationPeriod":"2020-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://openai.com/blog/image-gpt/#rfref53"},{"id":"epoch:75bc94ec2ab97dd825d1:training-compute","entityId":"epoch:75bc94ec2ab97dd825d1","metric":"training-compute","value":3.3e+22,"unit":"FLOP","observationPeriod":"2020-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Taken from here\nhttps://www.lesswrong.com/posts/wfpdejMWog4vEDLDg/ai-and-compute-trend-isn-t-predictive-of-what-is-happening (\"There's no compute data for the largest model, iGPT-XL. But based on the FLOP/s increase from GPT-3 XL (same num of params as iGPT-L) to GPT-3 6.7B (same num of params as iGPT-XL), I think it required 5 times more compute: 3.3 * 10^22 FLOP.\")"},{"id":"epoch:3f4bd1e64aa4f4c06905:parameters","entityId":"epoch:3f4bd1e64aa4f4c06905","metric":"parameters","value":1362000000.0,"unit":"parameters","observationPeriod":"2020-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://openai.com/blog/image-gpt/#rfref53"},{"id":"epoch:3f4bd1e64aa4f4c06905:training-compute","entityId":"epoch:3f4bd1e64aa4f4c06905","metric":"training-compute","value":8.91e+21,"unit":"FLOP","observationPeriod":"2020-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"We have that \"iGPT-L was trained for roughly 2500 V100-days\" [1]\n\nI assume this is the NVIDIA Tesla V100 GPU. In the specifications, the NVIDIA Tesla V100 has 7 to 8.2 TFLOPS of peak double precision performance and 14 to 16.4 TFLOPS of peak single precision performance and 112 to 130 TFLOPS of peak tensor performance [2].\n\nI suppose the one that makes sense using if peak tensor performance, for ~125 TFLOPS peak tensor performance more or less.\nFollowing OpenAIs AI and compute we apply a 0.33 utitilization factor [3].\n\nIn total we get 2500 V100-days * (24*60*60) seconds/day * 125 TFLOPS * 0.33 = 8.91e+21 FLOPS = 89.1 PF-days.\n\n[1] https://openai.com/blog/image-gpt/\n[2] https://images.nvidia.com/content/technologies/volta/pdf/volta-v100-datasheet-update-us-1165301-r5.pdf\n[3] https://openai.com/blog/ai-and-compute/"},{"id":"epoch:f3c08cc420de98eb067c:parameters","entityId":"epoch:f3c08cc420de98eb067c","metric":"parameters","value":256999999.99999997,"unit":"parameters","observationPeriod":"2020-06-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9936b7f12fd375c78a52:parameters","entityId":"epoch:9936b7f12fd375c78a52","metric":"parameters","value":256999999.99999997,"unit":"parameters","observationPeriod":"2020-06-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2cbbf35695663a3e0870:parameters","entityId":"epoch:2cbbf35695663a3e0870","metric":"parameters","value":51100000.0,"unit":"parameters","observationPeriod":"2020-06-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Rados"},{"id":"epoch:f8de8170c6750b99b599:parameters","entityId":"epoch:f8de8170c6750b99b599","metric":"parameters","value":85500000.0,"unit":"parameters","observationPeriod":"2020-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\" Following [35], a three-layer tensorized transformer core-1 for PTB and a six-layer tensorized transformer core-1 for Wikitext-103 are used in the experiments\"\n\n[35] Ma, X.; Zhang, P.; Zhang, S.; Duan, N.; Hou, Y.; Zhou,\nM.; and Song, D. 2019. A tensorized transformer for\nlanguage modeling. In Advances in Neural Information\nProcessing Systems, 2229\u20132239.\n\nhttps://arxiv.org/pdf/1906.09777\n\nTensorized transformer has 85.5M parameters"},{"id":"epoch:f8de8170c6750b99b599:training-compute","entityId":"epoch:f8de8170c6750b99b599","metric":"training-compute","value":1.58e+18,"unit":"FLOP","observationPeriod":"2020-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 85500000 parameters * 103000000 tokens * 30 epochs = 1.58517e+18 FLOP"},{"id":"epoch:e094cfe34cc090abb3fe:parameters","entityId":"epoch:e094cfe34cc090abb3fe","metric":"parameters","value":174600000000.0,"unit":"parameters","observationPeriod":"2020-05-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we train GPT-3, an autoregressive language model with 175 billion parameters\"\nRather, it's 174.6 billion.\nTable D.1: 174,600 million parameters"},{"id":"epoch:e094cfe34cc090abb3fe:training-compute","entityId":"epoch:e094cfe34cc090abb3fe","metric":"training-compute","value":3.14e+23,"unit":"FLOP","observationPeriod":"2020-05-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table D.1\nhttps://arxiv.org/abs/2005.14165"},{"id":"epoch:fcb06caf1bf69516d79f:parameters","entityId":"epoch:fcb06caf1bf69516d79f","metric":"parameters","value":6700000000.0,"unit":"parameters","observationPeriod":"2020-05-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6.7B"},{"id":"epoch:fcb06caf1bf69516d79f:training-compute","entityId":"epoch:fcb06caf1bf69516d79f","metric":"training-compute","value":1.2e+22,"unit":"FLOP","observationPeriod":"2020-05-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 6.7 * 10^9 parameters * 300 * 10^9 tokens = 1.206e+22 FLOP"},{"id":"epoch:e4ba797ce4ecfe64267b:parameters","entityId":"epoch:e4ba797ce4ecfe64267b","metric":"parameters","value":60000000.0,"unit":"parameters","observationPeriod":"2020-05-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"60M per Table 1"},{"id":"epoch:e4ba797ce4ecfe64267b:training-compute","entityId":"epoch:e4ba797ce4ecfe64267b","metric":"training-compute","value":4e+20,"unit":"FLOP","observationPeriod":"2020-05-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Training the baseline model for 300 epochs on 16 V100 GPUs takes 3 days, with 4 images per GPU (hence a total batch size of 64). For the longer schedule used to compare with Faster R-CNN we train for 500 epochs with learning rate drop after 400 epochs. This schedule adds 1.5 AP compared to the shorter schedule.\"\n\n48 V100-days for baseline DETR model. Larger model had 1.5x the params and 5/3 as many epochs, so required ~2.5x as much training compute.\n\n125 teraflop/s * 2.5 * 48 * 24 * 3600 * 0.3 (assumed utilization) ~ 4e20"},{"id":"epoch:769ac90d053d56465f5b:parameters","entityId":"epoch:769ac90d053d56465f5b","metric":"parameters","value":626000000.0,"unit":"parameters","observationPeriod":"2020-05-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our RAG models contain the trainable parameters for the BERT-base query and document encoder of DPR, with 110M parameters each (although we do not train the document encoder ourselves) and 406M trainable parameters from BART-large, 406M parameters, making a total of 626M trainable parameters\""},{"id":"epoch:c9a3cf01b2ef0a9edd2a:parameters","entityId":"epoch:c9a3cf01b2ef0a9edd2a","metric":"parameters","value":69000000.0,"unit":"parameters","observationPeriod":"2020-05-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" We adopted a 2-layer LSTM language model architec- ture with 1500 hidden units per layer\n\nOkay, let's break this down:\n\n- You have a 2-layer LSTM model\n- Each LSTM layer has 1500 hidden units\n- The vocabulary size is 10k\n\nFor each LSTM layer:\n\n- There are 4 weight matrices (input, forget, output, cell state) per layer \n- Each weight matrix has dimensions (input size x 1500 hidden units)\n- Plus 1500 bias terms per matrix\n\nLayer 1:\n- Input size is 10k (vocab size)\n- Weight matrices: 4 * (10k * 1500) = 60 million parameters \n- Biases: 4 * 1500 = 6000 parameters\n\nLayer 2: \n- Input size is 1500 (output of layer 1)  \n- Weight matrices: 4 * (1500 * 1500) = 9 million parameters\n- Biases: 4 * 1500 = 6000 parameters\n\nTotal parameters = 60 million + 6000 + 9 million + 6000 \n             = 69,012,000\n\nSo the total number of parameters in this 2-layer 1500-unit LSTM with 10k vocab size is approximately 69 million."},{"id":"epoch:c9a3cf01b2ef0a9edd2a:training-compute","entityId":"epoch:c9a3cf01b2ef0a9edd2a","metric":"training-compute","value":1.4352996e+16,"unit":"FLOP","observationPeriod":"2020-05-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 69000000 parameters * 912344 tokens * 38 epochs = 1.4352996e+16 FLOP"},{"id":"epoch:491f8cbbe49a42edba80:parameters","entityId":"epoch:491f8cbbe49a42edba80","metric":"parameters","value":110000000.0,"unit":"parameters","observationPeriod":"2020-05-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We initialize the model from SciBERT pretrained weights (Beltagy et al., 2019) since it is the stateof-the-art pretrained language model on scientific text\"\n\nSciBERT has 110M parameters"},{"id":"epoch:138dd3d67b287a9236e0:parameters","entityId":"epoch:138dd3d67b287a9236e0","metric":"parameters","value":1400000.0,"unit":"parameters","observationPeriod":"2020-05-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:af338a663fbf2728f3df:parameters","entityId":"epoch:af338a663fbf2728f3df","metric":"parameters","value":118800000.0,"unit":"parameters","observationPeriod":"2020-05-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"118.8M for Conformer(L)"},{"id":"epoch:0bf3c2d354b10a9fe7cc:parameters","entityId":"epoch:0bf3c2d354b10a9fe7cc","metric":"parameters","value":25000000.0,"unit":"parameters","observationPeriod":"2020-05-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"25M\nTable 1 "},{"id":"epoch:0bf3c2d354b10a9fe7cc:training-compute","entityId":"epoch:0bf3c2d354b10a9fe7cc","metric":"training-compute","value":1.368516e+17,"unit":"FLOP","observationPeriod":"2020-05-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 25 * 10^6 parameters * 912 344 tokens * 1000 epochs = 1.368516e+17 FLOP"},{"id":"epoch:47877fd9ba6da1ce3f35:parameters","entityId":"epoch:47877fd9ba6da1ce3f35","metric":"parameters","value":112700000.0,"unit":"parameters","observationPeriod":"2020-05-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 5"},{"id":"epoch:d47de3eaba7cee620c59:parameters","entityId":"epoch:d47de3eaba7cee620c59","metric":"parameters","value":156000000.0,"unit":"parameters","observationPeriod":"2020-05-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"156M (Table 2)"},{"id":"epoch:d47de3eaba7cee620c59:training-compute","entityId":"epoch:d47de3eaba7cee620c59","metric":"training-compute","value":2.89e+18,"unit":"FLOP","observationPeriod":"2020-05-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 156 * 10^6 parameters * 103000000 tokens * 30 epochs = 2.89224e+18 FLOP"},{"id":"epoch:7dae58133ce42b3d828d:parameters","entityId":"epoch:7dae58133ce42b3d828d","metric":"parameters","value":23000000.0,"unit":"parameters","observationPeriod":"2020-05-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a4a696178fce68b414f5:parameters","entityId":"epoch:a4a696178fce68b414f5","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2020-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"11B from appendix A.2 : Model sizes: \"Most of the experiments are done on T5(11B) which has 11 billion parameters. We also report experiments with BART (large) with 440 million parameters.\""},{"id":"epoch:a4a696178fce68b414f5:training-compute","entityId":"epoch:a4a696178fce68b414f5","metric":"training-compute","value":1.65e+19,"unit":"FLOP","observationPeriod":"2020-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"A.2: \"In the experiments, we use v3-8 TPUs for T5 models... pretraining UNIFIEDQA approximately takes about 36 hours on T5(11B)\"\n(Note that v3-8 refers to 8 TPUv3 cores, or 4 TPUv3 chips. From the JAX github repo: \"In a TPU v3-8 you have 4 chips, 2 cores per chip, and each core has 16GB of HBM. So this looks to JAX like 8 devices with 16GB each.\")\n\n4 * 1.23e14 * 36 * 3600 * 0.3 = 1.91e19\n\nAlternatively, input (ouput) size of 512 (100) tokens, batch size of 8, trained for 100k steps. Input tokens only need the forward pass.\n((2 * 11B * 512 * 8) + (6 * 11B * 100 * 8)) * 100k = 1.43e19\n\nTook geometric mean of these estimates:\nsqrt(1.91e19*1.43e19) = 1.65e19"},{"id":"epoch:09a8bfee80758efe38ac:parameters","entityId":"epoch:09a8bfee80758efe38ac","metric":"parameters","value":256999999.99999997,"unit":"parameters","observationPeriod":"2020-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Following Transformer-XL, we train a base size model and a large size model. <..> The large model is an 18 layer\nTransformer with a hidden size of 1024 and 16 attention\nheads\""},{"id":"epoch:09a8bfee80758efe38ac:training-compute","entityId":"epoch:09a8bfee80758efe38ac","metric":"training-compute","value":2.6527334e+19,"unit":"FLOP","observationPeriod":"2020-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 257000000 parameters * 350000 steps * 128 sequences per batch * 384 tokens per sequence = 2.6527334e+19 FLOP"},{"id":"epoch:87c38c18d0b8b8d75dd0:parameters","entityId":"epoch:87c38c18d0b8b8d75dd0","metric":"parameters","value":257000000.0,"unit":"parameters","observationPeriod":"2020-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:87c38c18d0b8b8d75dd0:training-compute","entityId":"epoch:87c38c18d0b8b8d75dd0","metric":"training-compute","value":2.653e+19,"unit":"FLOP","observationPeriod":"2020-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" The large model is an 18 layer\nTransformer with a hidden size of 1024 and 16 attention\nheads. This model is trained with 350K steps with a batch\nsize of 128. The sequence length and memory length dur-\ning training and testing all equal 150 for the base model\nand 384 for the large model\n\nTotal tokens: 350000*128*384=17203200000\n\nTraining FLOP: 6*17203200000*257000000=2.6527334e+19"},{"id":"epoch:7d4545c9db0c311ecc76:parameters","entityId":"epoch:7d4545c9db0c311ecc76","metric":"parameters","value":7700000.0,"unit":"parameters","observationPeriod":"2020-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Since\nall of these sub-networks share the same weights (i.e., Wo) (Cheung et al., 2019), we only require\n7.7M parameters to store all of them. Without sharing, the total model size will be prohibitive\""},{"id":"epoch:7d4545c9db0c311ecc76:training-compute","entityId":"epoch:7d4545c9db0c311ecc76","metric":"training-compute","value":6.237e+20,"unit":"FLOP","observationPeriod":"2020-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4.2k V100-hours (table 1)\n0.33 utilization rate\nV100 FP16 Tensor FLOPs: 125000000000000\n\n4200 hours*60*60* 125000000000000 FLOP/s *0.33 utilization =623700000000000000000\n"},{"id":"epoch:17d17781c68825376175:training-compute","entityId":"epoch:17d17781c68825376175","metric":"training-compute","value":1.0660896e+17,"unit":"FLOP","observationPeriod":"2020-04-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(1) * (4113 * 10**9) * (24 * 3600) * (0.3) = 106608960000000000\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate) = \n\n\"For each domain, we searched for 100 iterations (\u224815min/iter, 24 hours total) on a Tesla K80\"\n4.113 TFLOPS from https://www.techpowerup.com/gpu-specs/tesla-k80.c2616\n\n\n"},{"id":"epoch:576486cbbb1243764c74:parameters","entityId":"epoch:576486cbbb1243764c74","metric":"parameters","value":907264.0,"unit":"parameters","observationPeriod":"2020-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c3bb188582e2893c7be5:parameters","entityId":"epoch:c3bb188582e2893c7be5","metric":"parameters","value":25300000.0,"unit":"parameters","observationPeriod":"2020-04-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Rados"},{"id":"epoch:072d3c6b8c095f3de181:parameters","entityId":"epoch:072d3c6b8c095f3de181","metric":"parameters","value":1010000.0,"unit":"parameters","observationPeriod":"2020-04-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"All models trained on this dataset consisted of 100 hidden units\n\nOkay, let's break this down:\n\n- RNN with 1 hidden layer\n- 100 hidden units\n- Vocabulary size = 10,000\n\nThe RNN layer will have:\n\n- Input weight matrix U: dimensions (vocab size x hidden size) = (10,000 x 100) = 1,000,000 parameters\n\n- Recurrent weight matrix W: dimensions (hidden size x hidden size) = (100 x 100) = 10,000 parameters \n\n- Bias vector b: (hidden size) = (100) = 100 parameters\n\nSo the total parameters is:\n\nU = 1,000,000 \nW = 10,000\nb = 100\n\nTotal parameters = U + W + b \n            = 1,000,000 + 10,000 + 100\n            = 1,010,100\n\nTherefore, the total number of parameters for this 1-layer RNN with 100 hidden units and 10,000 vocab size is 1,010,100."},{"id":"epoch:072d3c6b8c095f3de181:training-compute","entityId":"epoch:072d3c6b8c095f3de181","metric":"training-compute","value":276440230000000.0,"unit":"FLOP","observationPeriod":"2020-04-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 1010000 parameters * 912 344 tokens * 50 epochs = 2.7644023e+14 FLOP"},{"id":"epoch:cecd2fcaeaf610c81874:parameters","entityId":"epoch:cecd2fcaeaf610c81874","metric":"parameters","value":110000000.0,"unit":"parameters","observationPeriod":"2020-03-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"We use the BERTbase configuration that has 12 encoder blocks, 768 hidden dimensions, 12 attention heads, 512 maximum sequence\nlength, and a total of \u223c110M parameters"},{"id":"epoch:cecd2fcaeaf610c81874:training-compute","entityId":"epoch:cecd2fcaeaf610c81874","metric":"training-compute","value":3.1765134e+19,"unit":"FLOP","observationPeriod":"2020-03-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 110000000 parameters * 81920000000 total training tokens [see dataset size notes] = 5.40672e+19 FLOP\n\n45000000000000 FLOP / sec / chip * 4 chips [=8 cores] * 96 hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.86624e+19 FLOP \n\nsqrt(5.40672e+19*1.86624e+19) = 3.1765134e+19 FLOP"},{"id":"epoch:ebfde7da14e00260ec03:parameters","entityId":"epoch:ebfde7da14e00260ec03","metric":"parameters","value":371000000.0,"unit":"parameters","observationPeriod":"2020-03-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"371M"},{"id":"epoch:ebfde7da14e00260ec03:training-compute","entityId":"epoch:ebfde7da14e00260ec03","metric":"training-compute","value":1.5399499e+21,"unit":"FLOP","observationPeriod":"2020-03-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * (128*13440*250000 + 512*2056*300000) total training tokens [see dataset size notes] * 371000000 parameters = 1.6603324e+21 FLOP\n\n123000000000000 FLOP / chip / sec * 64 chips [=128 cores] * 168 hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.4282957e+21 FLOP\n\nsqrt(1.6603324e+21*1.4282957e+21) = 1.5399499e+21 FLOP"},{"id":"epoch:276ed5dd9808d3ef899a:parameters","entityId":"epoch:276ed5dd9808d3ef899a","metric":"parameters","value":225000000.0,"unit":"parameters","observationPeriod":"2020-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\" In total this setting for MetNet has 225M parameters.\""},{"id":"epoch:276ed5dd9808d3ef899a:training-compute","entityId":"epoch:276ed5dd9808d3ef899a","metric":"training-compute","value":9.510912e+18,"unit":"FLOP","observationPeriod":"2020-03-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 7045120000 updates [see training dataset size notes, \"Likely\" confidence] * 225000000 parameters = 9.510912e+18 FLOP\n\ncheck: they report training in 256 TPUs (TPU v3 most likely) \n9.510912e+18 FLOP / (123000000000000 FLOP / sec / chip * 256 TPUs * 3600 sec / hour * 0.3 [assumed utilization]) = 0.27 hours (seems to be an underestimation) -> 'Speculative'  confidence"},{"id":"epoch:a16b64438e269f24af8e:parameters","entityId":"epoch:a16b64438e269f24af8e","metric":"parameters","value":335000000.0,"unit":"parameters","observationPeriod":"2020-03-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://github.com/google-research/electra"},{"id":"epoch:a16b64438e269f24af8e:training-compute","entityId":"epoch:a16b64438e269f24af8e","metric":"training-compute","value":3.1e+21,"unit":"FLOP","observationPeriod":"2020-03-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 8: \"ELECTRA-1.75M\" used 3.1e21 train FLOPs. Note that the actual parameter count is 335M. The 1.75M refers to the number of training steps.\n\nThis doesn't quite line up with a 6ND estimate, \n6 * 335M * (1.75M * 2048 * 128) = 9.22e20 FLOPs\nI'm inferring 128 sequence length, possibly this is 256 or 512?"},{"id":"epoch:a2328cd71773fd140da1:parameters","entityId":"epoch:a2328cd71773fd140da1","metric":"parameters","value":85300000.0,"unit":"parameters","observationPeriod":"2020-03-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"six layers tensorized transformer core-1 for Wikitext-103, following (Ma et al., 2019).\n\nhttps://arxiv.org/abs/1906.09777"},{"id":"epoch:a2328cd71773fd140da1:training-compute","entityId":"epoch:a2328cd71773fd140da1","metric":"training-compute","value":1.58e+18,"unit":"FLOP","observationPeriod":"2020-03-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 85300000 parameters * 103000000 tokens * 30 epochs = 1.581462e+18 FLOP"},{"id":"epoch:55bdb43af3ec94dd4c56:parameters","entityId":"epoch:55bdb43af3ec94dd4c56","metric":"parameters","value":672080.0,"unit":"parameters","observationPeriod":"2020-03-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:55bdb43af3ec94dd4c56:training-compute","entityId":"epoch:55bdb43af3ec94dd4c56","metric":"training-compute","value":2.4362208e+18,"unit":"FLOP","observationPeriod":"2020-03-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"31330000000000 FLOP / GPU/ sec [Tesla V100] * 1 GPU * 72 hours * 3600 sec / hour * 0.3 [assumed utilization] = 2.4362208e+18 FLOP"},{"id":"epoch:40fb5d88efbcab2a3732:parameters","entityId":"epoch:40fb5d88efbcab2a3732","metric":"parameters","value":1200000000.0,"unit":"parameters","observationPeriod":"2020-03-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train a 1.2B-parameter language model, ProGen, on \u223c280M protein sequences\""},{"id":"epoch:40fb5d88efbcab2a3732:training-compute","entityId":"epoch:40fb5d88efbcab2a3732","metric":"training-compute","value":3.7e+20,"unit":"FLOP","observationPeriod":"2020-03-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Our model was implemented in TensorFlow (Abadi et al.,\n2016) and trained with a global batch size of 64 distributed\nacross 256 cores of a Cloud TPU v3 Pod for 1M iterations. Training took approximately two weeks using Adagrad (Duchi et al., 2011)\n\n4.00E+12*256*60**2*24*14*0.3 = 3.7e20\n\n7.6E+21 FLOPs from Table 9 https://www.biorxiv.org/content/10.1101/2023.07.05.547496v1"},{"id":"epoch:f48990ebace3f82ec5e1:parameters","entityId":"epoch:f48990ebace3f82ec5e1","metric":"parameters","value":79500000.0,"unit":"parameters","observationPeriod":"2020-03-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f2d633ea7ab2bf1d4629:parameters","entityId":"epoch:f2d633ea7ab2bf1d4629","metric":"parameters","value":151000000.0,"unit":"parameters","observationPeriod":"2020-03-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"151M (Table 2)\n\n\" On the large WikiText-103 dataset, we implement our method based on the state-of-the-art Transformer-XL based models (Dai et al., 2019). We follow the same settings reported in (Dai et al., 2019), and our implementation is based on the\nofficial code for Transformer-XL.\""},{"id":"epoch:f2d633ea7ab2bf1d4629:training-compute","entityId":"epoch:f2d633ea7ab2bf1d4629","metric":"training-compute","value":2.6289761e+19,"unit":"FLOP","observationPeriod":"2020-03-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token *  151000000 parameters * 103000000 tokens * 250 epochs = 2.33295e+19 FLOP\n\n31330000000000 FLOP / sec / GPU [fp16] * 4 GPUs * 3152 sec per epoch * 250 epochs * 0.3 [assumed precision] = 2.9625648e+19 FLOP\n\nsqrt(2.33295e+19*2.9625648e+19) = 2.6289761e+19 FLOP\n\n'speculative' confidence due to the assumption of the amount of epochs"},{"id":"epoch:d080c2a4f16f509f3778:parameters","entityId":"epoch:d080c2a4f16f509f3778","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2020-03-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"24M imputed from Merity et al. 2017 https://arxiv.org/abs/1708.02182"},{"id":"epoch:d080c2a4f16f509f3778:training-compute","entityId":"epoch:d080c2a4f16f509f3778","metric":"training-compute","value":2.6275507e+16,"unit":"FLOP","observationPeriod":"2020-03-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 24000000 parameters * 912344 tokens * 200 epochs = 2.6275507e+16 FLOP"},{"id":"epoch:7c48f2c57c8bd74c7f7d:parameters","entityId":"epoch:7c48f2c57c8bd74c7f7d","metric":"parameters","value":33000000.0,"unit":"parameters","observationPeriod":"2020-02-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:74921702da8e493aed79:parameters","entityId":"epoch:74921702da8e493aed79","metric":"parameters","value":13000000.0,"unit":"parameters","observationPeriod":"2020-02-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:13fc6f983d829587433b:parameters","entityId":"epoch:13fc6f983d829587433b","metric":"parameters","value":126000000.0,"unit":"parameters","observationPeriod":"2020-02-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 3 shows 126M.\n\nThere is another instance of the Feedback Transformer mentioned in Table 9 with 139M parameters."},{"id":"epoch:13fc6f983d829587433b:training-compute","entityId":"epoch:13fc6f983d829587433b","metric":"training-compute","value":7.690547e+18,"unit":"FLOP","observationPeriod":"2020-02-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 126*10^6 parameters * 256 tokens per sequences * 512 sequences per batch * 210000 steps = 2.0808991e+19 FLOP\n\nassuming V100 GPU fp16:\n\n31330000000000 FLOP/sec/GPU * 1 GPU * 84 hours * 3600 sec / hour * 0.3 [assumed utilization] = 2.8422576e+18 FLOP\n\nsqrt(2.0808991e+19*2.8422576e+18) = 7.690547e+18 FLOP\n\n___________\nin the Algorithmic progress paper they used estimation of 4.41e+19 FLOP also with low confidence"},{"id":"epoch:a05e0be300e82af6fe94:parameters","entityId":"epoch:a05e0be300e82af6fe94","metric":"parameters","value":220000000.0,"unit":"parameters","observationPeriod":"2020-02-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use the same code base, model architecture, and training task as the base model from [Raffel et al.,2019]. The encoder and decoder each consist of 12 layers, with dmodel = 768. For the attention layers, h = 12 and dk = dv = 64. The FFN layers have hidden size df f = 3072.\"\n\n[Raffel et al.,2019]: \"Specifically, both the encoder and decoder consist of 12 blocks (each block comprising self-attention, optional encoder-decoder attention, and a feed-forward network). The feed-forward networks in each block consist of a dense layer with an output dimensionality of dff = 3072 followed by a\nReLU nonlinearity and another dense layer. The \u201ckey\u201d and \u201cvalue\u201d matrices of all attention mechanisms have an inner dimensionality of dkv = 64 and all attention mechanisms have 12 heads. All other sub-layers and embeddings have a dimensionality of dmodel = 768. In total, this results in a model with about 220 million parameters.\""},{"id":"epoch:a05e0be300e82af6fe94:training-compute","entityId":"epoch:a05e0be300e82af6fe94","metric":"training-compute","value":3.3687317e+19,"unit":"FLOP","observationPeriod":"2020-02-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 220000000 parameters * 50600083456 tokens = 66792110161920000000 FLOP (10^19)\n\n45000000000000 FLOP/GPU/sec * 21.85 hours * 3600 sec / hour * 16 GPUs * 0.3 [assumed utilization] = 16990560000000002000 FLOP (10^19)\n\nsqrt(66792110161920000000*16990560000000002000) = 3.3687317e+19"},{"id":"epoch:4f94997bae01afc2b750:parameters","entityId":"epoch:4f94997bae01afc2b750","metric":"parameters","value":375000000.0,"unit":"parameters","observationPeriod":"2020-02-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://openai.com/blog/image-gpt/"},{"id":"epoch:d18079e5aa70ffa6cc01:parameters","entityId":"epoch:d18079e5aa70ffa6cc01","metric":"parameters","value":17000000000.0,"unit":"parameters","observationPeriod":"2020-02-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d18079e5aa70ffa6cc01:training-compute","entityId":"epoch:d18079e5aa70ffa6cc01","metric":"training-compute","value":1.57e+22,"unit":"FLOP","observationPeriod":"2020-02-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://lair.lighton.ai/akronomicon/\n157 PF-days * 3600 * 24 * 10^15  = 1.35648e+22\n\narchived: https://github.com/lightonai/akronomicon/tree/main/akrodb\n\n6ND=6*17000000000*46400000000=4.7328e+21 (confidence regarding dataset size - likely)\n\nAuthors of \"AI and Memory Wall\" (https://github.com/amirgholami/ai_and_memory_wall) estimated model's training compute as 28,000,000 PFLOP = 2.8*10^22 FLOP"},{"id":"epoch:4f21e61a7b9455d031fd:parameters","entityId":"epoch:4f21e61a7b9455d031fd","metric":"parameters","value":235000000.0,"unit":"parameters","observationPeriod":"2020-02-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:4f21e61a7b9455d031fd:training-compute","entityId":"epoch:4f21e61a7b9455d031fd","metric":"training-compute","value":2.39e+21,"unit":"FLOP","observationPeriod":"2020-02-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32 hours of training\n512 TPU V3s\n0.33 utilization rate\n\n123000000000000 FLOP / chip / sec * 512 TPUs * 32 hours * 3600 sec / hour * 0.33 [assumed utilization] = 2.3940956e+21 FLOP\n\n\"We train all models for 125,000 steps unless otherwise specified\"\n\"All the model updates use a batch size of 4096 \"\n\"We always limit the maximum input length to 512, and randomly generate input sequences shorter than 512 with a probability of 10%.\"\n\n6 FLOP / parameter / token * 235000000 parameters * 512 tokens per sequence * 4096 sequences per batch * 125000 steps =  3.6962304e+20 FLOP\n\nAuthors of \"AI and Memory Wall\" (https://github.com/amirgholami/ai_and_memory_wall) estimated model's training compute as 31,000,000 PFLOP = 3.1*10^22 FLOP"},{"id":"epoch:a4d176f6f3c86d827167:parameters","entityId":"epoch:a4d176f6f3c86d827167","metric":"parameters","value":27900000.0,"unit":"parameters","observationPeriod":"2020-02-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Optical flow model (SOTA) was 27.9M params. There are other, larger models described in this paper, e.g. for language.\n\n\"For the pixel- and patch-based models, total computational\ncomplexity for a forward pass on a 368 \u00d7 496 image is roughly 987 billion FLOPs, and there are\nroughly 27.9 million parameters.\""},{"id":"epoch:da35ddc31def2f2f45ae:parameters","entityId":"epoch:da35ddc31def2f2f45ae","metric":"parameters","value":240000000.0,"unit":"parameters","observationPeriod":"2020-02-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 5\n\n\"For the language model, we followed the same configuration\nas Baevski & Auli (2019). We used 17 decoding layers, each\nlayer with a 1024 hidden size, a 4096 feed-forward hidden\nsize and 8 heads. The adaptive input factor was set to 4.\""},{"id":"epoch:da35ddc31def2f2f45ae:training-compute","entityId":"epoch:da35ddc31def2f2f45ae","metric":"training-compute","value":2.6990346e+19,"unit":"FLOP","observationPeriod":"2020-02-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 240000000 parameters * 286000 steps * 65536 tokens per batch = 2.6990346e+19 FLOP"},{"id":"epoch:d15e5765466034bb6247:parameters","entityId":"epoch:d15e5765466034bb6247","metric":"parameters","value":66000000.0,"unit":"parameters","observationPeriod":"2020-02-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"66M, Table 1"},{"id":"epoch:063419e6526fb101675d:parameters","entityId":"epoch:063419e6526fb101675d","metric":"parameters","value":2600000000.0,"unit":"parameters","observationPeriod":"2020-01-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We present Meena, a multi-turn open-domain chatbot trained end-to-end on data mined and filtered from public domain social media conversations. This 2.6B parameter neural network is simply trained to minimize perplexity of the next token.\""},{"id":"epoch:063419e6526fb101675d:training-compute","entityId":"epoch:063419e6526fb101675d","metric":"training-compute","value":1.12e+23,"unit":"FLOP","observationPeriod":"2020-01-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://arxiv.org/ftp/arxiv/papers/2104/2104.10350.pdf\nTable 4\n\nIn the paper: \"We trained our best model for 30 days on a TPUv3 Pod (2,048 TPU cores) on the Meena dataset containing 40B words (or 61B BPE tokens) [...] by the end of training, the model had traversed the full\ntraining set 164 times (or epochs) and observed a total of about 10T tokens\"\n\nHardware: 30 * 24 * 3600 * (2048/2) * 1.23e14 * 0.3 = 9.794e22\nOps counting: 6 * 10T * 2.6B = 1.56E23\nGeometric mean: sqrt(9.79e22*1.56E23) = 1.24e23, very close to the figure in the link above."},{"id":"epoch:4e7932711eea329d0a16:training-compute","entityId":"epoch:4e7932711eea329d0a16","metric":"training-compute","value":8.16e+21,"unit":"FLOP","observationPeriod":"2020-01-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train 6 generations of models numbered 0 to 5, where\nwe count the baseline model trained with the supervised set\nas the zeroth generation. Each generation is trained ... on 32 Google\nCloud TPU chips for 10 days.\"\n\nThe TPU version is likely v3 given this is a 2020 paper.\n\nwe get 6 * 10 * 24 * 3600 * 32 * 123 tflops * 0.4  (assumed utilization) = 8.16e21"},{"id":"epoch:b0ba69e06f534e9ff457:parameters","entityId":"epoch:b0ba69e06f534e9ff457","metric":"parameters","value":16340840.0,"unit":"parameters","observationPeriod":"2020-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Neural network hyperparameters\" section of https://www.nature.com/articles/s41586-019-1923-7:\n\u201c7 \u00d7 4 Blocks with 256 channels, cycling through dilations 1, 2, 4, 8\u201d\n\u201c48 \u00d7 4 Blocks with 128 channels, cycling through dilations 1, 2, 4, 8\u201d\n\n\"Distogram prediction\" section:\n\"For the final layer, a position-specific bias was used\"\n\nExtended Data Fig.1 (b): \nShows that each block consists of 9 layers:\n(1) Batch norm\n(2) Elu\n(3) Project down (halves number of dimensions)\n(4) Batch norm\n(5) Elu\n(6) 3x3 kernel with dilation\n(7) Batch norm\n(8) Elu\n(9) Project up (doubles number of dimensions)\n\nDilations don't change the number of parameters in each filter\nAssuming that projection layers are convolutional layers with 1x1 kernels\n\nParameter estimate for each layer in a 256 channel block:\n(1) 256*2            = 512\n(2) 0\n(3) 1*1*256*128 = 32768\n(4) 128*2            = 256 \n(5) 0\n(6) 3*3*128*128 = 147456\n(7) 128*2            = 256 \n(8) 0\n(9) 1*1*128*256 + 256 = 33024\nTotal                             = 214272\n\nParameter estimate for each layer in a 128 channel block:\n(1) 128*2            = 256\n(2) 0\n(3) 1*1*128*64   = 8192\n(4) 64*2              = 128\n(5) 0\n(6) 3*3*64*64     = 36864\n(7) 64*2              = 128\n(8) 0\n(9) 1*1*64*128 + 128 = 8320\nTotal                   = 53897\n\nEstimate total network = 7*4*214272 + 48*4*53897 = 5992616 + 10348224\n                                     = 16340840\n                                     ~ 16e6\n\nWithin a factor of 2 of the estimate of 21M parameters stated in: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7305407/\n\n[Previous approximation: 7 * 4 * 256 * 3 * 3 * 256 + 48 * 4 * 128 * 3 * 3 * 128 = 44826624]"},{"id":"epoch:b0ba69e06f534e9ff457:training-compute","entityId":"epoch:b0ba69e06f534e9ff457","metric":"training-compute","value":1e+20,"unit":"FLOP","observationPeriod":"2020-01-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Estimated in the blogpost below\n\nhttps://www.lesswrong.com/posts/wfpdejMWog4vEDLDg/ai-and-compute-trend-isn-t-predictive-of-what-is-happening\n\n\"AlphaFold: they say they trained on GPU and not TPU. Assuming V100 GPU, it's 5 days * 24 hours/day * 3600 sec/hour * 8 V100 GPU * 100*10^12 FLOP/s * 33% actual GPU utilization = 10^20 FLOP.\""},{"id":"epoch:a4d2d80f704bd3e5726d:parameters","entityId":"epoch:a4d2d80f704bd3e5726d","metric":"parameters","value":8300000.000000001,"unit":"parameters","observationPeriod":"2020-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ca14a68c89fd9a3df952:parameters","entityId":"epoch:ca14a68c89fd9a3df952","metric":"parameters","value":928000000.0,"unit":"parameters","observationPeriod":"2019-12-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:acd622488ee12f749e7b:parameters","entityId":"epoch:acd622488ee12f749e7b","metric":"parameters","value":928000000.0,"unit":"parameters","observationPeriod":"2019-12-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"928 million\" stated directly in paper"},{"id":"epoch:738ee90a971e428ad795:training-compute","entityId":"epoch:738ee90a971e428ad795","metric":"training-compute","value":7.8e+20,"unit":"FLOP","observationPeriod":"2019-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Using DD-PPO, we train agents for 2.5 Billion steps of experience with 64 Tesla V100 GPUs in 2.75 days \u2013 180 GPU-days of training\"\n\n125 teraFLOP/s (exact V100 model not specified) * 180 * 24 * 3600 * 0.4 (assumed utilization) = 7.8e20"},{"id":"epoch:5fa959f68138c1ffd76b:parameters","entityId":"epoch:5fa959f68138c1ffd76b","metric":"parameters","value":93000000.0,"unit":"parameters","observationPeriod":"2019-12-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The model had about 93\u2009M (mega/million) free parameters\""},{"id":"epoch:5fa959f68138c1ffd76b:training-compute","entityId":"epoch:5fa959f68138c1ffd76b","metric":"training-compute","value":4.1e+19,"unit":"FLOP","observationPeriod":"2019-12-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3 weeks, 5 NVIDIA Titan GPUs (Assuming NVIDIA Titan V and 30% utilization rate for calculation) with 12 GB memory, "},{"id":"epoch:3127d2533b13a856d92b:parameters","entityId":"epoch:3127d2533b13a856d92b","metric":"parameters","value":159000000.0,"unit":"parameters","observationPeriod":"2019-12-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We define a policy (\u03c0) as a function from the history of observations to a probability distribution over actions, which we parameterize as a recurrent neural network with approximately 159 million parameters (\u03b8).\" pg. 3 of paper\n\nsource: https://docs.google.com/spreadsheets/d/1Kj4Q5WADcDXtUJLIOfGTCE3tGvxNczEMwyy8QtgSkHk/edit#gid=54587040&fvid=1361937389"},{"id":"epoch:3127d2533b13a856d92b:training-compute","entityId":"epoch:3127d2533b13a856d92b","metric":"training-compute","value":6.7e+22,"unit":"FLOP","observationPeriod":"2019-12-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"770\u00b150 PFlops/s\u00b7days of compute\" for the model that played against world champions. They did a single training run that took 10 months.\n\nWhile the model was playing against world champions, they continued training for a few days, so that the resulting model used even more training compute: 820\u00b150 PFlops/s\u00b7days.\n\nFinally, they also trained a Rerun model with 150\u00b15 PFlops/s\u00b7days of compute.\n\nSource: Dota 2 with Large Scale Deep Reinforcement Learning\nhttps://arxiv.org/abs/1912.06680\n\nYou cannot multiply the hardware quantity by training time to get the quantity of GPU-hours! Page 5: \" the number of GPUs (up to 1536 at the peak)\"\n\nFrom this NVIDIA blogpost, it appears they were using P100s:\nhttps://developer.nvidia.com/blog/ai-learns-to-play-dota-2-with-human-precision/#:~:text=AI%20Learns%20to%20Play%20Dota,The%20neural"},{"id":"epoch:066ea7ccabff42090962:parameters","entityId":"epoch:066ea7ccabff42090962","metric":"parameters","value":159000000.0,"unit":"parameters","observationPeriod":"2019-12-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We define a policy (\u03c0) as a function from the history of observations to a probability distribution\nover actions, which we parameterize as a recurrent neural network with approximately 159 million\nparameters (\u03b8).\" pg. 3 of paper\n\nsource: https://docs.google.com/spreadsheets/d/1Kj4Q5WADcDXtUJLIOfGTCE3tGvxNczEMwyy8QtgSkHk/edit#gid=54587040&fvid=1361937389"},{"id":"epoch:066ea7ccabff42090962:training-compute","entityId":"epoch:066ea7ccabff42090962","metric":"training-compute","value":1.3e+22,"unit":"FLOP","observationPeriod":"2019-12-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"THIS CALCULATION IS FOR RERUN\n\n\"Rerun took 2 months and 150 \u00b1 5 PFlops/s\u00b7days of compute (see Figure 4)\"\n\n\n\nsource: https://docs.google.com/spreadsheets/d/1Kj4Q5WADcDXtUJLIOfGTCE3tGvxNczEMwyy8QtgSkHk/edit#gid=54587040&fvid=1361937389"},{"id":"epoch:8f56a29b4c347d484e18:parameters","entityId":"epoch:8f56a29b4c347d484e18","metric":"parameters","value":75000000.0,"unit":"parameters","observationPeriod":"2019-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table VII"},{"id":"epoch:8f56a29b4c347d484e18:training-compute","entityId":"epoch:8f56a29b4c347d484e18","metric":"training-compute","value":2.3175e+18,"unit":"FLOP","observationPeriod":"2019-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 75000000 parameters * 103000000 tokens * 50 epochs [assumption] = 2.3175e+18 FLOP"},{"id":"epoch:6b009687e5049c763486:parameters","entityId":"epoch:6b009687e5049c763486","metric":"parameters","value":32300000.0,"unit":"parameters","observationPeriod":"2019-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table III"},{"id":"epoch:6b009687e5049c763486:training-compute","entityId":"epoch:6b009687e5049c763486","metric":"training-compute","value":1.938e+17,"unit":"FLOP","observationPeriod":"2019-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 32300000 parameters * 2000000 tokens * 500 epochs [assumption] = 1.938e+17 FLOP"},{"id":"epoch:fee0ab3f06b60c4f59d9:parameters","entityId":"epoch:fee0ab3f06b60c4f59d9","metric":"parameters","value":21300000.0,"unit":"parameters","observationPeriod":"2019-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table II"},{"id":"epoch:fee0ab3f06b60c4f59d9:training-compute","entityId":"epoch:fee0ab3f06b60c4f59d9","metric":"training-compute","value":5.8298782e+16,"unit":"FLOP","observationPeriod":"2019-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 21300000 parameters * 912344 tokens * 500 epochs [assumption] = 5.8298782e+16 FLOP"},{"id":"epoch:e127ce4a6dafc4a8b7df:parameters","entityId":"epoch:e127ce4a6dafc4a8b7df","metric":"parameters","value":30000000.0,"unit":"parameters","observationPeriod":"2019-12-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"30M for the generator, 29M for the discriminator"},{"id":"epoch:efda5f71fd1637c40897:parameters","entityId":"epoch:efda5f71fd1637c40897","metric":"parameters","value":2550000.0,"unit":"parameters","observationPeriod":"2019-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:efda5f71fd1637c40897:training-compute","entityId":"epoch:efda5f71fd1637c40897","metric":"training-compute","value":275400000000000.0,"unit":"FLOP","observationPeriod":"2019-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 2550000 parameters * 300 tokens per batch [speculatively, I am not sure if it is sequences per batch or tokens per batch] * 60000 steps = 2.754e+14 FLOP"},{"id":"epoch:bae3659bde27c67df9d6:parameters","entityId":"epoch:bae3659bde27c67df9d6","metric":"parameters","value":141000000.0,"unit":"parameters","observationPeriod":"2019-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2b"},{"id":"epoch:bae3659bde27c67df9d6:training-compute","entityId":"epoch:bae3659bde27c67df9d6","metric":"training-compute","value":1.74276e+18,"unit":"FLOP","observationPeriod":"2019-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 141000000 parameters * 103000000 tokens * 20 epochs [assumption based on number of epochs for LSTMs] =  1.74276e+18 FLOP\n\n_______________\nolder estimation:  6.2 \u00d7 10^18 (no explantion how it was calculated)"},{"id":"epoch:83e71e8252542e145856:parameters","entityId":"epoch:83e71e8252542e145856","metric":"parameters","value":107370000.0,"unit":"parameters","observationPeriod":"2019-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b562b8aeeb7311296410:parameters","entityId":"epoch:b562b8aeeb7311296410","metric":"parameters","value":48690000.0,"unit":"parameters","observationPeriod":"2019-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1a"},{"id":"epoch:0bd6ea4b7013a6391e83:parameters","entityId":"epoch:0bd6ea4b7013a6391e83","metric":"parameters","value":20000000.0,"unit":"parameters","observationPeriod":"2019-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b8e3b5c35192fa4fcd13:parameters","entityId":"epoch:b8e3b5c35192fa4fcd13","metric":"parameters","value":30100000.0,"unit":"parameters","observationPeriod":"2019-11-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"30.1M"},{"id":"epoch:b8e3b5c35192fa4fcd13:training-compute","entityId":"epoch:b8e3b5c35192fa4fcd13","metric":"training-compute","value":7.1712e+18,"unit":"FLOP","observationPeriod":"2019-11-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"125000000000000*53.12*3600*0.3 = 7.1712e+18"},{"id":"epoch:ad2f737912ac18768387:parameters","entityId":"epoch:ad2f737912ac18768387","metric":"parameters","value":36864000.0,"unit":"parameters","observationPeriod":"2019-11-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Both the representation and dynamics function use the same architecture asAlphaZero, but with 16 instead of20 residual blocks [15]. We use 3x3 kernels and 256 hidden planes for each convolution.\n\nPrevious downsampling:\n\u2022  1 convolution with stride 2 and 128 output planes, output resolution 48x48.\u2022  2 residual blocks with 128 planes\u2022  1 convolution with stride 2 and 256 output planes, output resolution 24x24.\u2022  3 residual blocks with 256 planes.\u2022  Average pooling with stride 2, output resolution 12x12.\u2022  3 residual blocks with 256 planes.\u2022  Average pooling with stride 2, output resolution 6x6."},{"id":"epoch:ad2f737912ac18768387:training-compute","entityId":"epoch:ad2f737912ac18768387","metric":"training-compute","value":4.8e+19,"unit":"FLOP","observationPeriod":"2019-11-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"third-generation Google Cloud TPU\n(For each board game, we used 16 TPUs for training and 1000 TPUs for self-play)\nFor each game in Atari, we used 8 TPUs for training and 32 TPUs for self-play\nTraining for 12 hours (for Atari)\nData from Parameter, Compute and Data Trends in Machine Learning\nGoogle v3 TPU: 1.23E+14 FLOP/s (although with the caveat that it might be not applicable)\nUtilization rate \nIn LaMDA: Language Models for Dialog Applications, they report for TPU V3: 56.5%\nCalculations for Atari:\n12 hours \u2192 43200 seconds\n(8 TPUs for training) * (1.23*10^14 FLOP/s) * (43.2 *10^3 s) * (0.565 utilization rate) = 2.4017472 * 10^19 FLOP\nTraining time missing for boardgames\nAssumption also 12 hours \nAlso: 2.4017472 * 10^19 FLOP\nTotal cost \u2248 4.8 * 10^19 FLOP"},{"id":"epoch:bf741d44f3368f6b57f9:parameters","entityId":"epoch:bf741d44f3368f6b57f9","metric":"parameters","value":296000000.0,"unit":"parameters","observationPeriod":"2019-11-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:67e61a442db13808e5ab:parameters","entityId":"epoch:67e61a442db13808e5ab","metric":"parameters","value":375000000.0,"unit":"parameters","observationPeriod":"2019-11-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://openai.com/blog/image-gpt/#rfref53"},{"id":"epoch:88884b8d26e63a2e9548:training-compute","entityId":"epoch:88884b8d26e63a2e9548","metric":"training-compute","value":1.0202112e+20,"unit":"FLOP","observationPeriod":"2019-11-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"123000000000000 FLOP / second/ GPU * 64 GPUs * 12 hours * 3600 sec/hour * 0.3 [assumed utilization] = 1.0202112e+20 FLOP"},{"id":"epoch:0e230d9dd449b14eef9f:parameters","entityId":"epoch:0e230d9dd449b14eef9f","metric":"parameters","value":480000000.0,"unit":"parameters","observationPeriod":"2019-11-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0e230d9dd449b14eef9f:training-compute","entityId":"epoch:0e230d9dd449b14eef9f","metric":"training-compute","value":2.612e+22,"unit":"FLOP","observationPeriod":"2019-11-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our largest model, EfficientNet-L2, needs to be trained for 6 days on a Cloud TPU v3 Pod, which has 2048 cores, if the unlabeled batch size is 14x the labeled batch size\"\nTPU v3 gets 1.23e14 FLOP/s per chip, with 2 cores per chip\n\n1024 * 1.23e14 * 6 * 24 * 3600 * 0.4 = 2.612e22"},{"id":"epoch:dc548c774db10f79545c:parameters","entityId":"epoch:dc548c774db10f79545c","metric":"parameters","value":209000000.0,"unit":"parameters","observationPeriod":"2019-11-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"209M\n\"All of our experiments use the same hyperparameters as Baevski and Auli\u2019s original model.\""},{"id":"epoch:dc548c774db10f79545c:training-compute","entityId":"epoch:dc548c774db10f79545c","metric":"training-compute","value":2.3504093e+19,"unit":"FLOP","observationPeriod":"2019-11-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 209000000 parameters * 286000 steps * 65536 tokens per batch [same as Baevski and Auli (2019) = 2.3504093e+19 FLOP\n\n__________\n\nin the Algorithmic progress paper they assumed 180 epochs (same as Baevski and Auli 2019 Transformer, but that one was trained on WT 103 not the book corpus) -> training compute was estimated to be 1.58E+20 FLOP"},{"id":"epoch:63fc584f4f80ab678159:parameters","entityId":"epoch:63fc584f4f80ab678159","metric":"parameters","value":335000000.0,"unit":"parameters","observationPeriod":"2019-11-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"CamemBERT Large, Table 4"},{"id":"epoch:63fc584f4f80ab678159:training-compute","entityId":"epoch:63fc584f4f80ab678159","metric":"training-compute","value":8.3e+20,"unit":"FLOP","observationPeriod":"2019-11-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Unless otherwise specified, our models use the BASE architecture, and are pretrained for 100k backpropagation steps on 256 Nvidia V100 GPUs (32GB each) for a day\"\n\n256 V100-days\n\n256 * 125 teraflops * 24 * 3600 * 0.3 (assumed utilization)\n= 8.3e20\n\n\n\"Following (Liu et al., 2019), we optimize the model using Adam (Kingma and Ba, 2014) (\u03b21 = 0.9, \u03b22 = 0.98) for 100k steps with large batch sizes of 8192 sequences, each sequence containing at most 512 tokens\"\n\nUsing compute = 6*N*D, that's 6 * (100k * 8192 * 512) * 335M= 8.43e20"},{"id":"epoch:1a49d518e48fb5b46925:parameters","entityId":"epoch:1a49d518e48fb5b46925","metric":"parameters","value":29500000.0,"unit":"parameters","observationPeriod":"2019-11-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from Table 1"},{"id":"epoch:1a49d518e48fb5b46925:training-compute","entityId":"epoch:1a49d518e48fb5b46925","metric":"training-compute","value":6.75e+17,"unit":"FLOP","observationPeriod":"2019-11-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(15e9) * (300) * (50000) * 3 = 6.75e+17\n(inference compute) * (epochs) * (dataset size) * (constant to account for backpropagation)\n \nepochs from appendix B table 2\ninference compute from table 1"},{"id":"epoch:671249e8e678187f3619:parameters","entityId":"epoch:671249e8e678187f3619","metric":"parameters","value":550000000.0,"unit":"parameters","observationPeriod":"2019-11-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The number of parameters in the model is specified as \"550M params\" for XLM-R."},{"id":"epoch:671249e8e678187f3619:training-compute","entityId":"epoch:671249e8e678187f3619","metric":"training-compute","value":2.076e+22,"unit":"FLOP","observationPeriod":"2019-11-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use the multilingual MLM loss and train our XLM-R model for\n1.5 Million updates on five-hundred 32GB Nvidia\nV100 GPUs with a batch size of 8192. \"\n\n6ND:\nSequence length was probably 512, based on follow up paper (XLM-R XXL)\n6 * 550e6 * 1.5e6 * 8192 * 512 = 2.076e22\n"},{"id":"epoch:aeaf33e79c13021c5f7f:parameters","entityId":"epoch:aeaf33e79c13021c5f7f","metric":"parameters","value":247000000.00000003,"unit":"parameters","observationPeriod":"2019-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we take the exact architecture and optimization described by Baevski & Auli (2019) and use it to create a kNN-LM for inference. This model consists of 16 layers, each with 16 self-attention heads, 1024 dimensional hidden states, and 4096 dimensional feedforward layers, amounting to 247M trainable parameters.\""},{"id":"epoch:aeaf33e79c13021c5f7f:training-compute","entityId":"epoch:aeaf33e79c13021c5f7f","metric":"training-compute","value":3.05292e+19,"unit":"FLOP","observationPeriod":"2019-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 247*10^6 parameters * 103000000 tokens * 200 epochs = 3.05292e+19 FLOP\n\n\n__________\nfor the Algorithmic progress paper 7.3 \u00d7 10^18 FLOP was estimated similar to supposedly base model (transformer) "},{"id":"epoch:1c2915c799fcef853d6d:parameters","entityId":"epoch:1c2915c799fcef853d6d","metric":"parameters","value":139000000.0,"unit":"parameters","observationPeriod":"2019-10-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"AlphaStar has 139 million weights, but only 55 million weights are required during inference."},{"id":"epoch:1c2915c799fcef853d6d:training-compute","entityId":"epoch:1c2915c799fcef853d6d","metric":"training-compute","value":1.0773400001e+23,"unit":"FLOP","observationPeriod":"2019-10-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(Estimate by James Sanders, checked by Robi Rahman)\nFig 6 indicates that the learner uses 16 TPU \"devices\" which are 128 TPU cores total, which matches 4 TPUs per device, and 2 cores per TPU. Fig 6 indicates that 64 TPUs are used for training, and 64 are used for inference. (128 TPUs)*(12 agents)*(44 days)*(123 TFLOPS)*(0.3 utilization) = 2.15e23 FLOP (total), of which 1.07e23 FLOP are for training."},{"id":"epoch:def1ba2c0fbac3c7206b:parameters","entityId":"epoch:def1ba2c0fbac3c7206b","metric":"parameters","value":406291456.0,"unit":"parameters","observationPeriod":"2019-10-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In total, BART contains roughly 10% more parameters than the equivalently sized BERT model.\"\n\nI counted the parameters in the huggingface model\nhttps://huggingface.co/facebook/bart-large/tree/main\n\nfrom transformers import AutoTokenizer, AutoModel\ntokenizer = AutoTokenizer.from_pretrained(\"facebook/bart-large\")\nmodel = AutoModel.from_pretrained(\"facebook/bart-large\")\nsum(p.numel() for p in model.parameters() if p.requires_grad)"},{"id":"epoch:48753f7176cd2bde9713:parameters","entityId":"epoch:48753f7176cd2bde9713","metric":"parameters","value":23000000.0,"unit":"parameters","observationPeriod":"2019-10-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f12d15a009cc5690cbec:parameters","entityId":"epoch:f12d15a009cc5690cbec","metric":"parameters","value":2800000000.0,"unit":"parameters","observationPeriod":"2019-10-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"page 37, 3B and 11B. \"To further explore what kind of performance is possible when using larger models, we consider two additional variants. In both cases, we use d_model = 1024, a 24 layer encoder and decoder, and dkv = 128. For the \u201c3B\u201d variant, we use dff = 16,384 with 32-headed attention, which results in around 2.8 billion parameters; for \u201c11B\u201d we use dff = 65,536 with 128-headed attention producing a model with about 11 billion parameters\""},{"id":"epoch:f12d15a009cc5690cbec:training-compute","entityId":"epoch:f12d15a009cc5690cbec","metric":"training-compute","value":9.0000000001e+21,"unit":"FLOP","observationPeriod":"2019-10-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Akronomicon states 1.04e+22 FLOP. Archived source: https://github.com/lightonai/akronomicon/tree/main/akrodb\nHowever, this seems dubiously high.\n\n\"We pre-train each model for 2^19 = 524,288 steps on C4 before fine-tuning.\"\n\"In total, this batch size and number of steps corresponds to pre-training on 2^35 \u2248 34B tokens.\"\n\"To compare these mixing strategies on equal footing with our baseline pre-train-then-fine-tune results, we train multi-task models for the same total number of steps: 2^19 + 2^18 = 786,432\"\nUsing the 6DN approximation gives: 6 FLOP/token/param * 2^35 pretrain tokens * (1+1/2 finetune tokens per pretrain token) * 1 iteration of training data* 2.8 billion parameters = 8.659e20 FLOP\nhttps://www.wolframalpha.com/input?i=6+*+2%5E35+*+2.8+billion+*+1.5\n\nupdate: 9.0E+21 per FLAN paper from Google \nhttps://arxiv.org/pdf/2210.11416.pdf"},{"id":"epoch:53aafc0472ac0c16be86:parameters","entityId":"epoch:53aafc0472ac0c16be86","metric":"parameters","value":11000000000.0,"unit":"parameters","observationPeriod":"2019-10-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The full 11-billion parameter model"},{"id":"epoch:53aafc0472ac0c16be86:training-compute","entityId":"epoch:53aafc0472ac0c16be86","metric":"training-compute","value":3.3e+22,"unit":"FLOP","observationPeriod":"2019-10-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://arxiv.org/ftp/arxiv/papers/2104/2104.10350.pdf\nTable 4, 4.05e22\n\nupdate: 3.3e22 per FLAN paper from Google \nhttps://arxiv.org/pdf/2210.11416.pdf\n\n6ND rule suggests somewhat more FLOPs:\n6 * 1T * 11B = 6.6e22"},{"id":"epoch:717c5c7253eaf9cdcfef:parameters","entityId":"epoch:717c5c7253eaf9cdcfef","metric":"parameters","value":48000000.0,"unit":"parameters","observationPeriod":"2019-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Based on the details in the paper, here is the calculation for the number of parameters in the LSTM(large)+Sememe+cell model:\n\nThe model has 2 LSTM layers, each with 1500 hidden units.\n\nIt uses the \"+cell\" method to incorporate sememes, which introduces an additional sememe-LSTM cell in each LSTM unit.\n\nFor each original 1500 unit LSTM layer:\n\nInput size is 300 (embedding dimension)\nHidden size is 1500\n4 weight matrices of size 300 x 1500 = 450,000 params each\n4 recurrent weight matrices of size 1500 x 1500 = 2,250,000 params each\n4 bias vectors of size 1500 = 6,000 params each\nTotal = 450,000 x 4 + 2,250,000 x 4 + 6,000 x 4 = 12,006,000\nFor each sememe-LSTM cell:\n\nInput size is 2000 (number of sememes)\nHidden size is 1500\nSame weight and bias computations as above\nTotal = 12,006,000\nSince there are 2 LSTM layers each with a sememe-LSTM cell:\n\nTotal params = 2 * (12,006,000 + 12,006,000)\nTotal = 48,024,000\nTherefore, the total number of parameters for the LSTM(large)+Sememe+cell model is approximately 48 million."},{"id":"epoch:717c5c7253eaf9cdcfef:training-compute","entityId":"epoch:717c5c7253eaf9cdcfef","metric":"training-compute","value":2.304e+16,"unit":"FLOP","observationPeriod":"2019-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 48000000 parameters * 2000000 tokens * 40 epochs = 2.304e+16 FLOP"},{"id":"epoch:4c70e322de645f779f8e:parameters","entityId":"epoch:4c70e322de645f779f8e","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2019-10-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\" the \u201clarge\u201d vanilla LSTM has much more parameters\nthan \u201cmedium\u201d LSTM+cell (76M vs. 24M), it is still outper-\nformed by the latter.\""},{"id":"epoch:b32ebc483e1a9bdb709e:parameters","entityId":"epoch:b32ebc483e1a9bdb709e","metric":"parameters","value":65000000.0,"unit":"parameters","observationPeriod":"2019-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Transformer-base has 65M parameters (Table 3 from https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf)"},{"id":"epoch:b32ebc483e1a9bdb709e:training-compute","entityId":"epoch:b32ebc483e1a9bdb709e","metric":"training-compute","value":2.7570535e+18,"unit":"FLOP","observationPeriod":"2019-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP/parameter/token * 65000000 parameters * 7500000000 tokens = 2.925e18 FLOP\n\n125000000000000 FLOP/GPU/sec * 19.25 hours * 3600 sec / hour * 1 GPUs * 0.3 [assumed utilization] = 2.59875e18 FLOP\n\nsqrt(2.925e18 * 2.59875e18) = 2.7570535e+18 FLOP"},{"id":"epoch:b96bef798a6176aa0438:parameters","entityId":"epoch:b96bef798a6176aa0438","metric":"parameters","value":27769565.0,"unit":"parameters","observationPeriod":"2019-10-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 13 on pg. 44 of the Cube paper, saved in \"RL papers\" folder. Sum of all the trainable parameters (dominated by the value and policy networks).\n\nsource: https://docs.google.com/spreadsheets/d/1Kj4Q5WADcDXtUJLIOfGTCE3tGvxNczEMwyy8QtgSkHk/edit#gid=54587040&fvid=1361937389"},{"id":"epoch:b96bef798a6176aa0438:training-compute","entityId":"epoch:b96bef798a6176aa0438","metric":"training-compute","value":8.54e+20,"unit":"FLOP","observationPeriod":"2019-10-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://docs.google.com/spreadsheets/d/1Kj4Q5WADcDXtUJLIOfGTCE3tGvxNczEMwyy8QtgSkHk/edit#gid=54587040&fvid=1361937389"},{"id":"epoch:840a8ec5fb265ccfcbf6:parameters","entityId":"epoch:840a8ec5fb265ccfcbf6","metric":"parameters","value":44100000.0,"unit":"parameters","observationPeriod":"2019-10-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0b17379045f74ec3a840:parameters","entityId":"epoch:0b17379045f74ec3a840","metric":"parameters","value":50000000000.0,"unit":"parameters","observationPeriod":"2019-10-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(sparse architecture)\n\n\"By modifying the Transformer architecture through the substitution of the vanilla feed-forward layers with sparsely-gated mixture of experts, we drastically scale up the model capacity, allowing us to successfully train and pass 50 billion parameters, which further improved translation quality across the board.\""},{"id":"epoch:c3226e251dcdc474d064:parameters","entityId":"epoch:c3226e251dcdc474d064","metric":"parameters","value":5400000.0,"unit":"parameters","observationPeriod":"2019-10-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 3: multiadds for AlphaX-1 579M, parameters 5.4M"},{"id":"epoch:c3226e251dcdc474d064:training-compute","entityId":"epoch:c3226e251dcdc474d064","metric":"training-compute","value":8.89344e+17,"unit":"FLOP","observationPeriod":"2019-10-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\" Our models for ImageNet use polynomial learning rate\nschedule, starting with 0.05 and decay through 200 epochs.\"\n\n1280000 images * 200 epochs *3 forward-backward adjustment * 1158000000 forward FLOP =889344000000000000"},{"id":"epoch:d715f0c0642bc24b9700:parameters","entityId":"epoch:d715f0c0642bc24b9700","metric":"parameters","value":66000000.0,"unit":"parameters","observationPeriod":"2019-10-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 3"},{"id":"epoch:d715f0c0642bc24b9700:training-compute","entityId":"epoch:d715f0c0642bc24b9700","metric":"training-compute","value":1.24416e+19,"unit":"FLOP","observationPeriod":"2019-10-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 3: DistilBERT was trained on 8 16GB V100 GPUs for approximately 90 hours.\n\n1.6e13*8*60**2*90*0.3 = 1.2e19"},{"id":"epoch:4ebbd7cb62bb5d3d02e7:parameters","entityId":"epoch:4ebbd7cb62bb5d3d02e7","metric":"parameters","value":18000000.0,"unit":"parameters","observationPeriod":"2019-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 3.2 of paper"},{"id":"epoch:dfc344e8f467dccc592f:parameters","entityId":"epoch:dfc344e8f467dccc592f","metric":"parameters","value":423000000.00000006,"unit":"parameters","observationPeriod":"2019-09-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:05a88684c49f7589e378:parameters","entityId":"epoch:05a88684c49f7589e378","metric":"parameters","value":256999999.99999997,"unit":"parameters","observationPeriod":"2019-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0c0ca030c044e2148250:parameters","entityId":"epoch:0c0ca030c044e2148250","metric":"parameters","value":3900000000.0,"unit":"parameters","observationPeriod":"2019-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 4"},{"id":"epoch:0c0ca030c044e2148250:training-compute","entityId":"epoch:0c0ca030c044e2148250","metric":"training-compute","value":2.2e+22,"unit":"FLOP","observationPeriod":"2019-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"A third-party source: https://lair.lighton.ai/akronomicon/ claims 5.7e22\n\nThe authors report experimenting on 1 V100 GPU and achieving throughput of 39 TFLOPS which is 30% of the peak throughput. Therefore the GPU has a peak throughput of 130 TFLOPS so it is specifically the NVIDIA V100S PCIe.\nhttps://images.nvidia.com/content/technologies/volta/pdf/volta-v100-datasheet-update-us-1165301-r5.pdf\n\nParam-based calculation:\n6ND = 6*3.9e9*(2e6+1e4)*1024*512 = 2.5e22 FLOP\n\n1024 is the batch size, 512 is the sequence length (not explicitly stated but they say non-specified hyperparameters follow cited papers).\n\nTime-based calculation:\nThe 8.3B GPT-like arch took 2.1 days per epoch on 512 GPUs, batch size 512. An epoch was 68.5k iterations with sequence length 1024.\n\nHalving the model size should ~halve the iteration time.\nDoubling the batch size should ~double the iteration time.\nHalving the sequence length should ~quarter the iteration time (quadratic scaling).\n\nHence 3.1e-5 days/iteration * 2 * 1/2 * 1/4 = 7.8e-6 days/iteration.\n\n2e6 iterations => seems like 15.6 days training.\n\nOn 512 GPUs they achieve a peak throughput of 15.1 PFLOPS.\nC=15.1 PFLOPS * 58 days = 2.0e22 FLOP.\n\nIf we disregard the Akronomicon estimate and aggregate our two, geometric mean is 2.2e22 FLOP"},{"id":"epoch:4a5307d8fe2144a0105b:parameters","entityId":"epoch:4a5307d8fe2144a0105b","metric":"parameters","value":8300000000.0,"unit":"parameters","observationPeriod":"2019-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Source: https://lair.lighton.ai/akronomicon/\n\nArchived source: https://web.archive.org/web/20211220142906/https://lair.lighton.ai/akronomicon/\n\nData also available on GitHub: https://github.com/lightonai/akronomicon/blob/main/akrodb/NVIDIA/Megatron-LM.json"},{"id":"epoch:4a5307d8fe2144a0105b:training-compute","entityId":"epoch:4a5307d8fe2144a0105b","metric":"training-compute","value":9.1e+21,"unit":"FLOP","observationPeriod":"2019-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://lair.lighton.ai/akronomicon/\n\narchived: https://github.com/lightonai/akronomicon/tree/main/akrodb\n\nother estimates:\n\n8.3B is a GPT-2-based model (Table 2). \"For GPT-2 models, all training is performed with sequences of 1024 subword units at a batch size of 512 for 300k iterations\" \n\nI interpret the above as 1024*512*300k = 157B training tokens \n\n6 * 157 billion * 8.3 billion  = 7.8e21\n\nAlso, their training setup achieved 15.1 petaFLOPS or 1.5e16 FLOPS.\n(512 V100s is 512 * 125 teraflops = 64 petaFLOPS so they had ~25% utilization)\n2.1 days per epoch, ~4.4 epochs\n2.1 * 4.4 * 24 * 3600 * 1.5e16 = 1.197e22\n\nThese are both close to the akronomicon estimate\n\nAuthors of \"AI and Memory Wall\" (https://github.com/amirgholami/ai_and_memory_wall) estimated model's training compute as 8,100,000 PFLOP = 8.1*10^21 FLOP"},{"id":"epoch:dd5e645fefad36a3bcab:parameters","entityId":"epoch:dd5e645fefad36a3bcab","metric":"parameters","value":1600000.0,"unit":"parameters","observationPeriod":"2019-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The default model, which uses a batch size of 64,000 and 1.6 million parameters,...\" pg. 7 of the Hide and Seek paper, stored in \"RL papers\"\n\nsource: https://docs.google.com/spreadsheets/d/1Kj4Q5WADcDXtUJLIOfGTCE3tGvxNczEMwyy8QtgSkHk/edit#gid=54587040&fvid=1361937389"},{"id":"epoch:dd5e645fefad36a3bcab:training-compute","entityId":"epoch:dd5e645fefad36a3bcab","metric":"training-compute","value":1.15e+18,"unit":"FLOP","observationPeriod":"2019-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.6e6 * 2 * 120e9 * 3 ~= 1.15e18 FLOP. We assume the single convolution on the lidar input is negligible, and the rest of the model (which consists of MLPs, self-attention and LSTM) has roughly 1 multiply-add per parameter. The following source has a similar estimate but does not adjust for the full number of episodes: https://docs.google.com/spreadsheets/d/1Kj4Q5WADcDXtUJLIOfGTCE3tGvxNczEMwyy8QtgSkHk/edit#gid=54587040&fvid=1361937389"},{"id":"epoch:5c2da27bd26efe39259f:parameters","entityId":"epoch:5c2da27bd26efe39259f","metric":"parameters","value":2500000000.0,"unit":"parameters","observationPeriod":"2019-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5c2da27bd26efe39259f:training-compute","entityId":"epoch:5c2da27bd26efe39259f","metric":"training-compute","value":2.359e+21,"unit":"FLOP","observationPeriod":"2019-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"300k steps, batch size 512, sequence length 1024\n\nTraining tokens: 300000*512*1024=157286400000\n\nEpochs: 157286400000/46400000000=3.4\n\nTraining FLOP: 6*2500000000*157286400000=2.359296e+21"},{"id":"epoch:4fb19eaf4cde40372fc6:parameters","entityId":"epoch:4fb19eaf4cde40372fc6","metric":"parameters","value":355000000.0,"unit":"parameters","observationPeriod":"2019-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"355M"},{"id":"epoch:4fb19eaf4cde40372fc6:training-compute","entityId":"epoch:4fb19eaf4cde40372fc6","metric":"training-compute","value":3.35e+20,"unit":"FLOP","observationPeriod":"2019-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"355M is a GPT-2-based model (Table 2).\n\n\"For GPT-2 models, all training is performed with sequences\nof 1024 subword units at a batch size of 512 for 300k iterations\" \n\nI interpret the above as 1024*512*300k = 157B training tokens (or things that require a forward+backward pass). \n\n6 * 1024*512*300,000 * 355,000,000  = 3.35e20"},{"id":"epoch:ccc366489272da74d77e:parameters","entityId":"epoch:ccc366489272da74d77e","metric":"parameters","value":1200000000.0,"unit":"parameters","observationPeriod":"2019-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1 in https://arxiv.org/pdf/1909.08053"},{"id":"epoch:ccc366489272da74d77e:training-compute","entityId":"epoch:ccc366489272da74d77e","metric":"training-compute","value":1.1304e+21,"unit":"FLOP","observationPeriod":"2019-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"300,000*512*1,024 = 1.57e11 tokens\n6*1.57e11*1.2e9 = 1.13e21 FLOPs"},{"id":"epoch:13265b11a0b4c2dfc998:parameters","entityId":"epoch:13265b11a0b4c2dfc998","metric":"parameters","value":38000000.0,"unit":"parameters","observationPeriod":"2019-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:13265b11a0b4c2dfc998:training-compute","entityId":"epoch:13265b11a0b4c2dfc998","metric":"training-compute","value":1.94e+19,"unit":"FLOP","observationPeriod":"2019-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3-5 days of training (say, 4.5), 50 teraFLOP/second at 50% utilization rate (reported) = 1.94E19"},{"id":"epoch:e26525fe6325fa26d147:parameters","entityId":"epoch:e26525fe6325fa26d147","metric":"parameters","value":35000000.0,"unit":"parameters","observationPeriod":"2019-09-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:89c875b9c03d998ba878:parameters","entityId":"epoch:89c875b9c03d998ba878","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2019-09-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0acbf58f6f307cfb612c:parameters","entityId":"epoch:0acbf58f6f307cfb612c","metric":"parameters","value":28303800.0,"unit":"parameters","observationPeriod":"2019-09-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Python code:  \n# Given LSTM parameters\nemb_sz = 400  # embedding size, typically equal to the input size for the first layer\nnh = 1150     # number of hidden units\nnl = 3        # number of layers\n\n# The formula for a single LSTM layer parameters is:\n# P = 4 * ((input_dim + hidden_dim) * hidden_dim + hidden_dim)\n\n# First layer parameters (input_dim is the embedding size)\nfirst_layer_params = 4 * ((emb_sz + nh) * nh + nh)\n\n# For subsequent layers, input_dim is equal to hidden_dim (nh)\nsubsequent_layer_params = 4 * ((nh + nh) * nh + nh)\n\n# Total parameters for all layers\ntotal_params = first_layer_params + (nl - 1) * subsequent_layer_params\n\nprint(total_params)"},{"id":"epoch:0acbf58f6f307cfb612c:training-compute","entityId":"epoch:0acbf58f6f307cfb612c","metric":"training-compute","value":6.37e+17,"unit":"FLOP","observationPeriod":"2019-09-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Pretraining:\nTable 7 gives max of 499k sequences each at (seemingly) L=1024:\n499k * 1024 * 28.3M * 6 = 8.7e16\n\nFinetuning:\nLargest downstream task has 104940 sequences (Table 5), each sequence has L=1024 residues, 28.3M parameters, and 30 epochs.\n105k * 1024 * 30 * 28.3 * 6 = 5.5e17."},{"id":"epoch:e9e408fab15cc4da10dc:parameters","entityId":"epoch:e9e408fab15cc4da10dc","metric":"parameters","value":110000000.00000001,"unit":"parameters","observationPeriod":"2019-09-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"110M Table 3"},{"id":"epoch:e9e408fab15cc4da10dc:training-compute","entityId":"epoch:e9e408fab15cc4da10dc","metric":"training-compute","value":8.1576e+17,"unit":"FLOP","observationPeriod":"2019-09-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 110000000 parameters * 103000000 tokens * 12 epochs = 8.1576e+17 FLOP"},{"id":"epoch:092d033b319b4bea4168:parameters","entityId":"epoch:092d033b319b4bea4168","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2019-09-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b15896e549be2077702b:parameters","entityId":"epoch:b15896e549be2077702b","metric":"parameters","value":180000000.0,"unit":"parameters","observationPeriod":"2019-09-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:70798de97710fca59847:parameters","entityId":"epoch:70798de97710fca59847","metric":"parameters","value":67000000.0,"unit":"parameters","observationPeriod":"2019-08-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c1c0b68ee7eb91c87796:parameters","entityId":"epoch:c1c0b68ee7eb91c87796","metric":"parameters","value":27000000.0,"unit":"parameters","observationPeriod":"2019-08-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6c971294e1a803389eb8:parameters","entityId":"epoch:6c971294e1a803389eb8","metric":"parameters","value":20000000.0,"unit":"parameters","observationPeriod":"2019-08-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e4009499ca268e62da85:training-compute","entityId":"epoch:e4009499ca268e62da85","metric":"training-compute","value":3.8047968e+19,"unit":"FLOP","observationPeriod":"2019-08-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32620000000000 FLOP / GPU / sec [Nvidia Titan RTX, FP16 assumed] * 1 GPU * 1080 hours [see training time notes] * 3600 sec / hour * 0.3 [assumed utilization] = 3.8047968e+19 FLOP"},{"id":"epoch:1379a9c379ec4e50799e:parameters","entityId":"epoch:1379a9c379ec4e50799e","metric":"parameters","value":247000000.00000003,"unit":"parameters","observationPeriod":"2019-08-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f8190cfe6d7cc08af523:parameters","entityId":"epoch:f8190cfe6d7cc08af523","metric":"parameters","value":23000000.0,"unit":"parameters","observationPeriod":"2019-07-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e1a5d13567fce0807b20:parameters","entityId":"epoch:e1a5d13567fce0807b20","metric":"parameters","value":7190000.0,"unit":"parameters","observationPeriod":"2019-07-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:bc2a58e66f63374eb2ef:parameters","entityId":"epoch:bc2a58e66f63374eb2ef","metric":"parameters","value":225999999.99999997,"unit":"parameters","observationPeriod":"2019-07-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"226M"},{"id":"epoch:4632e385fdb3f400daae:parameters","entityId":"epoch:4632e385fdb3f400daae","metric":"parameters","value":175000000.0,"unit":"parameters","observationPeriod":"2019-07-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"175M"},{"id":"epoch:02b7953ad4d097c244ec:parameters","entityId":"epoch:02b7953ad4d097c244ec","metric":"parameters","value":15800000.0,"unit":"parameters","observationPeriod":"2019-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://colab.research.google.com/drive/1afDpFvO3Dtry2Mdg8C1aUBV_sMF7LGb7?usp=sharing"},{"id":"epoch:02b7953ad4d097c244ec:training-compute","entityId":"epoch:02b7953ad4d097c244ec","metric":"training-compute","value":8649021100000000.0,"unit":"FLOP","observationPeriod":"2019-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 15800000 parameters * 912344 tokens * 100 epochs = 8.6490211e+15 FLOP"},{"id":"epoch:55ffec07f00c109b2d8a:training-compute","entityId":"epoch:55ffec07f00c109b2d8a","metric":"training-compute","value":6.6e+16,"unit":"FLOP","observationPeriod":"2019-07-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Trained in 8 days on a 64 core CPU\nhttps://ai.facebook.com/blog/pluribus-first-ai-to-beat-pros-in-6-player-poker/\n\n\"We trained the blueprint strategy for Pluribus in eight days on a 64-core server and required less than 512 GB of RAM. No GPUs were used. At typical cloud computing instance rates, it would cost less than $150 to train.\"\n\nGuess: trained on i7 Intel CPU, approx 5e9 FLOP/s for each core.\n\n https://epoch.ai/blog/estimating-training-compute\n8 days, 64 cores, 5e9 FLOP/s, 30% utilization"},{"id":"epoch:b19732d064e172525b7f:parameters","entityId":"epoch:b19732d064e172525b7f","metric":"parameters","value":86000000.0,"unit":"parameters","observationPeriod":"2019-07-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://openai.com/blog/image-gpt/#rfref53"},{"id":"epoch:d0318d9db2501ae80a7f:parameters","entityId":"epoch:d0318d9db2501ae80a7f","metric":"parameters","value":133000000.0,"unit":"parameters","observationPeriod":"2019-07-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"133M (Table 3)\n\nFor word level language modeling, we use a model with d = 512 and 36 layers, each with 8 heads\nand 2048 persistent vectors."},{"id":"epoch:d0318d9db2501ae80a7f:training-compute","entityId":"epoch:d0318d9db2501ae80a7f","metric":"training-compute","value":1.1657733e+19,"unit":"FLOP","observationPeriod":"2019-07-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"31330000000000 FLOP / sec / GPU [fp16] * 64 GPUs * 24 hours * 3600 sec/hour * 0.3 [assumed utilization] = 5.197271e+19 FLOP\n\n6 FLOP / token / parameter * 200000 steps * 64 sequences per batch * 256 tokens per sequence * 133000000 parameters = 2.6148864e+18 FLOP\n\nsqrt(5.197271e+19*2.6148864e+18) = 1.1657733e+19 FLOP\n\nlikely confidence: i am uncertain about number of steps and v100 precision \n\n_________\nThe Algorithmic Progress paper estimation was 4.6 \u00d7 10^19 FLOP"},{"id":"epoch:675f2555455c55eec8d0:parameters","entityId":"epoch:675f2555455c55eec8d0","metric":"parameters","value":355000000.0,"unit":"parameters","observationPeriod":"2019-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"355M \nhttps://github.com/facebookresearch/fairseq/blob/main/examples/roberta/README.md"},{"id":"epoch:675f2555455c55eec8d0:training-compute","entityId":"epoch:675f2555455c55eec8d0","metric":"training-compute","value":8.5067e+21,"unit":"FLOP","observationPeriod":"2019-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 5: We pretrain our model using 1024 V100 GPUs for approximately one day.\n\nNote this is the base pretraining comparable to BERT, 100k steps. Subsequently they do more: \"increasing the number of pretraining steps\nfrom 100K to 300K, and then further to 500K\".\n\nSo assume 5x the 1024 V100 GPUs for 1d estimate. Mixed precision tensor cores get 1.25e14 FLOP/s.\n\n1024 * 1.25e14 * 5 * 24 * 3600 * 0.3 = 1.65888e22\n\n6ND estimate: batches are 8k sequences of 512 tokens; 500k updates means the model saw 500k * 8k * 512 = 2.048T tokens\n6 * 2.048T * 355M = 4.36224e21\n\ngeometric mean: sqrt(1.65888e22 * 4.36224e21) = 8.5067e21\n\nAuthors of \"AI and Memory Wall\" estimated model's training compute as 4,300,000 PFLOP = 4.3*10^21 FLOP\n(https://github.com/amirgholami/ai_and_memory_wall)"},{"id":"epoch:fc8b1b29cc206e16c23c:parameters","entityId":"epoch:fc8b1b29cc206e16c23c","metric":"parameters","value":125000000.0,"unit":"parameters","observationPeriod":"2019-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Specifically, we begin by training BERT models with the same configuration as BERTBASE (L = 12, H = 768, A = 12, 110M params\"\n\n\" This adds approximately 15M and 20M additional parameters for\nBERTBASE and BERTLARGE, respectively\"\n\n110M+15M = 125M "},{"id":"epoch:fc8b1b29cc206e16c23c:training-compute","entityId":"epoch:fc8b1b29cc206e16c23c","metric":"training-compute","value":1.536e+21,"unit":"FLOP","observationPeriod":"2019-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 125 * 10^6 parameters * 500000 steps * 8000 sequences per batch * 512 tokens per sequence = 1.536e+21 FLOP"},{"id":"epoch:addf39e8153114d78007:parameters","entityId":"epoch:addf39e8153114d78007","metric":"parameters","value":256999999.99999997,"unit":"parameters","observationPeriod":"2019-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"257M (Table 5) "},{"id":"epoch:addf39e8153114d78007:training-compute","entityId":"epoch:addf39e8153114d78007","metric":"training-compute","value":4.76e+18,"unit":"FLOP","observationPeriod":"2019-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 257000000 parameters * 103000000 tokens * 30 epochs = 4.76478e+18 FLOP\n"},{"id":"epoch:0393b36a6e7788879e3e:parameters","entityId":"epoch:0393b36a6e7788879e3e","metric":"parameters","value":160000000.0,"unit":"parameters","observationPeriod":"2019-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0393b36a6e7788879e3e:training-compute","entityId":"epoch:0393b36a6e7788879e3e","metric":"training-compute","value":2.388e+19,"unit":"FLOP","observationPeriod":"2019-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"FLOP calculation: 6*829250940*30*160000000=2.3882427e+19"},{"id":"epoch:33e7c4a77686bc94a075:parameters","entityId":"epoch:33e7c4a77686bc94a075","metric":"parameters","value":12000000.0,"unit":"parameters","observationPeriod":"2019-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:33e7c4a77686bc94a075:training-compute","entityId":"epoch:33e7c4a77686bc94a075","metric":"training-compute","value":2000000000000000.0,"unit":"FLOP","observationPeriod":"2019-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*30*929000*12000000=2.00664e+15"},{"id":"epoch:9c7eeb7567cb2077e381:parameters","entityId":"epoch:9c7eeb7567cb2077e381","metric":"parameters","value":85300000.0,"unit":"parameters","observationPeriod":"2019-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9c7eeb7567cb2077e381:training-compute","entityId":"epoch:9c7eeb7567cb2077e381","metric":"training-compute","value":1.58e+18,"unit":"FLOP","observationPeriod":"2019-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Tokens: 30 epochs * 103000000 tokens = 3090000000 tokens\n\nFLOP: 6 FLOP / parameter / token * 85300000* parameters * 3090000000 tokens = 1.581462e+18 FLOP"},{"id":"epoch:dbcd65cbb2509c609eb9:parameters","entityId":"epoch:dbcd65cbb2509c609eb9","metric":"parameters","value":38000000.0,"unit":"parameters","observationPeriod":"2019-06-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use a 12-layer Transformer with a hidden size of 512 units and 8 attention heads, leading to a 38M-parameter model\""},{"id":"epoch:dbcd65cbb2509c609eb9:training-compute","entityId":"epoch:dbcd65cbb2509c609eb9","metric":"training-compute","value":3e+19,"unit":"FLOP","observationPeriod":"2019-06-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"All self-supervised models are trained on four NVIDIA V100 GPUs for one week\"\n\n(7 * 24 * 3600) s * 4 GPUs * 3.1e13 FLOP/s * 0.4 (utilization assumption) = 3e19"},{"id":"epoch:e36f0547ab76a2a7f039:parameters","entityId":"epoch:e36f0547ab76a2a7f039","metric":"parameters","value":6160000.0,"unit":"parameters","observationPeriod":"2019-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2cbbddea5f24f8ca4910:parameters","entityId":"epoch:2cbbddea5f24f8ca4910","metric":"parameters","value":44100000.0,"unit":"parameters","observationPeriod":"2019-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"44.1M"},{"id":"epoch:d9c9cc59325dc87bafeb:parameters","entityId":"epoch:d9c9cc59325dc87bafeb","metric":"parameters","value":829000000.0,"unit":"parameters","observationPeriod":"2019-06-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:150c06c008495f72bdd7:training-compute","entityId":"epoch:150c06c008495f72bdd7","metric":"training-compute","value":2.2374144e+20,"unit":"FLOP","observationPeriod":"2019-06-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming (!) they used V100s:\n\n125000000000000 FLOP / GPU / sec * 336 hours [see training time notes] * 3600 sec / hour * 4 GPUs * 0.3 [assumed utilization] = 1.8144e+20 FLOP "},{"id":"epoch:5b6024fd6928f9a18d42:parameters","entityId":"epoch:5b6024fd6928f9a18d42","metric":"parameters","value":158000000.0,"unit":"parameters","observationPeriod":"2019-06-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"158M"},{"id":"epoch:ab27e10192324fd67b15:parameters","entityId":"epoch:ab27e10192324fd67b15","metric":"parameters","value":35000000.0,"unit":"parameters","observationPeriod":"2019-06-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"35M (Table 2)"},{"id":"epoch:ab27e10192324fd67b15:training-compute","entityId":"epoch:ab27e10192324fd67b15","metric":"training-compute","value":3.15e+17,"unit":"FLOP","observationPeriod":"2019-06-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 35000000 parameters * 2000000 tokens * 750 epochs = 3.15e+17 FLOP"},{"id":"epoch:1e191bff0766f5e77356:parameters","entityId":"epoch:1e191bff0766f5e77356","metric":"parameters","value":256999999.99999997,"unit":"parameters","observationPeriod":"2019-06-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1e191bff0766f5e77356:training-compute","entityId":"epoch:1e191bff0766f5e77356","metric":"training-compute","value":3.7848651e+20,"unit":"FLOP","observationPeriod":"2019-06-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Fine-tuned from pre-trained Transformer-XL Large (upd 3.7832771e+20 FLOP, old estimation 1.09e19 FLOP). \n\nTotal: 3.7832771e20 + 1.588e17 = 3.7848651e+20 FLOP (Speculative confidence same as Transformer XL)"},{"id":"epoch:8a2bd90c36f69a738d2e:parameters","entityId":"epoch:8a2bd90c36f69a738d2e","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2019-06-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:8d50a42f2fa650da188b:parameters","entityId":"epoch:8d50a42f2fa650da188b","metric":"parameters","value":626000000.0,"unit":"parameters","observationPeriod":"2019-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://openai.com/blog/image-gpt/#rfref13e"},{"id":"epoch:92015e04135f264a9885:parameters","entityId":"epoch:92015e04135f264a9885","metric":"parameters","value":340000000.0,"unit":"parameters","observationPeriod":"2019-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Same size as BERT-Large, which was 340M"},{"id":"epoch:92015e04135f264a9885:training-compute","entityId":"epoch:92015e04135f264a9885","metric":"training-compute","value":6.19e+21,"unit":"FLOP","observationPeriod":"2019-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Specifically, we train on 512 TPU v3 chips for 500K steps with an Adam weight decay optimizer, linear learning rate decay, and a batch size of 8192, which takes about 5.5 days.\"\n\n123 teraflops * 5.5 days * 24 * 3600 * 512 * 0.3 utilization (assumption) ~= 8977858560*10^12=8.9*10^21\n\nAlternatively, 500k steps * batch size 8192 * sequence length 512 = 2.1T training passes. 340 million * 6 * 2 trillion = 4.3e21 FLOP. \n\nGeometric mean: sqrt(8.9e21 * 4.3e21) = 6.19e21"},{"id":"epoch:279825751ae21b461c51:parameters","entityId":"epoch:279825751ae21b461c51","metric":"parameters","value":665000000.0,"unit":"parameters","observationPeriod":"2019-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:56ac27f639e1d7295c4d:parameters","entityId":"epoch:56ac27f639e1d7295c4d","metric":"parameters","value":100000000000.0,"unit":"parameters","observationPeriod":"2019-05-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 1\n\nhttps://arxiv.org/abs/2104.05158"},{"id":"epoch:56ac27f639e1d7295c4d:training-compute","entityId":"epoch:56ac27f639e1d7295c4d","metric":"training-compute","value":4e+18,"unit":"FLOP","observationPeriod":"2019-05-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 1\n\nhttps://arxiv.org/abs/2104.05158"},{"id":"epoch:288ec19ec05d5c1d3561:parameters","entityId":"epoch:288ec19ec05d5c1d3561","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2019-05-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our largest model, Grover-Mega, has 48 layers and 1.5 billion parameters, on par with GPT2\""},{"id":"epoch:288ec19ec05d5c1d3561:training-compute","entityId":"epoch:288ec19ec05d5c1d3561","metric":"training-compute","value":4.6386183e+21,"unit":"FLOP","observationPeriod":"2019-05-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We trained Grover-Mega for 800k iterations, using a batch size of 512 and 256 TPU v3 cores. Training time was two weeks.\"\n\nTPUv3 is 123 teraflops, but for 2 cores. \n\nSo 256/2 * 123 teraflops * 14 days * 24 * 3600 * 0.3 (utilization assumption) = 5.7e21\n\n6 FLOP / parameter / token * 1.5 * 10^9 parameters * 800000 steps * 512 [batch size] * 1024 [sequence length] =  3.7748736e+21 FLOP\n\nsqrt(5.7e21*3.7748736e+21) = 4.6386183e+21\n"},{"id":"epoch:d2d1acf929cf1d6f0038:parameters","entityId":"epoch:d2d1acf929cf1d6f0038","metric":"parameters","value":5200000.0,"unit":"parameters","observationPeriod":"2019-05-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From https://arxiv.org/pdf/1807.11626.pdf"},{"id":"epoch:d2d1acf929cf1d6f0038:training-compute","entityId":"epoch:d2d1acf929cf1d6f0038","metric":"training-compute","value":1.5e+21,"unit":"FLOP","observationPeriod":"2019-05-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"each architecture search takes 4.5 days on 64 TPUv2 devices\"\nThis seems to be referring to a TPUv2 pod, consisting of 64 four-chip modules. The total performance is 11.5 petaFLOPS.\nhttps://en.wikipedia.org/wiki/Tensor_Processing_Unit#Second_generation_TPU\nAssuming a 33% utilization rate:\n\n4.5 days * 64 * 180 teraFLOPS * 0.33 = 1.48*10^21 FLOP\n\nHowever, it is unclear if \"64 TPUv2 devices\" refers to chips or modules, so the true compute might be 1/4 of this amount."},{"id":"epoch:8a89a1f327621a223014:parameters","entityId":"epoch:8a89a1f327621a223014","metric":"parameters","value":4900000.0,"unit":"parameters","observationPeriod":"2019-05-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From https://arxiv.org/pdf/1807.11626.pdf"},{"id":"epoch:8a89a1f327621a223014:training-compute","entityId":"epoch:8a89a1f327621a223014","metric":"training-compute","value":1.5e+21,"unit":"FLOP","observationPeriod":"2019-05-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"each architecture search takes 4.5 days on 64 TPUv2 devices\"\nThis seems to be referring to a TPUv2 pod, consisting of 64 four-chip modules. The total performance is 11.5 petaFLOPS.\nhttps://en.wikipedia.org/wiki/Tensor_Processing_Unit#Second_generation_TPU\nAssuming a 33% utilization rate:\n\n4.5 days * 64 * 180 teraFLOPS * 0.33 = 1.48*10^21 FLOP\n\nHowever, it is unclear if \"64 TPUv2 devices\" refers to chips or modules, so the true compute might be 1/4 of this amount."},{"id":"epoch:318fe8e31b034c02e886:parameters","entityId":"epoch:318fe8e31b034c02e886","metric":"parameters","value":480000000.0,"unit":"parameters","observationPeriod":"2019-05-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:77f17256b5fb03328c7a:parameters","entityId":"epoch:77f17256b5fb03328c7a","metric":"parameters","value":7800000.0,"unit":"parameters","observationPeriod":"2019-05-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.8M (Figure 1)"},{"id":"epoch:75f6200949392446f10f:parameters","entityId":"epoch:75f6200949392446f10f","metric":"parameters","value":303000000.0,"unit":"parameters","observationPeriod":"2019-05-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://openai.com/blog/image-gpt/#rfref25d"},{"id":"epoch:92ef7dc0666827029445:parameters","entityId":"epoch:92ef7dc0666827029445","metric":"parameters","value":31400000.0,"unit":"parameters","observationPeriod":"2019-05-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2:\n 31.4M"},{"id":"epoch:4f658c98c83b89d4cc97:parameters","entityId":"epoch:4f658c98c83b89d4cc97","metric":"parameters","value":34000000.0,"unit":"parameters","observationPeriod":"2019-05-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"34M, Table 2"},{"id":"epoch:4f658c98c83b89d4cc97:training-compute","entityId":"epoch:4f658c98c83b89d4cc97","metric":"training-compute","value":4.08e+17,"unit":"FLOP","observationPeriod":"2019-05-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 34000000 parameters * 2000000 tokens * 1000 epochs = 4.08e+17 FLOP"},{"id":"epoch:be45e03d0275e91ea09f:parameters","entityId":"epoch:be45e03d0275e91ea09f","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2019-05-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:3bd61a6cccb2a0d2b9cb:parameters","entityId":"epoch:3bd61a6cccb2a0d2b9cb","metric":"parameters","value":25000000.0,"unit":"parameters","observationPeriod":"2019-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"25M parameters vanilla ResNet50"},{"id":"epoch:be7f5f657d6cf1de8aed:parameters","entityId":"epoch:be7f5f657d6cf1de8aed","metric":"parameters","value":193000000.0,"unit":"parameters","observationPeriod":"2019-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:91fb13c80ad90ae4b078:parameters","entityId":"epoch:91fb13c80ad90ae4b078","metric":"parameters","value":2038431744.0,"unit":"parameters","observationPeriod":"2019-04-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From article:\n72 Layers\n24 Attention heads\n\nAssumptions: \nD_head = 64 (GPT-2)\nD_model = 24*64=1536 (based on D_head)\nD_mlp = 4*1536=6144\n\nAttention: H*(W*(2*D+N)+N*M)\n24*(1536*(2*64+64)+64*1536)=9437184\nMLP: 1536*6144+6144*1536=18874368\nTotal parameters: 72*(9437184+18874368)=2038431744\n"},{"id":"epoch:91fb13c80ad90ae4b078:training-compute","entityId":"epoch:91fb13c80ad90ae4b078","metric":"training-compute","value":2.208301056e+20,"unit":"FLOP","observationPeriod":"2019-04-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From article:\n72 Layers\n24 Attention heads\n4096 ctx length\n\nAssumptions: \nD_head = 64 (GPT-2)\nD_model = 24*64=1536 (based on D_head)\nD_mlp = 4*1536=6144\nAverage sequence length: 5000\nEpochs: 5\nTraining data: 500000\n\nForward FLOP:\nAttention: 2*H*(W*(2*D+N)+L*(D+N)+N*M)\n2*24*(1536*(2*64+64)+4096*(64+64)+64*1536)=44040192\nMLP: 2*(1536*6144+6144*1536)=37748736\nTotal forward flop: 72*(44040192+37748736)=5888802816\nTraining compute\n5888802816*3*500000*5000*5=220830105600000000000=2.2e20\n"},{"id":"epoch:6c6f54bc7704caceab2d:parameters","entityId":"epoch:6c6f54bc7704caceab2d","metric":"parameters","value":95000000.0,"unit":"parameters","observationPeriod":"2019-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"95M"},{"id":"epoch:6c6f54bc7704caceab2d:training-compute","entityId":"epoch:6c6f54bc7704caceab2d","metric":"training-compute","value":4.104e+18,"unit":"FLOP","observationPeriod":"2019-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 95000000 parameters * 90000000 tokens * 80 epochs = 4.104e+18 FLOP"},{"id":"epoch:94cdc0f95206c400c3f8:parameters","entityId":"epoch:94cdc0f95206c400c3f8","metric":"parameters","value":59000000.0,"unit":"parameters","observationPeriod":"2019-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"59M"},{"id":"epoch:9eb0fff4d4b6c6cd3f1a:parameters","entityId":"epoch:9eb0fff4d4b6c6cd3f1a","metric":"parameters","value":152000000.0,"unit":"parameters","observationPeriod":"2019-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"152M"},{"id":"epoch:9eb0fff4d4b6c6cd3f1a:training-compute","entityId":"epoch:9eb0fff4d4b6c6cd3f1a","metric":"training-compute","value":1.45152e+21,"unit":"FLOP","observationPeriod":"2019-04-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"125000000000000 FLOP / GPU /sec [V100 reported, bf16 assumed] * 64 GPUs * 168 hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.45152e+21 FLOP"},{"id":"epoch:e9e6b3f2ea05e5fff256:parameters","entityId":"epoch:e9e6b3f2ea05e5fff256","metric":"parameters","value":395000000.0,"unit":"parameters","observationPeriod":"2019-04-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"395M (Table 6)"},{"id":"epoch:e9e6b3f2ea05e5fff256:training-compute","entityId":"epoch:e9e6b3f2ea05e5fff256","metric":"training-compute","value":1.5405e+20,"unit":"FLOP","observationPeriod":"2019-04-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 395000000 parameters * 1300000000 parameters * 50 epochs = 1.5405e+20 FLOP\n\n________\nin the Algorithmic progress paper, the estimation was 5.21E+20 FLOP"},{"id":"epoch:85c44a488827e44af264:parameters","entityId":"epoch:85c44a488827e44af264","metric":"parameters","value":340000000.0,"unit":"parameters","observationPeriod":"2019-04-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Taken from Bert paper."},{"id":"epoch:f670ba6db0b1a3afb34f:parameters","entityId":"epoch:f670ba6db0b1a3afb34f","metric":"parameters","value":340000000.0,"unit":"parameters","observationPeriod":"2019-04-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e3375e8b9c83fefca8e0:parameters","entityId":"epoch:e3375e8b9c83fefca8e0","metric":"parameters","value":256999999.99999997,"unit":"parameters","observationPeriod":"2019-04-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:dd9c4dca549c91ffc430:parameters","entityId":"epoch:dd9c4dca549c91ffc430","metric":"parameters","value":8720.0,"unit":"parameters","observationPeriod":"2019-04-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Calculations here: https://docs.google.com/document/d/1BTmyZ9KVTIwkp9z9tRRFsWM5A0QcXiufXrIhujffyso/edit?tab=t.0#heading=h.dcux1bvmijlm"},{"id":"epoch:dd9c4dca549c91ffc430:training-compute","entityId":"epoch:dd9c4dca549c91ffc430","metric":"training-compute","value":4.536e+17,"unit":"FLOP","observationPeriod":"2019-04-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Calculations here: \nhttps://docs.google.com/document/d/1BTmyZ9KVTIwkp9z9tRRFsWM5A0QcXiufXrIhujffyso/edit?tab=t.0#heading=h.c0h2t2icf5xr\n\nAssuming a single GPU (the model is quite small):\n1.0e+3*100*1.1e+13*0.4=4.5e+17"},{"id":"epoch:2359bb090e1baa8e13fa:parameters","entityId":"epoch:2359bb090e1baa8e13fa","metric":"parameters","value":26128.0,"unit":"parameters","observationPeriod":"2019-04-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Calculations here: https://docs.google.com/document/d/1BTmyZ9KVTIwkp9z9tRRFsWM5A0QcXiufXrIhujffyso/edit?tab=t.0#heading=h.913mrln2g0cv "},{"id":"epoch:2359bb090e1baa8e13fa:training-compute","entityId":"epoch:2359bb090e1baa8e13fa","metric":"training-compute","value":4.536e+17,"unit":"FLOP","observationPeriod":"2019-04-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Calculations here: https://docs.google.com/document/d/1BTmyZ9KVTIwkp9z9tRRFsWM5A0QcXiufXrIhujffyso/edit?tab=t.0#heading=h.obxgi47b0fxe\n\nUpdated calculation assuming 100s per epoch and a single GPU.\n1000*100*11340000000000*0.4=453600000000000000"},{"id":"epoch:f5f1a98e2a06545edab5:parameters","entityId":"epoch:f5f1a98e2a06545edab5","metric":"parameters","value":26128.0,"unit":"parameters","observationPeriod":"2019-04-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Calculations here:\nhttps://docs.google.com/document/d/1BTmyZ9KVTIwkp9z9tRRFsWM5A0QcXiufXrIhujffyso/edit?tab=t.0#heading=h.913mrln2g0cv"},{"id":"epoch:f5f1a98e2a06545edab5:training-compute","entityId":"epoch:f5f1a98e2a06545edab5","metric":"training-compute","value":4.536e+17,"unit":"FLOP","observationPeriod":"2019-04-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Calculations here:\nhttps://docs.google.com/document/d/1BTmyZ9KVTIwkp9z9tRRFsWM5A0QcXiufXrIhujffyso/edit?tab=t.0#heading=h.lk6l1te7vrkv\n\n\nUpdated calculation assuming 100s per epoch and a single GPU.\n1000*100*11340000000000*0.4=453600000000000000"},{"id":"epoch:6c617201b59607fd91cb:parameters","entityId":"epoch:6c617201b59607fd91cb","metric":"parameters","value":26128.0,"unit":"parameters","observationPeriod":"2019-04-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Calculations here:\nhttps://docs.google.com/document/d/1BTmyZ9KVTIwkp9z9tRRFsWM5A0QcXiufXrIhujffyso/edit?tab=t.0#heading=h.913mrln2g0cv\n"},{"id":"epoch:6c617201b59607fd91cb:training-compute","entityId":"epoch:6c617201b59607fd91cb","metric":"training-compute","value":4.536e+17,"unit":"FLOP","observationPeriod":"2019-04-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Calculations here:\nhttps://docs.google.com/document/d/1BTmyZ9KVTIwkp9z9tRRFsWM5A0QcXiufXrIhujffyso/edit?tab=t.0#heading=h.3ria9ya1ty1o\n\n\nUpdated calculation assuming 100s per epoch and a single GPU.\n1000*100*11340000000000*0.4=453600000000000000\n"},{"id":"epoch:204af3f99cd39e970dde:parameters","entityId":"epoch:204af3f99cd39e970dde","metric":"parameters","value":7000000.0,"unit":"parameters","observationPeriod":"2019-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b8569468a7e09ed559a9:parameters","entityId":"epoch:b8569468a7e09ed559a9","metric":"parameters","value":7000000.0,"unit":"parameters","observationPeriod":"2019-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:de95a39f9bcf91abb46d:parameters","entityId":"epoch:de95a39f9bcf91abb46d","metric":"parameters","value":33000000.0,"unit":"parameters","observationPeriod":"2019-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:3a35ee05b042451ca367:parameters","entityId":"epoch:3a35ee05b042451ca367","metric":"parameters","value":23000000.0,"unit":"parameters","observationPeriod":"2019-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:3a35ee05b042451ca367:training-compute","entityId":"epoch:3a35ee05b042451ca367","metric":"training-compute","value":3.530047365121324e+18,"unit":"FLOP","observationPeriod":"2019-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"GPU hours: 120*60*60*125000000000000*0.3=1.62e+19\nOperations: 6*23000000*929000*6000=7.69212e+17\nGeometric mean: 3530047365121324032 (3.5e18)"},{"id":"epoch:76e344630fcf88388b82:parameters","entityId":"epoch:76e344630fcf88388b82","metric":"parameters","value":23000000.0,"unit":"parameters","observationPeriod":"2019-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:76e344630fcf88388b82:training-compute","entityId":"epoch:76e344630fcf88388b82","metric":"training-compute","value":7.30000001e+17,"unit":"FLOP","observationPeriod":"2019-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"PTB has 912344 tokens. The model has 23M parameters and was trained for 6k epochs. If the model was dense, 6 FLOP/token/param/epoch * 6k epochs * 23M params * 912k tokens = 1.05e18 FLOP.\n\nAlternatively, the model was trained on 1 V100 GPU and \"In terms of efficiency, the overall cost... is within 1 GPU day\" so the training time was around or below 24 hours. Half precision and 30% utilization would be a pretty good match for the arithmetic estimate: 24 hours * 30% * 28 TFLOPS = 7.3e17 FLOP."},{"id":"epoch:892cbfb37acb3b58e796:training-compute","entityId":"epoch:892cbfb37acb3b58e796","metric":"training-compute","value":2.56e+18,"unit":"FLOP","observationPeriod":"2019-04-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From author communication:\n\nPrecision: float32\n\nHardware: 4 GPU  NVIDIA 1080Ti\n\nNVIDIA 1080Ti: 1.06E+13\n\nCompute: 7 GPU-days\n\n0.4 * 1.06E+13 FLOP/s * 7 days * 24h/day * 3600s/h\n= 2.56E+18"},{"id":"epoch:5ff85e8799ed9b40092f:parameters","entityId":"epoch:5ff85e8799ed9b40092f","metric":"parameters","value":247000000.00000003,"unit":"parameters","observationPeriod":"2019-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The first model has 16 blocks, inner dimension 4K and embedding dimension 1K\"\n\n247M as in Baevski and Auli (2019) Transformer"},{"id":"epoch:5ff85e8799ed9b40092f:training-compute","entityId":"epoch:5ff85e8799ed9b40092f","metric":"training-compute","value":3.1804274e+19,"unit":"FLOP","observationPeriod":"2019-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 247000000 parameters * 103000000 tokens * 180 epochs = 2.747628e+19 FLOP \n\nfor translation model, Table 2:\n\n31330000000000 FLOP / second / GPU * 128 GPUs * 8.5 hours * 3600 seconds / hour * 0.3 [assumed precision] = 3.6814003e+19 FLOP\n\nsqrt(2.747628e+19*3.6814003e+19) = 3.1804274e+19 FLOP\n\nSpeculative confidence since amount of parameters and epochs are assumed as well as hardware estimation is given for another model in the paper\n\n__________\nIn the Algorithmic progress paper the estimation was 7.30E+18 FLOP"},{"id":"epoch:9cfd9103b71194d01478:parameters","entityId":"epoch:9cfd9103b71194d01478","metric":"parameters","value":18200000.0,"unit":"parameters","observationPeriod":"2019-03-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"1,900-dimensional single-layer multiplicative LSTM (~18.2 million parameters)\""},{"id":"epoch:9cfd9103b71194d01478:training-compute","entityId":"epoch:9cfd9103b71194d01478","metric":"training-compute","value":2.2e+19,"unit":"FLOP","observationPeriod":"2019-03-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Training was performed using data parallelism on four Nvidia K80 GPUs (mLSTM-1,900) or two Nvidia K-40s (4\u00d7 mLSTM-256, 4\u00d7 mLSTM-64). The mLSTM-1,900 model was trained for ~770,000 weight updates, or ~3.5\u2009weeks wall clock time, corresponding to ~1 epoch.\" [Methods - Unsupervised training dataset]\n\nAssuming 30% utilization rate and single-precision performance\n\nEstimate: 3.5 weeks * 7 days/week * 24 hours/day * 60 min/hour * 60 sec/min * 4 GPUs *8.73e12 FLOP/sec * 0.3"},{"id":"epoch:d7fbd90c0c014df3ee7e:parameters","entityId":"epoch:d7fbd90c0c014df3ee7e","metric":"parameters","value":110000000.0,"unit":"parameters","observationPeriod":"2019-03-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"110M\nsize of bert base from https://huggingface.co/google-bert/bert-base-uncased\nrelevant citation: \n\"We use the original BERT code to\ntrain SCIBERT on our corpus with the same con-\nfiguration and size as BERT-Base. We train 4\ndifferent versions of SCIBERT: (i) cased or un-\ncased and (ii) BASEVOCAB or SCIVOCAB. The\ntwo models that use BASEVOCAB are finetuned\nfrom the corresponding BERT-Base models. The\nother two models that use the new SCIVOCAB are\ntrained from scratch.\""},{"id":"epoch:d7fbd90c0c014df3ee7e:training-compute","entityId":"epoch:d7fbd90c0c014df3ee7e","metric":"training-compute","value":8.926848e+19,"unit":"FLOP","observationPeriod":"2019-03-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4*123e12*0.3*(7*24*3600) = 8.926848e+19\n(num gpu) * (peak compute) * (assumed utilization rate) * (time in seconds)\nWe have:\n 4 TPUv3 chips.123teraFLOPS per chip.\n7 days of training\n\"We use a single TPU v3 with 8 cores. Training the SCIVOCAB models from scratch on our corpus takes 1 week (5 days with max length 128, then 2 days with max length 512). \"\n\nIf this compute estimate is accurate and BERT is approximately dense, then C=6eND -> e=C/6ND ~= 40 epochs."},{"id":"epoch:a3798f0e4fa23e91d27e:parameters","entityId":"epoch:a3798f0e4fa23e91d27e","metric":"parameters","value":67300000.0,"unit":"parameters","observationPeriod":"2019-03-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ec32c3e1f97cea6a8deb:parameters","entityId":"epoch:ec32c3e1f97cea6a8deb","metric":"parameters","value":437700000.0,"unit":"parameters","observationPeriod":"2019-02-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Regarding the model, for these experiments we\nuse a larger Transformer model with 6 layers in\nboth the encoder and the decoder, model dimension set to 1024, hidden dimension size of 8192,\nand 16 attention heads. This results in a model\nwith approximately 473.7M parameters.\""},{"id":"epoch:7d6ab42a2522f3e310d5:parameters","entityId":"epoch:7d6ab42a2522f3e310d5","metric":"parameters","value":2500000.0,"unit":"parameters","observationPeriod":"2019-02-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://arxiv.org/abs/2210.00849 gives parameter count for AlphaZero in Fig 1b."},{"id":"epoch:7d6ab42a2522f3e310d5:training-compute","entityId":"epoch:7d6ab42a2522f3e310d5","metric":"training-compute","value":2.32e+19,"unit":"FLOP","observationPeriod":"2019-02-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"[KataGo] surpasses the strength of ELF OpenGo after training on about 27 V100 GPUs for 19 days\"\n14.13 teraFLOP/s * 19 days = 2.32e+19 FLOP"},{"id":"epoch:d88d8e35c7b73fb3e2ad:training-compute","entityId":"epoch:d88d8e35c7b73fb3e2ad","metric":"training-compute","value":3.723192e+18,"unit":"FLOP","observationPeriod":"2019-02-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"For their searched Imagenet models, they used 200 GPU hours on a V100 GPU.\n\nAt FP32, a V100 GPU has a peak performance of 1.56E+13 FLOPS.\n\nUtilization rate of 0.33.\n\n200 h * 3600 second / hour* 1.6e+13 flop /second * 0.33 = 3.7e+18 flop"},{"id":"epoch:078ad6f93adfc618b868:training-compute","entityId":"epoch:078ad6f93adfc618b868","metric":"training-compute","value":1.296e+19,"unit":"FLOP","observationPeriod":"2019-02-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"All models were implemented in PyTorch and trained on a single NVIDIA Tesla V100 GPU. Each model took roughly 3 days to train and required 16 GB of GPU RAM\"\n1.3e+14*0.4*259200s=1.3e+19"},{"id":"epoch:9bc2627fa0719e0c751c:parameters","entityId":"epoch:9bc2627fa0719e0c751c","metric":"parameters","value":37000000.0,"unit":"parameters","observationPeriod":"2019-02-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"while our code2seq model had only 37M"},{"id":"epoch:9bc2627fa0719e0c751c:training-compute","entityId":"epoch:9bc2627fa0719e0c751c","metric":"training-compute","value":1.1513908e+19,"unit":"FLOP","observationPeriod":"2019-02-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*37*10^6*997393280*52 = 1.1513908e+19"},{"id":"epoch:0baa73425d4fddd30a0f:parameters","entityId":"epoch:0baa73425d4fddd30a0f","metric":"parameters","value":1500000000.0,"unit":"parameters","observationPeriod":"2019-02-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"GPT-2 is a large transformer-based language model with 1.5 billion parameters\""},{"id":"epoch:0baa73425d4fddd30a0f:training-compute","entityId":"epoch:0baa73425d4fddd30a0f","metric":"training-compute","value":1.920000000001e+21,"unit":"FLOP","observationPeriod":"2019-02-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Estimating based on compute = 6 FLOP/token/param * epochs * parameters * tokens.\n\n40GB dataset is approximately 8B words, or 1/0.75 * 8B = 10.66B tokens.\n\nThe number of epochs is not reported, but another paper [1] claims in table 1 that it is 20 or 100 epochs, and another paper [2] claims 12 epochs based on communication with the GPT-2 authors (page 4).\n\n12 epochs is the modal, most credible value. Mean of probability mass is probably around 20 epochs, so calculating from that value:\n\n6 * (40 * 200 million * 1/0.75 * 20) * 1.5 billion parameters = 1.92e21\nhttps://www.wolframalpha.com/input?i=6+FLOP+*+20+*+%2840+billion+%2F+5+*+%284%2F3%29%29+*+1.5+billion\n\n[1] https://arxiv.org/abs/1906.06669 One Epoch Is All You Need\n[2] https://www.usenix.org/system/files/sec21-carlini-extracting.pdf Extracting Data From Large Language Models\n\nIt also appears the model was trained on TPU v3 chips:\nhttps://huggingface.co/openai-community/gpt2"},{"id":"epoch:f9a68d50a50e57e258ae:parameters","entityId":"epoch:f9a68d50a50e57e258ae","metric":"parameters","value":124000000.0,"unit":"parameters","observationPeriod":"2019-02-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Note that the initial paper release stated GPT-2 small had 117M parameters. The official github repo notes that this was due to an error:\nhttps://github.com/openai/gpt-2?tab=readme-ov-file"},{"id":"epoch:f9a68d50a50e57e258ae:training-compute","entityId":"epoch:f9a68d50a50e57e258ae","metric":"training-compute","value":7.936e+20,"unit":"FLOP","observationPeriod":"2019-02-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"assuming 100 epochs (consistent with original GPT paper + reasoning from here: https://arxiv.org/pdf/1906.06669)\n\n6 FLOP / token / parameter * 124*10^6 parameters * 10666666666.7 tokens * 100 epochs = 7.936e+20 FLOP"},{"id":"epoch:b443e765a93ab5bb8a2f:parameters","entityId":"epoch:b443e765a93ab5bb8a2f","metric":"parameters","value":774000000.0,"unit":"parameters","observationPeriod":"2019-02-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Note that the initial paper release stated GPT-2 large had 762M parameters. The official github repo notes that this was due to an error:\nhttps://github.com/openai/gpt-2?tab=readme-ov-file"},{"id":"epoch:b443e765a93ab5bb8a2f:training-compute","entityId":"epoch:b443e765a93ab5bb8a2f","metric":"training-compute","value":4.9536e+21,"unit":"FLOP","observationPeriod":"2019-02-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"assuming 100 epochs (consistent with original GPT paper + reasoning from here: https://arxiv.org/pdf/1906.06669)\n\n6 FLOP / token / parameter * 774*10^6 parameters * 10666666666.7 tokens * 100 epochs = 4.9536e+21 FLOP"},{"id":"epoch:1113d7b2f89db6a1de24:parameters","entityId":"epoch:1113d7b2f89db6a1de24","metric":"parameters","value":355000000.0,"unit":"parameters","observationPeriod":"2019-02-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Note that the initial paper release stated GPT-2 medium had 345M parameters. The official github repo notes that this was due to an error:\nhttps://github.com/openai/gpt-2?tab=readme-ov-file"},{"id":"epoch:1113d7b2f89db6a1de24:training-compute","entityId":"epoch:1113d7b2f89db6a1de24","metric":"training-compute","value":2.272000000071e+21,"unit":"FLOP","observationPeriod":"2019-02-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"assuming 100 epochs (consistent with original GPT paper + reasoning from here: https://arxiv.org/pdf/1906.06669)\n\n6* 355000000*10666666667*100=2.272000e+21"},{"id":"epoch:325b004652165553cef7:parameters","entityId":"epoch:325b004652165553cef7","metric":"parameters","value":66000000.0,"unit":"parameters","observationPeriod":"2019-02-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:325b004652165553cef7:training-compute","entityId":"epoch:325b004652165553cef7","metric":"training-compute","value":3.31e+16,"unit":"FLOP","observationPeriod":"2019-02-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 66000000 parameters * 929000 tokens * 90 epochs = 3.310956e+16 FLOP"},{"id":"epoch:1600eb3e1f047d4e3bfb:parameters","entityId":"epoch:1600eb3e1f047d4e3bfb","metric":"parameters","value":4600000.0,"unit":"parameters","observationPeriod":"2019-02-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:95e2c97de83e46d9a2f9:parameters","entityId":"epoch:95e2c97de83e46d9a2f9","metric":"parameters","value":764000.0,"unit":"parameters","observationPeriod":"2019-02-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: https://docs.google.com/spreadsheets/d/1Kj4Q5WADcDXtUJLIOfGTCE3tGvxNczEMwyy8QtgSkHk/edit#gid=54587040&fvid=1361937389"},{"id":"epoch:95e2c97de83e46d9a2f9:training-compute","entityId":"epoch:95e2c97de83e46d9a2f9","metric":"training-compute","value":4.3e+18,"unit":"FLOP","observationPeriod":"2019-02-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"14.13e+12 FLOP/s * 7 days * 86400 s/day * 0.50 utilization = 4.3e+18 FLOP"},{"id":"epoch:7256c4a6c5db9393c89c:parameters","entityId":"epoch:7256c4a6c5db9393c89c","metric":"parameters","value":330000000.0,"unit":"parameters","observationPeriod":"2019-01-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:fa7292243c57fabba467:parameters","entityId":"epoch:fa7292243c57fabba467","metric":"parameters","value":9120000.0,"unit":"parameters","observationPeriod":"2019-01-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:57d23c9feb4aee015624:parameters","entityId":"epoch:57d23c9feb4aee015624","metric":"parameters","value":800000.0,"unit":"parameters","observationPeriod":"2019-01-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:370afe0f5593ecbaf096:training-compute","entityId":"epoch:370afe0f5593ecbaf096","metric":"training-compute","value":4.85606016e+18,"unit":"FLOP","observationPeriod":"2019-01-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(4) * (6.691 * 10**12) * (168* 3600) * (0.3) = \n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate) = \n\ntraining time - about one week = 168 hours from\n\"It takes about one week to train the 2D bounding box net-work, and two hours for the orientation/3D scale network\non KITTI with 4 TITAN-X GPUs. The landmark detector is\ntrained on Pascal3D. The training process for the monocu-\nlar depth estimation network is unsupervised using KITTI\nstereo-pairs, which takes around 5 to 12 hours depending\non the amount of data available. \"\n\ngpu flops - FP32 (float) 6.691 TFLOPS from https://www.techpowerup.com/gpu-specs/geforce-gtx-titan-x.c2632"},{"id":"epoch:07a3316be03c0c6f3bf6:parameters","entityId":"epoch:07a3316be03c0c6f3bf6","metric":"parameters","value":256999999.99999997,"unit":"parameters","observationPeriod":"2019-01-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Transformer-XL Large, Table 1"},{"id":"epoch:07a3316be03c0c6f3bf6:training-compute","entityId":"epoch:07a3316be03c0c6f3bf6","metric":"training-compute","value":3.7832771e+20,"unit":"FLOP","observationPeriod":"2019-01-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 257000000 parameters * 103000000 tokens * 1908 epochs [see dataset size notes] = 3.0304001e+20 FLOP\n\nfrom training code (https://github.com/kimiyoung/transformer-xl/blob/master/tf/scripts/wt103_large_tpu.sh) they used 64 tpv3 cores \n\n123000000000000 FLOP/s/chip* (64 cores / 2 cores per chip) * 4000000 steps *  0.1 sec / step [assumption] * 0.3 [assumed utilization] = 4.7232e+20 FLOP\n\ngeometric mean\nsqrt(3.0304001e+20 * 4.7232e+20) = 3.7832771e+20\n\nspeculative confidence given assumptions used\n_________\nprevious estimation in the algorithmic progress paper was 1.09 \u00d7 10^19 FLOP without explanation \n"},{"id":"epoch:ed61fb40182c1dbe072f:parameters","entityId":"epoch:ed61fb40182c1dbe072f","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2019-01-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"24M, Table 5"},{"id":"epoch:5291b19abf5d1c799c79:parameters","entityId":"epoch:5291b19abf5d1c799c79","metric":"parameters","value":36500000.0,"unit":"parameters","observationPeriod":"2019-01-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From author communication\n\nWideResNet 28-10 models with 36.5 million parameters (3.65E+07)"},{"id":"epoch:5291b19abf5d1c799c79:training-compute","entityId":"epoch:5291b19abf5d1c799c79","metric":"training-compute","value":4.716e+17,"unit":"FLOP","observationPeriod":"2019-01-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From author communication\n\nPer image: 5.24 billion FLOPs (5.24E+09)  Per training run: 50k times 5.24E+09 times 1800 epochs\n\n5240000000*50000*1800=471600000000000000=4.72e17"},{"id":"epoch:3ca618f198550e6cb130:parameters","entityId":"epoch:3ca618f198550e6cb130","metric":"parameters","value":56000000.0,"unit":"parameters","observationPeriod":"2019-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5a24a0ae5dc7713574a7:parameters","entityId":"epoch:5a24a0ae5dc7713574a7","metric":"parameters","value":26200000.0,"unit":"parameters","observationPeriod":"2018-12-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5a24a0ae5dc7713574a7:training-compute","entityId":"epoch:5a24a0ae5dc7713574a7","metric":"training-compute","value":3.93e+16,"unit":"FLOP","observationPeriod":"2018-12-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 * 26.2M * 25M ~= 39300000000000000"},{"id":"epoch:4aaa5638046d081129f5:training-compute","entityId":"epoch:4aaa5638046d081129f5","metric":"training-compute","value":8459100000000000.0,"unit":"FLOP","observationPeriod":"2018-12-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"for a 6-dimensional data set of size 500, the RL algorithm finishes in approximately 15 minutes with a single GPU\"\"\n\nAssuming they used V100, fp16 and trained for similar amount of time:\n\n31330000000000 FLOP / GPU / sec * 15 min * 60 sec / min * 0.3 [assumed utilization] = 8.4591e+15 FLOP "},{"id":"epoch:7158ddac4b9193f81f29:training-compute","entityId":"epoch:7158ddac4b9193f81f29","metric":"training-compute","value":8459100000000000.0,"unit":"FLOP","observationPeriod":"2018-12-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"for a 6-dimensional data set of size 500, the RL algorithm finishes in approximately 15 minutes with a single GPU\"\"\n\nAssuming they used V100, fp16 and trained for similar amount of time:\n\n31330000000000 FLOP / GPU / sec * 15 min * 60 sec / min * 0.3 [assumed utilization] = 8.4591e+15 FLOP "},{"id":"epoch:66c01ce223c7c60e435b:training-compute","entityId":"epoch:66c01ce223c7c60e435b","metric":"training-compute","value":8459100000000000.0,"unit":"FLOP","observationPeriod":"2018-12-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"for a 6-dimensional data set of size 500, the RL algorithm finishes in approximately 15 minutes with a single GPU\"\"\n\nAssuming they used V100, fp16 and trained for similar amount of time:\n\n31330000000000 FLOP / GPU / sec * 15 min * 60 sec / min * 0.3 [assumed utilization] = 8.4591e+15 FLOP "},{"id":"epoch:745d89d32b70cdc2039b:parameters","entityId":"epoch:745d89d32b70cdc2039b","metric":"parameters","value":250000000.0,"unit":"parameters","observationPeriod":"2018-12-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"250M (Table 4)"},{"id":"epoch:0fc70b2f5afdeecb2b0c:parameters","entityId":"epoch:0fc70b2f5afdeecb2b0c","metric":"parameters","value":50000000.0,"unit":"parameters","observationPeriod":"2018-12-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"50M (Table 4)"},{"id":"epoch:99344e2ad3bdf774dcef:parameters","entityId":"epoch:99344e2ad3bdf774dcef","metric":"parameters","value":3800000.0,"unit":"parameters","observationPeriod":"2018-11-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Based on Fig. 9:\n\nDistance predictor network: [Total = 17684]\n(1) Maxout2D: 0 parameters\n(2) ResBlock x 18: 2*64*18 = 16384 parameters (a shift and scale parameter for every InstanceNorm2D layer)\n(3) Conv2D 1x1: 64*20 + 20 = 1300 parameters\n(4) Softmax: 0 parameters\n\nHydrogen bond predictor network: [Total = 3691073]\n(1) Maxout2D: 0 parameters\n(2) ResBlock x 18: (2*64*2 + 5*5*64*64*2)*18 = 3 691 008 parameters\n(3) Conv2D 1x1: 64*1 + 1 = 65 parameters\n(4) Sigmoid: 0 parameters\n\nTorsion angles and errors network [Total = 115459]\n(1) Maxout2D: 0 parameters\n(2) ResBlock x 18: 2*64*18 = 16384 parameters (a shift and scale parameter for every InstanceNorm2D layer)\n(3) BLSTM: 4\u00d72\u00d7(N+M)\u00d7M where M is 128 (hidden units in BLSTM layer) and N is 64 (input dimensionality) = 98304\n(4) Conv1D 1x1: 256*3 + 3 = 771 parameters\n\nEstimate total parameters = 3.8e6 parameters\n\n[See Section 'Additional constraint types and iterative predictions': \"a bidirectional recurrent LSTM layer with 128 hidden units (BLSTM in Fig. 9c), which embeds each row of the final 2-D 64-channel feature map in a single 256-D vector (concatenation of 128-D final timestep output states of the forward and reverse direction LSTM passes)\"]"},{"id":"epoch:f6b995264d131449d039:parameters","entityId":"epoch:f6b995264d131449d039","metric":"parameters","value":6000000000.0,"unit":"parameters","observationPeriod":"2018-11-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 5: "},{"id":"epoch:3bf54514c0cc556cab72:parameters","entityId":"epoch:3bf54514c0cc556cab72","metric":"parameters","value":557000000.0,"unit":"parameters","observationPeriod":"2018-11-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 4"},{"id":"epoch:b5dfc1b5c6658e760f2e:parameters","entityId":"epoch:b5dfc1b5c6658e760f2e","metric":"parameters","value":7200000.0,"unit":"parameters","observationPeriod":"2018-11-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Based on the details in the paper, the number of parameters in the Multi-cell LSTM model can be calculated as follows:\n\nThe model has 2 hidden LSTM layers, each with 1500 hidden units\n\nEach LSTM unit is a multi-cell LSTM with 10 memory cells per unit\n\nFor a standard LSTM layer with n hidden units:\n\nW matrix: n x input_size\nU matrix: n x n\n4 bias vectors of size n (for input, forget, cell, output gates)\nSo for each multi-cell LSTM layer with 1500 units and 10 cells per unit:\n\nW matrix: 1500 x input_size\nU matrix: 1500 x 1500\n4 bias vectors of size 1500\nNumber of parameters is same as standard LSTM layer\n\nFor the 2 hidden layers:\n\nInput size for Layer 1: embedding dimension (estimated 300 in paper)\n\nInput size for Layer 2: 1500 (output of layer 1)\n\nTotal params =\nLayer 1: 1500 x (300 + 1500 + 4) = 2,706,000\nLayer 2: 1500 x (1500 + 1500 + 4) = 4,506,000\n\nTotal Parameters = 2,706,000 + 4,506,000 = 7,212,000\n\nSo the total number of parameters for the Multi-cell LSTM model with 2 layers of 1500 units and 10 cells per unit is approximately 7.2 million."},{"id":"epoch:b5dfc1b5c6658e760f2e:training-compute","entityId":"epoch:b5dfc1b5c6658e760f2e","metric":"training-compute","value":2006640000000000.0,"unit":"FLOP","observationPeriod":"2018-11-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 7200000 parameters * 929000 tokens * 50 epochs = 2.00664e+15 FLOP"},{"id":"epoch:393f15bece7e879056c8:parameters","entityId":"epoch:393f15bece7e879056c8","metric":"parameters","value":46000000.0,"unit":"parameters","observationPeriod":"2018-11-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"This is the model trained on Penn Treebank, which uses as a base model the 23M model from Table 7 in https://aclanthology.org/D18-1489.pdf\n\nThey additionally train a discriminator with the same architecture, so total parameters is 2*23M = 46M"},{"id":"epoch:393f15bece7e879056c8:training-compute","entityId":"epoch:393f15bece7e879056c8","metric":"training-compute","value":5.188e+16,"unit":"FLOP","observationPeriod":"2018-11-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Base model uses 4.323e16 FLOPs.\nThey then train a discriminator using the same architecture for 30 epochs, and then use the discriminator to fine-tune the base model for another 15 epochs. Both of these latter training steps require running forward passes on both the discriminator and the language model, but only doing a backward pass on one of them.\n\nDiscriminator training: \n2*23M*30*1044112 + 6*23M*30*1044112 = 5.763e15\n\nLM fine-tuning:\n2*23M*15*1044112 + 6*23M*15*1044112 = 2.882e15\n\nTotal:\n4.323e16 + 5.763e15 + 2.882e15 = 5.188e16"},{"id":"epoch:da25d29fd6d9af942cc1:parameters","entityId":"epoch:da25d29fd6d9af942cc1","metric":"parameters","value":111920000.0,"unit":"parameters","observationPeriod":"2018-11-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The largest experiment is the model trained in section 4.3 (\"Large scale experiment\"). The dataset here has a vocabulary of 100k words. They use a single-layer LSTM with 500 hidden units, with differentiated softmax for the output layer. d_k for distributed softmax varies from 16-150 depending on the relative frequency of the output word. \n\nInput layer:\n100k * 500 = 50M parameters\n\nLSTM:\n4 * (500 + 500) * 500 = 2M parameters\n\nOutput layer:\nHere we need to assume the structure of differentiated softmax. Let's (somewhat arbitrarily) assume the top 10k words get 150-dim embeddings, the middle 30k get 50-dim embeddings, and the bottom 60k get 16-dim embeddings. Then we have:\n(10k * 150) + (30k * 50) + (60k * 16) = 3.96M parameters\n\nSo in total, we have\n50M + 2M + 3.96M = 55.96M\n\nIn other experiments, the authors use the same architecture for the discriminator as for the language model. So total parameter count would appear to be 2 x 55.96M = 111.92M"},{"id":"epoch:b12a7e9738aa9b06cde5:parameters","entityId":"epoch:b12a7e9738aa9b06cde5","metric":"parameters","value":4900000000.0,"unit":"parameters","observationPeriod":"2018-11-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4.9B from section 9.1 : ''The largest model (4.9B parameters) took 13 hours to train on a 512-core TPUv2 cluster.'"},{"id":"epoch:b12a7e9738aa9b06cde5:training-compute","entityId":"epoch:b12a7e9738aa9b06cde5","metric":"training-compute","value":1.617408e+20,"unit":"FLOP","observationPeriod":"2018-11-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"flops = (256) * ( 45 * 10**12) * (13 * 3600) * (0.3) = 1.6e20\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)\n\nfrom section 9.1 : ''The largest model (4.9B parameters) took 13 hours to train on a 512-core TPUv2 cluster.'\nfrom https://en.wikipedia.org/wiki/Tensor_Processing_Unit \n45TFLOPs per chips"},{"id":"epoch:7b605384d74798d07b0f:parameters","entityId":"epoch:7b605384d74798d07b0f","metric":"parameters","value":2900000000.0,"unit":"parameters","observationPeriod":"2018-11-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.9B from section 9.1 : \"On the WMT14 En-Fr translation tasks (3), we trained the models for 3 epochs. The largest model\n(2.9B parameters) was trained for 22 hours on a 128-core TPUv2 cluster.\""},{"id":"epoch:7b605384d74798d07b0f:training-compute","entityId":"epoch:7b605384d74798d07b0f","metric":"training-compute","value":6.84288e+19,"unit":"FLOP","observationPeriod":"2018-11-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"flops = (64) * ( 45 * 10**12) * (22 * 3600) * (0.3) = 6.8e19\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate)\n\nfrom section 9.1 : \"On the WMT14 En-Fr translation tasks (3), we trained the models for 3 epochs. The largest model\n(2.9B parameters) was trained for 22 hours on a 128-core TPUv2 cluster.\"\nfrom https://en.wikipedia.org/wiki/Tensor_Processing_Unit \n45TFLOPs per chips"},{"id":"epoch:62d20023eadf865885b5:training-compute","entityId":"epoch:62d20023eadf865885b5","metric":"training-compute","value":3.1593888e+17,"unit":"FLOP","observationPeriod":"2018-10-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8126000000000.000 * 36 * 3600 * 0.3 = 3.1593888e+17"},{"id":"epoch:49f335358b525e0a719c:parameters","entityId":"epoch:49f335358b525e0a719c","metric":"parameters","value":180000000.0,"unit":"parameters","observationPeriod":"2018-10-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"180M, Table 2"},{"id":"epoch:49f335358b525e0a719c:training-compute","entityId":"epoch:49f335358b525e0a719c","metric":"training-compute","value":2.78e+18,"unit":"FLOP","observationPeriod":"2018-10-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 180000000 parameters * 103000000 tokens * 25 epochs = 2.781e+18 FLOP"},{"id":"epoch:62215a085149538943f3:parameters","entityId":"epoch:62215a085149538943f3","metric":"parameters","value":34000000.0,"unit":"parameters","observationPeriod":"2018-10-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"34M, table 1"},{"id":"epoch:60f527da230e7bfe732e:parameters","entityId":"epoch:60f527da230e7bfe732e","metric":"parameters","value":340000000.0,"unit":"parameters","observationPeriod":"2018-10-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"340M"},{"id":"epoch:60f527da230e7bfe732e:training-compute","entityId":"epoch:60f527da230e7bfe732e","metric":"training-compute","value":2.85e+20,"unit":"FLOP","observationPeriod":"2018-10-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"more info here https://docs.google.com/document/d/1B8x6XYcmB1u6Tmq3VcbAtj5bzhDaj2TcIPyK6Wpupx4/edit?usp=sharing\n285000000000000000000 = 2.85 \u00d7 10^20\n\n\"AI and Memory Wall\" paper (https://github.com/amirgholami/ai_and_memory_wall) made an estimation of 250,000 PFLOPS = 2.5*10^20 FLOP"},{"id":"epoch:2a5a4d0836fdb16353a4:parameters","entityId":"epoch:2a5a4d0836fdb16353a4","metric":"parameters","value":22000000.0,"unit":"parameters","observationPeriod":"2018-10-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"This representational demand motivates the introduction of high-capacity deep neural networks. We found the architecture, shown in Figure 3, with residual connections, 20 convolution layers with 512 channels\nfor a total of 22 million parameters, and instance normalization to drastically improve performance, as shown in Figure 6 of the Experiments section.\""},{"id":"epoch:a96353b904c84531c372:parameters","entityId":"epoch:a96353b904c84531c372","metric":"parameters","value":112694781.0,"unit":"parameters","observationPeriod":"2018-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"I used the publicly available implementation available at [1]\n\nThere I loaded the biggan-deep512/1 model, and ran script [2] to compute the number of parameters\n\n[1] https://colab.research.google.com/github/tensorflow/hub/blob/master/examples/colab/biggan_generation_with_tf_hub.ipynb\n\n[2]\nn_params = 0\nfor var in module.variables:\n  n_params += np.prod(var.shape.as_list())\n  pass\n\nprint(n_params)"},{"id":"epoch:a96353b904c84531c372:training-compute","entityId":"epoch:a96353b904c84531c372","metric":"training-compute","value":1.8e+21,"unit":"FLOP","observationPeriod":"2018-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3e21, estimate taken from:\n\nhttps://www.lesswrong.com/posts/wfpdejMWog4vEDLDg/ai-and-compute-trend-isn-t-predictive-of-what-is-happening"},{"id":"epoch:af5609536d65aba7b296:parameters","entityId":"epoch:af5609536d65aba7b296","metric":"parameters","value":247000000.0,"unit":"parameters","observationPeriod":"2018-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:af5609536d65aba7b296:training-compute","entityId":"epoch:af5609536d65aba7b296","metric":"training-compute","value":4.47e+19,"unit":"FLOP","observationPeriod":"2018-09-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8 V100s * 67 hours per Table 2.\n125e12 FLOP/sec * 8 * 67 * 3600 * 0.3 (utilization assumption) = 7.2e19 FLOP\n\nThey also say they trained for 286k steps in batches of 65,536 tokens.\n6 * 247M * (286k * 65536) = 2.78e19\n\ngeometric mean: sqrt(7.2e19 * 2.78e19) = 4.47e19"},{"id":"epoch:dcd9367a3bdbe472bc36:parameters","entityId":"epoch:dcd9367a3bdbe472bc36","metric":"parameters","value":2100000.0,"unit":"parameters","observationPeriod":"2018-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Given:\n\nHidden size (H) = 512\nNumber of hidden layers (L) = 1\nInput size (I) is not mentioned, so let's denote it as I\nThe total number of parameters P in an LSTM can be calculated as follows:\n\nP = 4 * ((I * H) + (H * H) + H)\n\nThis is for one layer of LSTM cells. Since the LSTM model described has only one layer, we don't need to multiply by the number of layers.\n\nTo calculate the exact number of parameters, we would need to know the input size I. However, if you are looking for the number of parameters just within a single LSTM cell (assuming I is equal to H), then you can substitute I with H in the above formula:\n\nP = 4 * ((H * H) + (H * H) + H)\n= 4 * (2 * (H^2) + H)\n\nFor H = 512, this becomes:\n\nP = 4 * (2 * (512^2) + 512)\n= 4 * (2 * 262144 + 512)\n= 4 * (524288 + 512)\n= 4 * 524800\n\u2248 2,099,200"},{"id":"epoch:dcd9367a3bdbe472bc36:training-compute","entityId":"epoch:dcd9367a3bdbe472bc36","metric":"training-compute","value":982800000000000.0,"unit":"FLOP","observationPeriod":"2018-09-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 2100000 parameters * 2000000 tokens * 39 epochs = 9.828e+14 FLOP"},{"id":"epoch:2c827351cc7195ffa0e6:parameters","entityId":"epoch:2c827351cc7195ffa0e6","metric":"parameters","value":35000000.0,"unit":"parameters","observationPeriod":"2018-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:48a543818b34482312a8:parameters","entityId":"epoch:48a543818b34482312a8","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2018-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0667dc9e112246957a52:parameters","entityId":"epoch:0667dc9e112246957a52","metric":"parameters","value":90000000.0,"unit":"parameters","observationPeriod":"2018-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"5-layer model, Table 3"},{"id":"epoch:0667dc9e112246957a52:training-compute","entityId":"epoch:0667dc9e112246957a52","metric":"training-compute","value":1.1e+19,"unit":"FLOP","observationPeriod":"2018-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use a single NVIDIA Tesla V100 GPU for each model. The published results were obtained\nusing 8 GPUs in parallel, which provide a large effective batch size during training. To approximate\nthe setup, we update the model parameters every 5\u00d75120 tokens and use 16,000 warm-up steps\nfollowing OpenNMT suggestions. We train each\nmodel for 40 epochs (250,000 steps), and perform\n3 independent trials for each model configuration.\nA single run takes about 3.5 days with a Tesla V100 GPU.\"\n\n125 trillion * 3.5 * 24 * 3600 * 0.3 = 1.1e19"},{"id":"epoch:452aede1a13c8c423038:parameters","entityId":"epoch:452aede1a13c8c423038","metric":"parameters","value":185000000.0,"unit":"parameters","observationPeriod":"2018-08-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"185M (table 8)"},{"id":"epoch:452aede1a13c8c423038:training-compute","entityId":"epoch:452aede1a13c8c423038","metric":"training-compute","value":6.66e+17,"unit":"FLOP","observationPeriod":"2018-08-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 185000000 parameters * 2000000 tokens * 300 epochs = 6.66e+17 FLOP"},{"id":"epoch:59ef49dbedabff68de67:parameters","entityId":"epoch:59ef49dbedabff68de67","metric":"parameters","value":37000000.0,"unit":"parameters","observationPeriod":"2018-08-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5a6be7d1281c06050c8e:parameters","entityId":"epoch:5a6be7d1281c06050c8e","metric":"parameters","value":23000000.0,"unit":"parameters","observationPeriod":"2018-08-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:5a6be7d1281c06050c8e:training-compute","entityId":"epoch:5a6be7d1281c06050c8e","metric":"training-compute","value":4.323e+16,"unit":"FLOP","observationPeriod":"2018-08-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From Figure 2, 300 epochs on PTB, which has 1044112 tokens.\n\n6 * 23M * 1044112 * 300 = 4.323e16"},{"id":"epoch:6789e25c244233b18a61:training-compute","entityId":"epoch:6789e25c244233b18a61","metric":"training-compute","value":4.7808e+20,"unit":"FLOP","observationPeriod":"2018-08-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(128) * (1.25e14) * (27*3600 + 40*60) * (0.3)  = 4.7808e20\n(number of gpus) * (peak flops) * (seconds) * (assumed utilization rate)  \n\n\"We run experiments on DGX-1 machines with 8Nvidia V100 GPUs and machines are interconnected by Infiniband. Experiments are run on 16\nmachines and we perform 30K synchronous updates.\"\n\"We also use the NCCL2 library [...] with 16-bit floating point\noperations\"\n\nNCCL2 supported tensor core operations at 1.25e14 FLOP/s on a V100 for FP16 \n\nin section 5.6 we have\n\n\"train this system we perform 300K training up-\ndates in 27h 40min on 128 GPUs;\""},{"id":"epoch:9516e3fd2f2e20a21991:parameters","entityId":"epoch:9516e3fd2f2e20a21991","metric":"parameters","value":35000000.0,"unit":"parameters","observationPeriod":"2018-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:069e0079bb118ca4de67:parameters","entityId":"epoch:069e0079bb118ca4de67","metric":"parameters","value":60852800.0,"unit":"parameters","observationPeriod":"2018-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"10k vocabulary\n\"For the mixture-of-softmax model, we use a 3-layer LSTM with dimensions 960, 960 and 620, embedding dimension of 280 and 15 experts for PTB\"\n\nEmbedding: 10000*280=2800000\nUnembedding: 10000*280*15=42000000\nLSTM 1: 4*(280+960)*960=4761600\nLSTM 2: 4*(960+960)*960=7372800\nLSTM 3: 4*(960+620)*620=3918400\n\nTotal: 2800000+42000000+4761600+7372800+3918400=60852800"},{"id":"epoch:069e0079bb118ca4de67:training-compute","entityId":"epoch:069e0079bb118ca4de67","metric":"training-compute","value":4.07e+17,"unit":"FLOP","observationPeriod":"2018-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Forward flop: 2*60852800=121705600\n\nFLOP: 1200*121705600*3*929000=4.0703221e+17"},{"id":"epoch:3b3cb27b7e9e853c29a7:parameters","entityId":"epoch:3b3cb27b7e9e853c29a7","metric":"parameters","value":209000000.0,"unit":"parameters","observationPeriod":"2018-08-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b2c66595b10027757a77:parameters","entityId":"epoch:b2c66595b10027757a77","metric":"parameters","value":53477376.0,"unit":"parameters","observationPeriod":"2018-08-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1: 204MB parameters. \n\nAssuming they used FP32: 204*1024*1024/4=53477376"},{"id":"epoch:b2c66595b10027757a77:training-compute","entityId":"epoch:b2c66595b10027757a77","metric":"training-compute","value":1.19e+16,"unit":"FLOP","observationPeriod":"2018-08-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*53477376*929000*40=1.1923316e+16"},{"id":"epoch:d7f7d6308943a2a4f48e:parameters","entityId":"epoch:d7f7d6308943a2a4f48e","metric":"parameters","value":3181588.0,"unit":"parameters","observationPeriod":"2018-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Figure 2 - Deployed model (D) consists of a Control Policy Network (B) and Vision Network (C), trained separately. (Training of (B) also includes training of a Value Network that is, ignoring input-output size differences, identical in architecture to that of the control policy; the value network is not included during deployment).\n\nControl Policy Network: 3181588 (23552 + 3147776 + 10260)\nsee Table 2, Figure 12, Table 10\ngiven:\n- input: 22 out [Table 2, 15+3+4]\n- normalization: 22 out\n- FC-RELU: 1024 out [Table 10, assume bias]\n- LSTM: 512 out [Table 10, assume bias]\n- FC (implied?): 20 out [Appendix C.1, Actions]\nparameters:\n- FC-RELU: 23552 [(22+1)*1024]\n- normalization: 0 [ignored]\n- LSTM: 3147776 [4*(1024+512+1)*512]\n- FC (Implied?): 10260 [(512+1)*20]"},{"id":"epoch:d7f7d6308943a2a4f48e:training-compute","entityId":"epoch:d7f7d6308943a2a4f48e","metric":"training-compute","value":2.16e+20,"unit":"FLOP","observationPeriod":"2018-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"In experiments, runs were typically conducted for 50 hours of wall time, with up to 32 GPUs. Table 10 specifies V100 GPUs for training the policy.\nAssume V100 base variant, BF16, 0.3 utilization\nV100 tensor FP16 performance = 1.25e14 FLOP/s\n\n32 GPUs * (0.3 * 1.25e14 FLOP/s/GPU) * (3600 s/1 hr) * 50 hr  = 2.16e+20 FLOP"},{"id":"epoch:d2156ebd2833f044142b:parameters","entityId":"epoch:d2156ebd2833f044142b","metric":"parameters","value":77360000.0,"unit":"parameters","observationPeriod":"2018-07-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:d2156ebd2833f044142b:training-compute","entityId":"epoch:d2156ebd2833f044142b","metric":"training-compute","value":2.46048e+17,"unit":"FLOP","observationPeriod":"2018-07-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Using the 6ND formula: \n6\u00d7number of tokens\u00d7number of parameters\u00d7number of epochs\n6\u00d71.28\u00d710^6\u00d777360000\u00d7110=6.5353728e+16 FLOPs\n\n9.32*10^9 (flops per inference)*1.28\u00d710^6(dataset size)/16 (batch size) * 110 epochs * 3 (to account for backpropagation)= 2.46048e+17 FLOPs"},{"id":"epoch:3ebff45ed346fbdbf0ad:parameters","entityId":"epoch:3ebff45ed346fbdbf0ad","metric":"parameters","value":77360000.0,"unit":"parameters","observationPeriod":"2018-07-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2 - fifth row"},{"id":"epoch:3ebff45ed346fbdbf0ad:training-compute","entityId":"epoch:3ebff45ed346fbdbf0ad","metric":"training-compute","value":6.2988288e+19,"unit":"FLOP","observationPeriod":"2018-07-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"number of epochs (appendix A1) times flops per inference (from table 2) times dataset size times 3 (to account for backpropagation)\n110 * 9.32e9 FLOPs * 256/16 * 1280000 * 3 = 6.3e19"},{"id":"epoch:3ae59e24ef8be1b5607b:parameters","entityId":"epoch:3ae59e24ef8be1b5607b","metric":"parameters","value":3320000.0,"unit":"parameters","observationPeriod":"2018-07-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"table 3"},{"id":"epoch:3ae59e24ef8be1b5607b:training-compute","entityId":"epoch:3ae59e24ef8be1b5607b","metric":"training-compute","value":4.290048e+17,"unit":"FLOP","observationPeriod":"2018-07-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"980000000 (number of FLOPs from table 3) * 27360000 (dataset size) * 16 (number of epochs from appendix B.1) = 429004800000000000"},{"id":"epoch:d43ed604f5e9d91922d7:parameters","entityId":"epoch:d43ed604f5e9d91922d7","metric":"parameters","value":16000000.0,"unit":"parameters","observationPeriod":"2018-07-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"EDSR has much larger number of parameters (43 M) than ours\n(16 M), but our RCAN obtains much better performance.\""},{"id":"epoch:1a0da3a6322109d8f380:parameters","entityId":"epoch:1a0da3a6322109d8f380","metric":"parameters","value":126001330.0,"unit":"parameters","observationPeriod":"2018-07-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Architecture described in figure S11 of the supplement\n\nThe architecture includes modules for visual embedding, reward prediction, recurrent processing, policy, baseline and pixel control.\n\nInput is 84x84x3 pixels as seen in figure S10 of the supplement\n\n\"We elected to use a resolution of 84x84 pixels as in previous related work in this environment. Each pixel is represented by a triple of three bytes\"\n\nVisual embedding (84x84x3 -> 256)\n32*(8*8*3+1)+64*(4*4*32+1)+64*(3*3*64+1)+64*(3*3*64+1) + (84/(S^4)*84/(S^4)*64+1)*256\nNote there is no information about the stride S used in the convolutions; we assume S = 1\n\nReward prediction (256 -> 3)\n(256+1)*128 + (128+1)*3\n\nRecurrent processing (n-> 512)\nVU1 (256 -> 512)\n4*(799+2*32)*((512+(32*2) + 3*32 + 5*2 + 3)+(799+2*32)+1) + 2*(256+1)*256\n\nVU2 (512 -> 512)\n4*(512+2*32)*((512+(32*2) + 3*32 + 5*2 + 3)+(512+2*32)+1) + 2*(256+1)*256\n\nLSTMs usually have 4*(n*m+n*n+n) parameters, where n=input size and m=output size.\n\nThis DNS + LSTM takes as input the concatenation of the previous layer of size n and R read vectors of size W=32; and outputs m units plus an interface vector of size (W*R) + 3*W + 5*R + 3, for a total of about 4*(n+R*W)*((m+(W*R) + 3*W + 5*R + 3)+(n+R*32)+1) parameters\n\nI assume R=2 since that seems implied by the previous paper (?)\n\nThe first VU has as input the visual embedding (size 256), the previous action (size 540) and the previous reward (size 3), for a total size of 256+540+3 = 799. The output is size 512.\n\nThe second VU has input size 512 and output size 512\n\nThe DNC memory architecture is described in https://www.nature.com/articles/nature20101.epdf\n\nPolicy (512 -> 5x3x3x3x2x2)\n6*(512+1)*256 + (256+1)*5 + 3*(256+1)*3 + 2*(256+1)*2\n\nBaseline\n(512+1)*256 + (256+1)*1\n\nPixel control\n(512+1)*32*7*7 + 32*(9*9+1) + 5*(4*4+1) + 3*2*(4*4+1) + 2*2*(4*4+1) + 1*(4*4+1)\n\"we trained independent pixel control policies for each of the six action groups\""},{"id":"epoch:1a0da3a6322109d8f380:training-compute","entityId":"epoch:1a0da3a6322109d8f380","metric":"training-compute","value":3.49e+19,"unit":"FLOP","observationPeriod":"2018-07-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Source: \nhttps://docs.google.com/spreadsheets/d/1Kj4Q5WADcDXtUJLIOfGTCE3tGvxNczEMwyy8QtgSkHk/edit#gid=54587040&fvid=1361937389\n\n\nAdditional estimate based on parameters (low confidence, had to make some assumptions about the architecture)\nForward pass operations\nVision\n84*84*32*8*8*3+20*20*32*64*4*4+7*7*64*64*3*3+7*7*64*64*3*3+7*7*64*256=60874752\n60874752*2=121749504\nPixel control (guess on the internal dimensions)\n256*32*7*7+7*7*32*32*9*9+20*20*32*4*4*(1+2+3+4+5)=7537664\nRemaining 2*(463616+33152+1310720)=3614976\nTotal: 121749504+7537664+3614976\nTraining FLOP estimate: 3*3614976*2000000000=21689856000000000\n"},{"id":"epoch:2e6ead926dfa8116272c:parameters","entityId":"epoch:2e6ead926dfa8116272c","metric":"parameters","value":2280000.0,"unit":"parameters","observationPeriod":"2018-06-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ee9a7d28b9b821e72273:parameters","entityId":"epoch:ee9a7d28b9b821e72273","metric":"parameters","value":1200000.0,"unit":"parameters","observationPeriod":"2018-06-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The Q-function Q\u03b8(s, a) is represented in our system by a large convolutional neural network with 1.2M parameters\""},{"id":"epoch:ee9a7d28b9b821e72273:training-compute","entityId":"epoch:ee9a7d28b9b821e72273","metric":"training-compute","value":1.395e+19,"unit":"FLOP","observationPeriod":"2018-06-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We distribute training across 10 GPUs, using asynchronous SGD with momentum... This system allows us to train the Q-function at 40 steps per second with a batch size of 32 across 10 NVIDIA P100 GPUs.\"\n\n\"We found empirically that a large number of gradient steps (up to 15M) were needed to train an effective Q-function...\"\n\n15M steps * 0.025 seconds/step *  9.30E+12 FLOP/sec/GPU * 0.4 utilization * 10 GPU = 1.395E+19"},{"id":"epoch:fa3cea664e437835a811:parameters","entityId":"epoch:fa3cea664e437835a811","metric":"parameters","value":33000000.0,"unit":"parameters","observationPeriod":"2018-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"33M (Table 4) - parameters reported for PTB, but they say it is the same for WT-2\n\n\" WIKITEXT-2\nWe use embedding and hidden sizes 700, weight decay 5\u00d710\u22127, and hidden-node variational dropout 0.15. Other hyperparameters remain the same as in our PTB experiments.\""},{"id":"epoch:fa3cea664e437835a811:training-compute","entityId":"epoch:fa3cea664e437835a811","metric":"training-compute","value":3.2366286e+17,"unit":"FLOP","observationPeriod":"2018-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 33000000 parameters * 2000000 tokens * 300 epochs = 1.188e+17 FLOP\n\n11340000000000 FLOP / second / GPU * 1 GPU * 72 hours * 3600 sec / hour * 0.3 [assumed utilization] = 8.817984e+17 FLOP\n\nsqrt(1.188e+17*8.817984e+17) = 3.2366286e+17 FLOP\n\n'Speculative' confidence since many variables are assumed based on PTB model training"},{"id":"epoch:117e237bc419baa69894:parameters","entityId":"epoch:117e237bc419baa69894","metric":"parameters","value":23000000.0,"unit":"parameters","observationPeriod":"2018-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:282903beaf9dcbc20554:parameters","entityId":"epoch:282903beaf9dcbc20554","metric":"parameters","value":3400000.0,"unit":"parameters","observationPeriod":"2018-06-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Rados"},{"id":"epoch:faa78fbb2a4eb20edede:parameters","entityId":"epoch:faa78fbb2a4eb20edede","metric":"parameters","value":117000000.0,"unit":"parameters","observationPeriod":"2018-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The model had 117M parameters in total.\"\n\nsource: https://medium.com/walmartglobaltech/the-journey-of-open-ai-gpt-models-32d95b7b7fb2"},{"id":"epoch:faa78fbb2a4eb20edede:training-compute","entityId":"epoch:faa78fbb2a4eb20edede","metric":"training-compute","value":1.7578125e+19,"unit":"FLOP","observationPeriod":"2018-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"COMPUTE = FORWARD COMPUTE PER TOKEN * 3 BACKWARD FORWARD ADJUSTMENT * EPOCHS * DATASET SIZE\n\n\"We train for 100 epochs on minibatches of 64 randomly sampled, contiguous sequences of 512 tokens.\"\n\nAuthors of \"AI and Memory Wall\" estimated model's training compute as 57,000 PFLOPS = 5.7*10^19 FLOP\n(https://github.com/amirgholami/ai_and_memory_wall)"},{"id":"epoch:4cf462861343e282db83:parameters","entityId":"epoch:4cf462861343e282db83","metric":"parameters","value":5400000.0,"unit":"parameters","observationPeriod":"2018-05-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1be045109d641ba31d80:parameters","entityId":"epoch:1be045109d641ba31d80","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2018-05-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f268c684c4057dbb7097:parameters","entityId":"epoch:f268c684c4057dbb7097","metric":"parameters","value":32000000.0,"unit":"parameters","observationPeriod":"2018-05-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32M (Table 3)"},{"id":"epoch:f268c684c4057dbb7097:training-compute","entityId":"epoch:f268c684c4057dbb7097","metric":"training-compute","value":7.296e+16,"unit":"FLOP","observationPeriod":"2018-05-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 32000000 parameters * 2000000 tokens * 190 epochs = 7.296e+16 FLOP"},{"id":"epoch:b0b898486ff8a0768f11:parameters","entityId":"epoch:b0b898486ff8a0768f11","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2018-05-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b0b898486ff8a0768f11:training-compute","entityId":"epoch:b0b898486ff8a0768f11","metric":"training-compute","value":2.4e+16,"unit":"FLOP","observationPeriod":"2018-05-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*24000000*929000*180=2.407968e16"},{"id":"epoch:910d334a0cad216a84af:parameters","entityId":"epoch:910d334a0cad216a84af","metric":"parameters","value":51000000.0,"unit":"parameters","observationPeriod":"2018-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The large network has 2 layers with 1500 hidden units each. This leads to a model complexity of 51 million parameters.\""},{"id":"epoch:910d334a0cad216a84af:training-compute","entityId":"epoch:910d334a0cad216a84af","metric":"training-compute","value":1.27e+17,"unit":"FLOP","observationPeriod":"2018-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 51000000 parameters * 2000000 tokens * 200 epochs = 1.224e+17 FLOP\n\n'Likely' confidence because I am not very sure that 51M paramters and 200 epochs relate to WT-2 model, but it is very likely"},{"id":"epoch:e8d39d4154c9a36d9c2f:parameters","entityId":"epoch:e8d39d4154c9a36d9c2f","metric":"parameters","value":51000000.0,"unit":"parameters","observationPeriod":"2018-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e8d39d4154c9a36d9c2f:training-compute","entityId":"epoch:e8d39d4154c9a36d9c2f","metric":"training-compute","value":1.27e+17,"unit":"FLOP","observationPeriod":"2018-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"6 FLOP / parameter / token * 51000000 parameters * 2000000 tokens * 200 epochs = 1.224e+17 FLOP\n\n'Likely' confidence because I am not very sure that 51M paramters and 200 epochs relate to WT-2 model, but it is very likely\""},{"id":"epoch:f7bfa7a17f01e1e19d43:parameters","entityId":"epoch:f7bfa7a17f01e1e19d43","metric":"parameters","value":22000000.0,"unit":"parameters","observationPeriod":"2018-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f7bfa7a17f01e1e19d43:training-compute","entityId":"epoch:f7bfa7a17f01e1e19d43","metric":"training-compute","value":4.9e+16,"unit":"FLOP","observationPeriod":"2018-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 22000000 parameters * 929000 tokens * 400 epochs = 4.90512e+16 FLOP"},{"id":"epoch:0bd8282d8b03c5aa8f7b:parameters","entityId":"epoch:0bd8282d8b03c5aa8f7b","metric":"parameters","value":13000000.0,"unit":"parameters","observationPeriod":"2018-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0bd8282d8b03c5aa8f7b:training-compute","entityId":"epoch:0bd8282d8b03c5aa8f7b","metric":"training-compute","value":3.12e+16,"unit":"FLOP","observationPeriod":"2018-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 13000000 parameters * 2000000 tokens * 200 epochs = 3.12e+16 FLOP"},{"id":"epoch:f3396e0d03ab2bb203d0:parameters","entityId":"epoch:f3396e0d03ab2bb203d0","metric":"parameters","value":51000000.0,"unit":"parameters","observationPeriod":"2018-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f3396e0d03ab2bb203d0:training-compute","entityId":"epoch:f3396e0d03ab2bb203d0","metric":"training-compute","value":5.68e+16,"unit":"FLOP","observationPeriod":"2018-05-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / paramter / token * 51000000 parameters * 929000 tokens * 200 epochs = 5.68548e+16 FLOP"},{"id":"epoch:168073ee02adc16c5f13:parameters","entityId":"epoch:168073ee02adc16c5f13","metric":"parameters","value":829000000.0,"unit":"parameters","observationPeriod":"2018-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 6\n"},{"id":"epoch:168073ee02adc16c5f13:training-compute","entityId":"epoch:168073ee02adc16c5f13","metric":"training-compute","value":8.74395e+21,"unit":"FLOP","observationPeriod":"2018-05-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 6: 153e9 mult-adds.\nSection 2.4: \"minibatches of 8,064 images\".\n\nCompute = 2 * 3 * mult-adds * dataset size = 2 * 3 * 153e9 * 9525e6 = 8.74e21 FLOP\n\nLikely trained on V100s, since Facebook had just upgraded their Big Basin GPU cluster to V100s as of March 2018. The previous iteration of Big Basin had 32 clusters of 8xP100s, while Big Basin v2 had 42 clusters of 8xV100s, which matches the 336 GPUs used in this paper."},{"id":"epoch:e27ef69721de55a448fa:training-compute","entityId":"epoch:e27ef69721de55a448fa","metric":"training-compute","value":9.5e+16,"unit":"FLOP","observationPeriod":"2018-05-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our training was conducted on Tesla K20 Nvidia GPUs each having 5 GB of GPU memory, on which, training one model took around 12\u2009h.\"\n\n\"We train each CNN for a total of 1600 epochs with each epoch of training taking around 2\u2009min.\"\n\nAssumptions:\npeakFLOP rate 3.52e12FLOP/s (from: https://www.techpowerup.com/gpu-specs/tesla-k20c.c564)\n30% utilization rate\n1 GPU\n\nEstimate 1: \"training one model took around 12h\" => unclear how many GPUs\n(12 *3600) s * 3.52e12 FLOP/s * 0.3 = 4.5e16 FLOP\n\nEstimate 2: \"We train each CNN for a total of 1600 epochs with each epoch of training taking around 2 min.\"\n(1600 epochs * 2 min/epoch * 60 s/min) * 3.52e12 FLOP/s * 0.3 =  2e17 FLOP\n\nGeometric mean: 9.5e16"},{"id":"epoch:19bc00d153a3a8d41c5d:parameters","entityId":"epoch:19bc00d153a3a8d41c5d","metric":"parameters","value":87600000.0,"unit":"parameters","observationPeriod":"2018-04-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Calculatable - 2-layer LSTM with 650 hidden units."},{"id":"epoch:19bc00d153a3a8d41c5d:training-compute","entityId":"epoch:19bc00d153a3a8d41c5d","metric":"training-compute","value":2.1024e+16,"unit":"FLOP","observationPeriod":"2018-04-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 87600000 parameters * 2000000 tokens * 20 epochs = 2.1024e+16 FLOP"},{"id":"epoch:b66fbf855d6d59af522b:parameters","entityId":"epoch:b66fbf855d6d59af522b","metric":"parameters","value":378900000.0,"unit":"parameters","observationPeriod":"2018-04-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from Table 3 RNMT+ "},{"id":"epoch:b66fbf855d6d59af522b:training-compute","entityId":"epoch:b66fbf855d6d59af522b","metric":"training-compute","value":1.83e+19,"unit":"FLOP","observationPeriod":"2018-04-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"32 * 9.3 TFLOPS * (120 * 3600) * 0.3 = 3.86e19\n(number of gpus) * (peak flops) * (seconds) * (assumed utilization rate)\n\"All models were trained with synchronous training. RNMT+ and ConvS2S were trained with 32 NVIDIA P100 GPUs\", training time is 120h from Table 1.\n\nAlternatively, typical inference is 2.81e10 FLOP, there are 36.3M examples in the dataset, and they train for 8.5 epochs: 2.81e10 * 36.3M * 8.5 = 8.7e18\n\nGeometric mean: sqrt(3.86e19 * 8.7e18) = 1.83e19"},{"id":"epoch:0fd272f1e56e30d85626:parameters","entityId":"epoch:0fd272f1e56e30d85626","metric":"parameters","value":8000000000.0,"unit":"parameters","observationPeriod":"2018-04-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For example, using five 3D-printed transmission layers, containing a total of 0.2 million neurons and ~8.0 billion connections that are trained using deep learning, we experimentally demonstrated the function of a handwritten digit classifier.\"\n\nMy understanding is that every connection correspond to the parameter to learn."},{"id":"epoch:2c3ea1941d194fa36bab:parameters","entityId":"epoch:2c3ea1941d194fa36bab","metric":"parameters","value":56933216.0,"unit":"parameters","observationPeriod":"2018-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Feature extractor (ignoring biases)\n32*3*3*3 +\n64*3*3*32 +\n32*1*1*64 +\n64*3*3*32 +\n128*3*3*64 +\n2*(64*1*1*128 +\n128*3*3*64) +\n256*3*3*128 +\n8*(128*1*1*256 +\n256*3*3*128) +\n512*3*3*256 + \n8*(256*1*1*512 + \n512*3*3*256) + \n1024*3*3*512 + \n4*(512*1*1*1024 +\n1024*3*3*512) +\n4*4*1024*1000\n\nsource: table 1\nThis is assuming the average pooling step changes the output size from 8x8 to 4x4.\n\nThe weights file is 237MB. If the weights are saved as float32, 4 bytes per weight, then there are approximately 237M/4=59M parameters, consistent with the calculation above."},{"id":"epoch:2c3ea1941d194fa36bab:training-compute","entityId":"epoch:2c3ea1941d194fa36bab","metric":"training-compute","value":1.3416380824e+19,"unit":"FLOP","observationPeriod":"2018-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"We use the formula training_compute = ops_per_forward_pass * 3.5 * n_epochs * n_examples\n\nAssuming 160 epochs of training as in https://arxiv.org/pdf/1612.08242.pdf\n\nTable 2: 18700000000 operations \n\n18700000000 ops * 3.5 *160 epochs * 1281167 images"},{"id":"epoch:48debf81cbb20d8861ab:parameters","entityId":"epoch:48debf81cbb20d8861ab","metric":"parameters","value":530442240.0,"unit":"parameters","observationPeriod":"2018-03-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Single layer LSTM with hidden dimension of 2048. Vocabulary for Gutenberg is 242,621; input and output embeddings are tied.\n\nEmbedding layer (tied): 242,621 * 2048 = 496,887,808\nLSTM layer: 4 * (2048 + 2048) * 2048 = 33,554,432\n\nTotal: 496,887,808 + 33,554,432 = 530,442,240"},{"id":"epoch:48debf81cbb20d8861ab:training-compute","entityId":"epoch:48debf81cbb20d8861ab","metric":"training-compute","value":3.33e+19,"unit":"FLOP","observationPeriod":"2018-03-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"They do training runs on a vision task and three language datasets. The largest dataset by size is GigaWord, but the largest training run is on the Gutenberg dataset, at 15B tokens. \n\nI assume the input embedding is done with an embedding lookup for efficiency rather than a dense matrix multiplication, so we only count FLOPs on the de-embedding.\n\nOps counting:\n6 * 15B * 530,442,240 = 4.774e19\n\nHardware:\n(8 * 1.87e13) * (6 * 24 * 3600) * 0.3 = 2.327e19\n\nGeometric mean: sqrt(4.774e19 * 2.327e19) = 3.33e19\n\n(Note they say 15B steps, but they also say it is 80 epochs on a 175M token dataset, and that it took 6 days on 8 P100s, both of which would agree with 15B tokens, not 15B steps)"},{"id":"epoch:93858c3ee35d306e4714:parameters","entityId":"epoch:93858c3ee35d306e4714","metric":"parameters","value":151000000.0,"unit":"parameters","observationPeriod":"2018-03-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 6"},{"id":"epoch:93858c3ee35d306e4714:training-compute","entityId":"epoch:93858c3ee35d306e4714","metric":"training-compute","value":5.9158815e+17,"unit":"FLOP","observationPeriod":"2018-03-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"20670000000000 FLOP / sec / GPU [fp16 assumed] * 1 GPU * 12 hours * 3600 sec / hour * 0.3 [assumed utilization] = 2.678832e+17FLOP\n\n6 FLOP / token / parameter * 151000000 parameters * 103000000 tokens * 14 epochs = 1.306452e+18 FLOP\n\nsqrt(2.678832e+17*1.306452e+18) = 5.9158815e+17 FLOP\n__________________\nin the algorithmic progress paper the estimation was 2.4 \u00d7 10^17 FLOP under assumption of 26M parameters"},{"id":"epoch:8faca20695177af5f8f6:parameters","entityId":"epoch:8faca20695177af5f8f6","metric":"parameters","value":86000000.0,"unit":"parameters","observationPeriod":"2018-03-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://openai.com/blog/image-gpt/#rfref53"},{"id":"epoch:879d29558b987e52fc61:parameters","entityId":"epoch:879d29558b987e52fc61","metric":"parameters","value":13000000.0,"unit":"parameters","observationPeriod":"2018-03-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c9c3a0f9a0aec286e09b:parameters","entityId":"epoch:c9c3a0f9a0aec286e09b","metric":"parameters","value":26000000.0,"unit":"parameters","observationPeriod":"2018-02-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d12c484c2b3c46f1a549:parameters","entityId":"epoch:d12c484c2b3c46f1a549","metric":"parameters","value":148000000.0,"unit":"parameters","observationPeriod":"2018-02-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:8194a17e6b50071344aa:parameters","entityId":"epoch:8194a17e6b50071344aa","metric":"parameters","value":13000000.0,"unit":"parameters","observationPeriod":"2018-02-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:87edd9096cb36a70ee9f:parameters","entityId":"epoch:87edd9096cb36a70ee9f","metric":"parameters","value":40000000.0,"unit":"parameters","observationPeriod":"2018-02-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:b46b42187042fecdfcf7:parameters","entityId":"epoch:b46b42187042fecdfcf7","metric":"parameters","value":42000.0,"unit":"parameters","observationPeriod":"2018-02-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:12e375f0c05fe84297ef:parameters","entityId":"epoch:12e375f0c05fe84297ef","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2018-02-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"24M\nTable 1"},{"id":"epoch:12e375f0c05fe84297ef:training-compute","entityId":"epoch:12e375f0c05fe84297ef","metric":"training-compute","value":2.00664e+16,"unit":"FLOP","observationPeriod":"2018-02-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training on PTB:\n6 FLOP / token / parameter * 24000000 parameters * 929000 tokens * 150 epochs = 2.00664e+16 FLOP"},{"id":"epoch:7b8433e86a7553910dbd:parameters","entityId":"epoch:7b8433e86a7553910dbd","metric":"parameters","value":469000000.0,"unit":"parameters","observationPeriod":"2018-02-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:7b8433e86a7553910dbd:training-compute","entityId":"epoch:7b8433e86a7553910dbd","metric":"training-compute","value":3.85296912e+20,"unit":"FLOP","observationPeriod":"2018-02-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"450 K40 GPUs for 20k models (approx. 7 days).\n(From Imagenet paper-data, Besiroglu et al., forthcoming) "},{"id":"epoch:30fb47e28a139feaf62f:parameters","entityId":"epoch:30fb47e28a139feaf62f","metric":"parameters","value":87000000.0,"unit":"parameters","observationPeriod":"2018-02-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:3301ace86ea82da753fc:parameters","entityId":"epoch:3301ace86ea82da753fc","metric":"parameters","value":1600000.0,"unit":"parameters","observationPeriod":"2018-02-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Figure 3 in the paper states that the large architecture has 1.6 million parameters. I am using the large model because it was the only one trained on all the Atari games at once, which seems like the most impressive task in the suite.\"\n\nSource: https://docs.google.com/spreadsheets/d/1Kj4Q5WADcDXtUJLIOfGTCE3tGvxNczEMwyy8QtgSkHk/edit#gid=54587040&fvid=1361937389"},{"id":"epoch:3301ace86ea82da753fc:training-compute","entityId":"epoch:3301ace86ea82da753fc","metric":"training-compute","value":1.68e+20,"unit":"FLOP","observationPeriod":"2018-02-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Source: Ajeya Cotra and Tom Davidson, https://docs.google.com/spreadsheets/d/1Kj4Q5WADcDXtUJLIOfGTCE3tGvxNczEMwyy8QtgSkHk/edit#gid=54587040&fvid=1361937389"},{"id":"epoch:7a2c9b65e1e6082b4210:parameters","entityId":"epoch:7a2c9b65e1e6082b4210","metric":"parameters","value":94000000.0,"unit":"parameters","observationPeriod":"2018-02-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:7a2c9b65e1e6082b4210:training-compute","entityId":"epoch:7a2c9b65e1e6082b4210","metric":"training-compute","value":3300100000000010.0,"unit":"FLOP","observationPeriod":"2018-02-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3300e12 - https://github.com/amirgholami/ai_and_memory_wall"},{"id":"epoch:b64ee40b30d409bb0736:parameters","entityId":"epoch:b64ee40b30d409bb0736","metric":"parameters","value":135000000.0,"unit":"parameters","observationPeriod":"2018-02-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Based on the details provided in the paper, the number of parameters in the QRNN model can be calculated as follows:\n\nEmbedding layer size: 400\nNumber of QRNN layers: 4\nNumber of nodes per QRNN layer: 2500\nVocabulary size: 267,735 (for WikiText-103 dataset)\nAdaptive softmax layer\nParameters:\n\nEmbedding layer: 400 x 267,735 = 107,094,000\nQRNN layers:\nInput to hidden weights per layer: 400 x 2500 = 1,000,000\nHidden to hidden weights per layer: 2500 x 2500 = 6,250,000\nBiases per layer: 2500 = 2,500\nTotal QRNN layers parameters: 4 x (1,000,000 + 6,250,000 + 2,500) = 28,000,000\nAdaptive softmax layer: No extra parameters due to weight tying\nTotal Parameters = Embedding + QRNN layers\n= 107,094,000 + 28,000,000\n= 135,094,000\n\nSo the total number of parameters in the QRNN model is approximately 135 million.\n\nThe majority of parameters are in the embedding layer, while the 4 QRNN layers contribute 28 million parameters. The adaptive softmax does not add any extra parameters due to weight tying."},{"id":"epoch:b64ee40b30d409bb0736:training-compute","entityId":"epoch:b64ee40b30d409bb0736","metric":"training-compute","value":6.8866472e+17,"unit":"FLOP","observationPeriod":"2018-02-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 135000000 parameters ['Likely' confidence] * 103000000 tokens * 14 epochs = 1.16802e+18 FLOP\n\n31330000000000 FLOP / sec / GPU [fp16 assumed] * 1 GPU * 12 hours * 3600 sec / hour * 0.3 [assumed utilization] = 4.060368e+17 FLOP \n\nsqrt(4.060368e+17*1.16802e+18) = 6.8866472e+17 FLOP \n"},{"id":"epoch:e0a764339e9c91920f56:parameters","entityId":"epoch:e0a764339e9c91920f56","metric":"parameters","value":20000000.0,"unit":"parameters","observationPeriod":"2018-01-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 3"},{"id":"epoch:306e5d1dd7296c16e462:parameters","entityId":"epoch:306e5d1dd7296c16e462","metric":"parameters","value":441000000.0,"unit":"parameters","observationPeriod":"2018-01-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://files.fast.ai/models/wt103/?C=S;O=D"},{"id":"epoch:306e5d1dd7296c16e462:training-compute","entityId":"epoch:306e5d1dd7296c16e462","metric":"training-compute","value":2.72538e+17,"unit":"FLOP","observationPeriod":"2018-01-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"=103000000*441000000*6=2.72538e+17"},{"id":"epoch:3803522a6259645af765:training-compute","entityId":"epoch:3803522a6259645af765","metric":"training-compute","value":2.6244e+16,"unit":"FLOP","observationPeriod":"2018-01-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12150000000000*3600*2*0.3=2.6244e+16 FLOP"},{"id":"epoch:c9226072a25e187d7d0e:parameters","entityId":"epoch:c9226072a25e187d7d0e","metric":"parameters","value":13275200.0,"unit":"parameters","observationPeriod":"2018-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2-layer LSTM, 650 hidden units per layer, embedding size = 650. Penn Treebank standard vocab = ~10k words (from Zaremba et al. 2014)\n\n2 * (4 * (650*650 + 650*650 + 650)) + (650*10000)+10000 = 13275200"},{"id":"epoch:c9226072a25e187d7d0e:training-compute","entityId":"epoch:c9226072a25e187d7d0e","metric":"training-compute","value":502900000000000.0,"unit":"FLOP","observationPeriod":"2018-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"((8*650*(650+650))*2)*2 * 930000 tokens/epoch * 20 epochs = 5.029e14. batch size assumed from prior literature"},{"id":"epoch:280e5d74a3be4eba1b46:training-compute","entityId":"epoch:280e5d74a3be4eba1b46","metric":"training-compute","value":1.0600000001e+20,"unit":"FLOP","observationPeriod":"2017-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The Go version (there were versions trained individually for three different games) was trained for 34 hours using\n5k TPUv1 for data generation\n64 TPUv2 for parameter updating\nI used the PCD chip entries for the lowest precision numerical format we had for each, which was INT8 for TPUv1 and FP16 for TPUv2\n\nDirect training compute: 1.06e20 FLOPs"},{"id":"epoch:c503f3b8dc463b6a7be4:parameters","entityId":"epoch:c503f3b8dc463b6a7be4","metric":"parameters","value":6020000.0,"unit":"parameters","observationPeriod":"2017-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Here is a summary of the calculations to determine the number of parameters in the 2-layer LSTM model with Deep Gradient Compression described in the paper:\n\nModel has 2 LSTM layers with 1500 hidden units each\n\nFor an LSTM layer:\n\nW matrix: number of hidden units * input dimension\nU matrix: number of hidden units * number of hidden units\n4 bias vectors with size = number of hidden units\nSo params per LSTM layer = hidden_units * (input_dim + hidden_units + 4)\n\nFor layer 1:\n\nInput dim is word embedding size (estimated 300)\nParams = 1500 * (300 + 1500 + 4) = 1,512,000\nFor layer 2:\n\nInput is previous layer output (1500)\nParams = 1500 * (1500 + 1500 + 4) = 4,506,000\nTotal Params = Layer 1 + Layer 2\n= 1,512,000 + 4,506,000\n= 6,018,000\n\nThe Deep Gradient Compression technique does not change number of parameters.\n\nSo total parameters for the model is approximately 6 million."},{"id":"epoch:c503f3b8dc463b6a7be4:training-compute","entityId":"epoch:c503f3b8dc463b6a7be4","metric":"training-compute","value":1340000000000000.0,"unit":"FLOP","observationPeriod":"2017-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 6020000 parameters ['Likely' confidence] * 929000 tokens * 40 epochs = 1.3422192e+15 FLOP"},{"id":"epoch:b895e8a53add1db35ced:parameters","entityId":"epoch:b895e8a53add1db35ced","metric":"parameters","value":86000000.0,"unit":"parameters","observationPeriod":"2017-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0da34668eef513f181a1:parameters","entityId":"epoch:0da34668eef513f181a1","metric":"parameters","value":86100000.0,"unit":"parameters","observationPeriod":"2017-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 5"},{"id":"epoch:0da34668eef513f181a1:training-compute","entityId":"epoch:0da34668eef513f181a1","metric":"training-compute","value":6.62904e+19,"unit":"FLOP","observationPeriod":"2017-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8 times less compute than Zoph (2018), which used 500 p100s for 4 days.\n(From Imagenet paper-data, Besiroglu et al., forthcoming) "},{"id":"epoch:339bdf10b05ec9e3db73:parameters","entityId":"epoch:339bdf10b05ec9e3db73","metric":"parameters","value":193000000.0,"unit":"parameters","observationPeriod":"2017-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9426d930834ff136c87a:parameters","entityId":"epoch:9426d930834ff136c87a","metric":"parameters","value":177000000.0,"unit":"parameters","observationPeriod":"2017-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:e69cb506fb34f507e0e3:parameters","entityId":"epoch:e69cb506fb34f507e0e3","metric":"parameters","value":121000000.0,"unit":"parameters","observationPeriod":"2017-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"We test models with parameter counts ranging from 89K to 121M."},{"id":"epoch:d4faf5d38eb941a0913e:parameters","entityId":"epoch:d4faf5d38eb941a0913e","metric":"parameters","value":35000000.0,"unit":"parameters","observationPeriod":"2017-11-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"35M (Table 2)"},{"id":"epoch:d4faf5d38eb941a0913e:training-compute","entityId":"epoch:d4faf5d38eb941a0913e","metric":"training-compute","value":3.36e+18,"unit":"FLOP","observationPeriod":"2017-11-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 35000000 parameters * 2000000 tokens * 8000 epochs = 3.36e+18 FLOP \n\n_____________\nin the algorithmic progress paper the estimation was 4.37 \u00d7 10^17 FLOP based on 1000 epochs assumption"},{"id":"epoch:af85b1ee1dc91556e88d:parameters","entityId":"epoch:af85b1ee1dc91556e88d","metric":"parameters","value":22000000.0,"unit":"parameters","observationPeriod":"2017-11-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b2713fa96a1e5ec126d4:parameters","entityId":"epoch:b2713fa96a1e5ec126d4","metric":"parameters","value":34000000.0,"unit":"parameters","observationPeriod":"2017-10-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"34M (Table 2)"},{"id":"epoch:b2713fa96a1e5ec126d4:training-compute","entityId":"epoch:b2713fa96a1e5ec126d4","metric":"training-compute","value":3.06e+17,"unit":"FLOP","observationPeriod":"2017-10-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / token / parameter * 34000000 parameters * 2000000 tokens * 750 epochs = 3.06e+17 FLOP\n\n_________________\nIn the Algorithmic Progress paper, the compute was estimated to be 9.85 \u00d7 10\u00b9\u2076 FLOP, assuming 520 epochs reported for the PTB dataset."},{"id":"epoch:db4c738d3502e4a45a0f:parameters","entityId":"epoch:db4c738d3502e4a45a0f","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2017-10-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:db4c738d3502e4a45a0f:training-compute","entityId":"epoch:db4c738d3502e4a45a0f","metric":"training-compute","value":6.95e+16,"unit":"FLOP","observationPeriod":"2017-10-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*24000000*929000*520=6.956352e+16"},{"id":"epoch:9c8efb93ef4c8a297930:parameters","entityId":"epoch:9c8efb93ef4c8a297930","metric":"parameters","value":11360000.0,"unit":"parameters","observationPeriod":"2017-10-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"This model has 24.56M parameters which is 2 times more parameters\nthan CapsNet with 11.36M parameters.\""},{"id":"epoch:19abbcb1fd60955026d3:parameters","entityId":"epoch:19abbcb1fd60955026d3","metric":"parameters","value":8200000.0,"unit":"parameters","observationPeriod":"2017-10-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In terms of number of parameters the baseline has 35.4M while CapsNet\nhas 8.2M parameters and 6.8M parameters without the reconstruction subnetwork\""},{"id":"epoch:35bfb841cc76af6ebcfc:training-compute","entityId":"epoch:35bfb841cc76af6ebcfc","metric":"training-compute","value":3.4100000001e+20,"unit":"FLOP","observationPeriod":"2017-10-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"This is a guess. There was no single journal publication that accompanied this model, that gave information about architecture/model training time etc. All I could find was that it has the same architecture as AlphaGo Zero, and that it had roughly the same power consumption as AGZ. See for instance: \nhttps://deepmind.com/blog/article/alphago-zero-starting-scratch\n\nSince AGZ reaches the ELO of AlphaGo Master in about 25-30 days (60-75% of the total training time), I estimate the compute to be around 60-75% that of AGZ. I round this to 2e23, and I expect this to only be accurate within an OOM."},{"id":"epoch:47e1c681ec7a34e252e3:parameters","entityId":"epoch:47e1c681ec7a34e252e3","metric":"parameters","value":46400244.0,"unit":"parameters","observationPeriod":"2017-10-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Input size: 19*19*17=6137, internal dimension 19*19\n1 conv block\n17*3*3*256=39168\n39 residual blocks\n2*3*3*256*256=1179648\nPolicy head\n256*2+2*19*19*(19*19+2)=262598\nValue head\n256*1+1*19*19*256+256*1=92928\nTotal: 39168+39*1179648+262598+92928=46400966"},{"id":"epoch:47e1c681ec7a34e252e3:training-compute","entityId":"epoch:47e1c681ec7a34e252e3","metric":"training-compute","value":6.49439910290227e+20,"unit":"FLOP","observationPeriod":"2017-10-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Updating this compute estimate to only account for direct training compute, not synthetic data generation compute.\n\nNumber of connections (to calculate forward FLOP)\nInput size: 19*19*17, Assuming internal dimension stays at 19*19\n1 conv block\n19*19*17*3*3*256=14139648\n40 residual blocks\n19*19*2*3*3*256*256=425852928\nPolicy head\n19*19*256*2+2*19*19*(19*19+2)=446918\nValue head\n19*19*256*1+1*19*19*256+256*1=185088\nTotal: 14139648+40*425852928+446918+185088=17048888774\nForward FLOP: 2*17048888774=34097777548\nParameter updates \u201cParameters were updated from 3.1 million mini-batches of 2,048 positions each.\u201d\nTotal updates: 3100000*2048=6348800000\n\nTraining compute: 3 * 34097777548 FLOP * 6348800000 = 6.5e20 FLOP"},{"id":"epoch:bc7e35c8148b4e49ed00:parameters","entityId":"epoch:bc7e35c8148b4e49ed00","metric":"parameters","value":53000000.0,"unit":"parameters","observationPeriod":"2017-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"53M (Table 3)"},{"id":"epoch:bc7e35c8148b4e49ed00:training-compute","entityId":"epoch:bc7e35c8148b4e49ed00","metric":"training-compute","value":3180000000000000.0,"unit":"FLOP","observationPeriod":"2017-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 53000000 parameters * 2000000 tokens * 5 epochs = 3.18e+15 FLOP"},{"id":"epoch:8f6c45764693b0957eea:parameters","entityId":"epoch:8f6c45764693b0957eea","metric":"parameters","value":30000000.0,"unit":"parameters","observationPeriod":"2017-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:bcd5c56fc22d5dce8237:parameters","entityId":"epoch:bcd5c56fc22d5dce8237","metric":"parameters","value":50000000.0,"unit":"parameters","observationPeriod":"2017-09-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"table 2"},{"id":"epoch:f0bdf3563e49c7ccb27d:parameters","entityId":"epoch:f0bdf3563e49c7ccb27d","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2017-09-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:844908f191a18c38ad2b:parameters","entityId":"epoch:844908f191a18c38ad2b","metric":"parameters","value":33000000.0,"unit":"parameters","observationPeriod":"2017-09-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6b02522833e7c3536b1b:parameters","entityId":"epoch:6b02522833e7c3536b1b","metric":"parameters","value":11100000.000000002,"unit":"parameters","observationPeriod":"2017-09-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"11.1M (Table 2)"},{"id":"epoch:6b02522833e7c3536b1b:training-compute","entityId":"epoch:6b02522833e7c3536b1b","metric":"training-compute","value":3400000000000000.0,"unit":"FLOP","observationPeriod":"2017-09-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 11100000 parameters * 929000 tokens * 55 epochs = 3.402927e+15 FLOP"},{"id":"epoch:9c18aa84328dc0a59ca1:parameters","entityId":"epoch:9c18aa84328dc0a59ca1","metric":"parameters","value":26000000.0,"unit":"parameters","observationPeriod":"2017-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"best model had 26M params"},{"id":"epoch:9c18aa84328dc0a59ca1:training-compute","entityId":"epoch:9c18aa84328dc0a59ca1","metric":"training-compute","value":2340000000000000.0,"unit":"FLOP","observationPeriod":"2017-09-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND=6*26000000*50000*300=2.34e+15"},{"id":"epoch:373adc1c372b146b2bc3:parameters","entityId":"epoch:373adc1c372b146b2bc3","metric":"parameters","value":28100000.0,"unit":"parameters","observationPeriod":"2017-09-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 16"},{"id":"epoch:b83c5cdc8055c9e22d52:parameters","entityId":"epoch:b83c5cdc8055c9e22d52","metric":"parameters","value":38000000.0,"unit":"parameters","observationPeriod":"2017-08-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"38M (Table 2)"},{"id":"epoch:b83c5cdc8055c9e22d52:training-compute","entityId":"epoch:b83c5cdc8055c9e22d52","metric":"training-compute","value":4.56e+17,"unit":"FLOP","observationPeriod":"2017-08-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 38000000 parameters * 2000000 tokens * 1000 epochs = 4.56e+17 FLOP"},{"id":"epoch:f2e13b6d9c47c99d4708:parameters","entityId":"epoch:f2e13b6d9c47c99d4708","metric":"parameters","value":26000000.0,"unit":"parameters","observationPeriod":"2017-08-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f2e13b6d9c47c99d4708:training-compute","entityId":"epoch:f2e13b6d9c47c99d4708","metric":"training-compute","value":1.45e+17,"unit":"FLOP","observationPeriod":"2017-08-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*26000000*929000*1000=1.44924e+17"},{"id":"epoch:74f6b313fb2e3bb1d1b7:parameters","entityId":"epoch:74f6b313fb2e3bb1d1b7","metric":"parameters","value":5180000.0,"unit":"parameters","observationPeriod":"2017-08-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:fd9010cb80a91584dbb4:parameters","entityId":"epoch:fd9010cb80a91584dbb4","metric":"parameters","value":7160000.0,"unit":"parameters","observationPeriod":"2017-08-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9295e3c2640e6d2225d8:parameters","entityId":"epoch:9295e3c2640e6d2225d8","metric":"parameters","value":5970000.0,"unit":"parameters","observationPeriod":"2017-08-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:b3909e3db1f9bae91c92:training-compute","entityId":"epoch:b3909e3db1f9bae91c92","metric":"training-compute","value":5.51e+20,"unit":"FLOP","observationPeriod":"2017-08-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In total, Libratus used about 25 million core hours. Of those, about 13 million core hours were used for exploratory experiments and evaluation. About 6 million core hours were spent on the initial abstraction and equilibrium finding component, another 3 million were used for nested subgame solving, and about 3 million were used on the self-improvement algorithm.\"\n\n\"Like many data-centric supercomputers, Bridges offers a relatively a modest number of FLOPS, but lots of memory: 895 teraflops and 130 TB, respectively.\"\n\nI just used the first bullet point (as those are usually independent systems and you only benchmark one of them).\nThe first system has 752 nodes a 2CPUs a 14cores each.\n\nsource: https://www.top500.org/news/bridges-supercomputer-boots-up-at-pittsburgh/\n\n\n\n1. 12M core hours for 196 cores\n2. We have  895 TFLOPS for 752 nodes a 2 CPUs a 14 cores\n2.1 That's 42.5 GFLOPS per core.\n3. Running this for 12M h\n3.1 12 * 10^6 * 60 * 60 * 42.5 * 10^9 FLOP/S = 1.823e21 FLOPs\n4. Assuming 30% utilization\n 1.823e21 * 0.3\n\u2192 5.51e20 FLOPs"},{"id":"epoch:90164fc5b21192f4bea6:parameters","entityId":"epoch:90164fc5b21192f4bea6","metric":"parameters","value":206000000.0,"unit":"parameters","observationPeriod":"2017-08-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:fb7f5b90c86d8ce42b91:parameters","entityId":"epoch:fb7f5b90c86d8ce42b91","metric":"parameters","value":60000000.0,"unit":"parameters","observationPeriod":"2017-08-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Model is based on ResNet (60m params), might have more parameters though\n\n\"We implement all deep methods based on the Caffe framework, and fine-tune from Caffe-provided models of AlexNet (Krizhevsky et al., 2012) and ResNet (He et al., 2016), both are pre-trained on the ImageNet 2012 dataset.\""},{"id":"epoch:3a3ea4e8741aa8bfef3d:parameters","entityId":"epoch:3a3ea4e8741aa8bfef3d","metric":"parameters","value":19000000.0,"unit":"parameters","observationPeriod":"2017-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"19M (Table 4)"},{"id":"epoch:3a3ea4e8741aa8bfef3d:training-compute","entityId":"epoch:3a3ea4e8741aa8bfef3d","metric":"training-compute","value":1.06e+16,"unit":"FLOP","observationPeriod":"2017-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 19000000 parameters * 929000 tokens * 100 epochs = 1.05906e+16 FLOP"},{"id":"epoch:4bd2b83d5e215a93decd:parameters","entityId":"epoch:4bd2b83d5e215a93decd","metric":"parameters","value":25000000.0,"unit":"parameters","observationPeriod":"2017-08-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:8d063f67d8d19d6c2dbb:training-compute","entityId":"epoch:8d063f67d8d19d6c2dbb","metric":"training-compute","value":6.046095222592002e+20,"unit":"FLOP","observationPeriod":"2017-08-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Extracted from AI and Compute (https://openai.com/blog/ai-and-compute/) charts by using https://automeris.io/WebPlotDigitizer/."},{"id":"epoch:392f7578856f43bcd337:parameters","entityId":"epoch:392f7578856f43bcd337","metric":"parameters","value":34000000.0,"unit":"parameters","observationPeriod":"2017-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: table 2 in https://arxiv.org/pdf/1911.09070.pdf"},{"id":"epoch:d8edb9dd0c0b20b7d160:parameters","entityId":"epoch:d8edb9dd0c0b20b7d160","metric":"parameters","value":53000000.0,"unit":"parameters","observationPeriod":"2017-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"source: table 2 in https://arxiv.org/pdf/1911.09070.pdf"},{"id":"epoch:d8edb9dd0c0b20b7d160:training-compute","entityId":"epoch:d8edb9dd0c0b20b7d160","metric":"training-compute","value":2.065392e+18,"unit":"FLOP","observationPeriod":"2017-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use synchronized SGD over 8 GPUs with a total of 16 images per minibatch (2 images per GPU). Unless otherwise specified, all models are trained for 90k iterations with an initial learning rate of 0.01, which is then divided by 10 at 60k and again at 80k iterations. We use horizontal image flipping as the only form of data augmentation unless otherwise noted. Weight decay of 0.0001 and momentum of 0.9 are used. The training loss is the sum the focal loss and the standard smooth L1 loss used for box regression [10]. Training time ranges between 10 and 35 hours for the models in Table 1e.\"\n\nNVIDIA M40 GPU\n\n35*60**2*0.3*8*6.83E+12 = 2.07e18"},{"id":"epoch:47d9deaf4c5f898790ed:parameters","entityId":"epoch:47d9deaf4c5f898790ed","metric":"parameters","value":33000000.0,"unit":"parameters","observationPeriod":"2017-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"33M (Table 2)"},{"id":"epoch:47d9deaf4c5f898790ed:training-compute","entityId":"epoch:47d9deaf4c5f898790ed","metric":"training-compute","value":2.97e+17,"unit":"FLOP","observationPeriod":"2017-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 33000000 parameters * 2000000 tokens * 750 epochs = 2.97e+17 FLOP"},{"id":"epoch:20bc909460de1206fe91:parameters","entityId":"epoch:20bc909460de1206fe91","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2017-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:20bc909460de1206fe91:training-compute","entityId":"epoch:20bc909460de1206fe91","metric":"training-compute","value":1e+17,"unit":"FLOP","observationPeriod":"2017-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*24000000*929000*750=1.00332e+17"},{"id":"epoch:d8acad09d009409008c1:training-compute","entityId":"epoch:d8acad09d009409008c1","metric":"training-compute","value":5.64e+19,"unit":"FLOP","observationPeriod":"2017-07-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"All models are implemented in Torch (Collobert et al., 2011) and trained on a single Nvidia M40 GPU except for WMT\u201914 English-French for which we use a multi-GPU setup on a single\nmachine. We train on up to eight GPUs synchronously by\nmaintaining copies of the model on each card and split the\nbatch so that each worker computes 1/8-th of the gradients;\nat the end we sum the gradients via Nvidia NCCL.\n\n1. English-Romanian: \"Training took between 6 and 7.5 days on a single GPU.\"\n7 days * 24 * 3600 * 6.8e12 FLOP/s (Nvidia M40, fp32) * 0.3 = 1.2e18 FLOP\n\n2. English-German: \" We trained this model on a single GPU over a\nperiod of 18.5 days with a batch size of 48\".\n18.5 days * 24 * 3600 * 6.8e12 FLOP/s (Nvidia M40, fp32) * 0.3 = 3.3e18 FLOP\n\n3. English-French: \"Our results are based on training\nwith 8 GPUs for about 37 days and batch size 32 on each\nworker.6 \"\n37 days * 24 * 3600 * 8 * 6.8e12 FLOP/s (Nvidia M40, fp32) * 0.3 = 5.2e19 FLOP\n\nthe minimum compute needed to train ensemble model: 1.2e18 FLOP + 3.3e18 FLOP + 5.2e19 FLOP = 5.65e19 FLOP\n\nI am not sure how much to add more (they say ensemble model consists of 8 models), probably summarization training takes at least 1.2e18 FLOP more. \n\n"},{"id":"epoch:34baab493e6e67cd3606:parameters","entityId":"epoch:34baab493e6e67cd3606","metric":"parameters","value":89000000.0,"unit":"parameters","observationPeriod":"2017-07-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:42bf316360331710d96a:parameters","entityId":"epoch:42bf316360331710d96a","metric":"parameters","value":23000000.0,"unit":"parameters","observationPeriod":"2017-07-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"23M (Table 1)"},{"id":"epoch:42bf316360331710d96a:training-compute","entityId":"epoch:42bf316360331710d96a","metric":"training-compute","value":1.28e+16,"unit":"FLOP","observationPeriod":"2017-07-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 23000000 parameters * 929000 tokens * 100 epochs = 1.28202e+16 FLOP"},{"id":"epoch:9af136e9bea896d548d6:parameters","entityId":"epoch:9af136e9bea896d548d6","metric":"parameters","value":14000000.0,"unit":"parameters","observationPeriod":"2017-07-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9af136e9bea896d548d6:training-compute","entityId":"epoch:9af136e9bea896d548d6","metric":"training-compute","value":7800000000000000.0,"unit":"FLOP","observationPeriod":"2017-07-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*14000000*929000*100=7.8036e+15"},{"id":"epoch:0232c213a873fd8c7948:parameters","entityId":"epoch:0232c213a873fd8c7948","metric":"parameters","value":24000000.0,"unit":"parameters","observationPeriod":"2017-07-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:44a688640f704b1ac9bb:parameters","entityId":"epoch:44a688640f704b1ac9bb","metric":"parameters","value":44654504.0,"unit":"parameters","observationPeriod":"2017-07-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Uses ResNet-101 architecture, which has 44,654,504 parameters:\nhttps://resources.wolframcloud.com/NeuralNetRepository/resources/ResNet-101-Trained-on-ImageNet-Competition-Data/"},{"id":"epoch:44a688640f704b1ac9bb:training-compute","entityId":"epoch:44a688640f704b1ac9bb","metric":"training-compute","value":8.43e+20,"unit":"FLOP","observationPeriod":"2017-07-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Tesla K80 performance: 8.13 TFLOP/s\n\nAssume 40% utilization\n\n60 days * 50 GPUs * 40% utilization * 8.13 TFLOP/s/GPU = 8.43*10^20 FLOP"},{"id":"epoch:746a1449d0644ea8406e:training-compute","entityId":"epoch:746a1449d0644ea8406e","metric":"training-compute","value":5.781024e+17,"unit":"FLOP","observationPeriod":"2017-07-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6691000000000 FLOP / GPU / sec [Titan X reported] * 1 GPU [assumed] * 80 hours * 3600 sec / hour * 0.3 [assumed utilization] = 5.781024e+17 FLOP"},{"id":"epoch:9086dee73ed0102389c8:parameters","entityId":"epoch:9086dee73ed0102389c8","metric":"parameters","value":2430000.0,"unit":"parameters","observationPeriod":"2017-07-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f0d2916f8d848e5616b3:parameters","entityId":"epoch:f0d2916f8d848e5616b3","metric":"parameters","value":213000000.0,"unit":"parameters","observationPeriod":"2017-06-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"This page suggests the transformer has 213M parameters.\n\n\"Although there are others architectures that make use of attention layers, none achieves so good results so fast. Not only that, but the only model that can compite against Transformer is the Slicenet22, proposed just fifteen days before. It takes much longer to train, due to the huge amount of parameters it requires (348 million against the 213 millions of Transformer), and the BLEU scores it achieves are slightly worse on average. In short, up to date it offers no profit over Transformer.\"\n\nhttps://ricardokleinklein.github.io/2017/11/16/Attention-is-all-you-need.html"},{"id":"epoch:f0d2916f8d848e5616b3:training-compute","entityId":"epoch:f0d2916f8d848e5616b3","metric":"training-compute","value":7.4245248e+18,"unit":"FLOP","observationPeriod":"2017-06-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The model was trained during 300000 steps, roughly 3.5 days, using 8 NVIDIA P100 GPUs.\"\n\nsource: https://ricardokleinklein.github.io/2017/11/16/Attention-is-all-you-need.html\n\nNVIDIA Tesla P100 has 9.3 teraFLOPS single-precision performance\n\nsource: https://www.nvidia.com/en-gb/data-center/tesla-p100/\n\nWe assume 0.33 utilization performance, in line with OpenAI's \"AI and compute\" article\n\nsource: https://openai.com/blog/ai-and-compute/\n\n9.3*10^12 FLOP / GPU / sec * 8 GPUs * 3.5 days * 24 hour / day * 3600 sec / hour * 0.33 [assumed utilization] = 7.4245248e+18 FLOP\n\nIn the \"AI and Memory Wall\" paper (https://github.com/amirgholami/ai_and_memory_wall) the estimation is 23,000 PFLOPS = 2.3*10^19 FLOPs. They estimate number of parameters as 65M\n\nNote that Table 2 provides a training compute estimate, but appears not to account for utilization."},{"id":"epoch:946d6da190bcd0389898:parameters","entityId":"epoch:946d6da190bcd0389898","metric":"parameters","value":4200000.0,"unit":"parameters","observationPeriod":"2017-04-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:437348494e814d3003e7:parameters","entityId":"epoch:437348494e814d3003e7","metric":"parameters","value":60000000.0,"unit":"parameters","observationPeriod":"2017-03-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The paper reimplements ResNet-20, AlexNet and GoogLeNet. The largest from these models is AlexNet. (data for these models taken from this db)\nAlexNet has 60000000.00 params.\nGoogLeNet has 6797700.00 params.\nSo max(60000000, 6797700) = 60000000\nsource \"The effectiveness of our approach is comprehensively evaluated in ResNets, AlexNet, and GoogLeNet. In AlexNet, for example, Force Regularization gains 2x speedup on modern GPU without accuracy loss and 4.05x speedup on CPU by paying small accuracy degradation. \""},{"id":"epoch:437348494e814d3003e7:training-compute","entityId":"epoch:437348494e814d3003e7","metric":"training-compute","value":4.7e+17,"unit":"FLOP","observationPeriod":"2017-03-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The paper reimplements ResNet-20, AlexNet and GoogLeNet. The largest from these models is AlexNet. (data for these models taken from this db)\ntraining of AlexNet taken 470000000000000000.00 FLOPs\ntraining of GoogLeNet taken 1557140125176000000.00 FLOPs\nmax(1557140125176000000, 470000000000000000) = 1557140125176000000\nsource \"The effectiveness of our approach is comprehensively evaluated in ResNets, AlexNet, and GoogLeNet. In AlexNet, for example, Force Regularization gains 2x speedup on modern GPU without accuracy loss and 4.05x speedup on CPU by paying small accuracy degradation. \""},{"id":"epoch:3a1ad36f681e9e014d97:parameters","entityId":"epoch:3a1ad36f681e9e014d97","metric":"parameters","value":5320000.0,"unit":"parameters","observationPeriod":"2017-03-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:37a644232191c83ef3c9:parameters","entityId":"epoch:37a644232191c83ef3c9","metric":"parameters","value":7800000.0,"unit":"parameters","observationPeriod":"2017-01-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7.8 M from Table 2."},{"id":"epoch:37a644232191c83ef3c9:training-compute","entityId":"epoch:37a644232191c83ef3c9","metric":"training-compute","value":5.7267e+17,"unit":"FLOP","observationPeriod":"2017-01-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"5050000000000 [peak FLOPs] * 15 [epochs] * 7 [h] * 3600 [s] * 0.3 [assumed utilization rate] = 5.7267e+17"},{"id":"epoch:46e4d0f5730c489022f9:parameters","entityId":"epoch:46e4d0f5730c489022f9","metric":"parameters","value":8700000000.0,"unit":"parameters","observationPeriod":"2017-01-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 5\n\nhttps://arxiv.org/abs/1701.06538"},{"id":"epoch:46e4d0f5730c489022f9:training-compute","entityId":"epoch:46e4d0f5730c489022f9","metric":"training-compute","value":9.393905664e+19,"unit":"FLOP","observationPeriod":"2017-01-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12 days \n64 NVIDIA K40 GPUs (see hardware data sheet for performance)\n0.33 util rate\n "},{"id":"epoch:65d05b12c5f7dd08c844:parameters","entityId":"epoch:65d05b12c5f7dd08c844","metric":"parameters","value":18200000.0,"unit":"parameters","observationPeriod":"2017-01-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"18.2M for largest OR-WideResNet model."},{"id":"epoch:acb29c57f39d9a736717:parameters","entityId":"epoch:acb29c57f39d9a736717","metric":"parameters","value":2500000.0,"unit":"parameters","observationPeriod":"2017-01-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 3, p.9\n\nsource: https://docs.google.com/spreadsheets/d/1Kj4Q5WADcDXtUJLIOfGTCE3tGvxNczEMwyy8QtgSkHk/edit#gid=54587040&fvid=1361937389"},{"id":"epoch:acb29c57f39d9a736717:training-compute","entityId":"epoch:acb29c57f39d9a736717","metric":"training-compute","value":1.446336e+19,"unit":"FLOP","observationPeriod":"2017-01-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The largest source of compute necessary for training seems to be the data generation job on 20 GPUs. We count this towards the training compute because it requires simulation using the network. This is analogous to the AlphaGo systems simulating Go games.\n\nFrom p.26: \"For the flop network, one million poker flop situations (from after the flop cards are dealt) were generated and solved. These situations were solved using DeepStack\u2019s depth limited solver with the turn network used for the counterfactual values at public states immediately after the turn card. We used a cluster of 20 GPUS and one-half of a GPU year of computation time.\"\n\nAssume they used P100 GPUs because they were common at the time (P100 was released in 2016 and this paper was published in 2017).\n\nBut assume low utilization of 10% to hedge on (a) lower-performing GPUs being used, (b) non-FLOP computations taking up a lot of the data generation job.\n\nCalculation:\n6 months * 30 days * 24 hours * 3600 seconds * 9.3e12 FLOP/s * 0.1 utilization = 1.446336e+19 FLOP."},{"id":"epoch:f3a354e9878eaf2d0ae0:parameters","entityId":"epoch:f3a354e9878eaf2d0ae0","metric":"parameters","value":51000000.0,"unit":"parameters","observationPeriod":"2016-12-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Source: https://resources.wolframcloud.com/NeuralNetRepository/resources/YOLO-V2-Trained-on-MS-COCO-Data_1"},{"id":"epoch:b44ea3936c2080204792:parameters","entityId":"epoch:b44ea3936c2080204792","metric":"parameters","value":814464.0,"unit":"parameters","observationPeriod":"2016-12-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2*3*3*3*64+22*3*3*64*64=814464\n24 CNN layers with 3x3 kernels and 64 channels and 3 input/output channels (see Table 1)\n"},{"id":"epoch:b44ea3936c2080204792:training-compute","entityId":"epoch:b44ea3936c2080204792","metric":"training-compute","value":1.3079231999999998e+17,"unit":"FLOP","observationPeriod":"2016-12-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Compute: 0.3*24*60*60*5046000000000=130792319999999980\nK40 FLOPs: 5046000000000\n\"We trained all models for a maximum of 24 hours on an Nvidia K40 GPU\""},{"id":"epoch:22b6432c821e0bf056fa:parameters","entityId":"epoch:22b6432c821e0bf056fa","metric":"parameters","value":42000000.0,"unit":"parameters","observationPeriod":"2016-12-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"42M (Table 2)"},{"id":"epoch:22b6432c821e0bf056fa:training-compute","entityId":"epoch:22b6432c821e0bf056fa","metric":"training-compute","value":3040000000000000.0,"unit":"FLOP","observationPeriod":"2016-12-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 42000000 parameters * 929000 tokens * 13 epochs = 3.043404e+15 FLOP"},{"id":"epoch:e4a0b2ce9afd87dd3335:parameters","entityId":"epoch:e4a0b2ce9afd87dd3335","metric":"parameters","value":44500000.0,"unit":"parameters","observationPeriod":"2016-12-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Based on ResNet 101 architecture"},{"id":"epoch:4ed7f5956d9a59f855c8:parameters","entityId":"epoch:4ed7f5956d9a59f855c8","metric":"parameters","value":300000000.0,"unit":"parameters","observationPeriod":"2016-12-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d416ecae3911e8257517:parameters","entityId":"epoch:d416ecae3911e8257517","metric":"parameters","value":44500000.0,"unit":"parameters","observationPeriod":"2016-12-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"ResNet 101 "},{"id":"epoch:d416ecae3911e8257517:training-compute","entityId":"epoch:d416ecae3911e8257517","metric":"training-compute","value":7.077e+18,"unit":"FLOP","observationPeriod":"2016-12-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"ResNet-101 training compute: 7004000000000000000\nFinetune compute: 73422840000000000\nTotal compute: 7004000000000000000+73422840000000000=7.0774228e+18"},{"id":"epoch:ec66b6d6932b0d65dc93:parameters","entityId":"epoch:ec66b6d6932b0d65dc93","metric":"parameters","value":92000000.0,"unit":"parameters","observationPeriod":"2016-11-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:ec66b6d6932b0d65dc93:training-compute","entityId":"epoch:ec66b6d6932b0d65dc93","metric":"training-compute","value":6.4e+19,"unit":"FLOP","observationPeriod":"2016-11-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 5: \"ResNet-500 [has] similar computation\ncosts to our Very Deep PolyNet\".\n\nResNet-152 has 11.3e9 FLOP per forward pass (https://arxiv.org/abs/1512.03385, Table 1). Hence ResNet-500 has approx 3.7e10 = 11.3e9*500/152 FLOP per forward pass.\n\n560k iterations, batch size 512:\nTrain compute = 3.7e10*3*2*560e3 * 512 = 6.4e19"},{"id":"epoch:0027267657c184a1e38f:parameters","entityId":"epoch:0027267657c184a1e38f","metric":"parameters","value":25000000.0,"unit":"parameters","observationPeriod":"2016-11-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"If you\u2019re thinking about ResNets, yes, they are related. ResNeXt-50 has 25M parameters (ResNet-50 has 25.5M).\"\n\nhttps://towardsdatascience.com/illustrated-10-cnn-architectures-95d78ace614d"},{"id":"epoch:48a43c728ed5d6415c8a:parameters","entityId":"epoch:48a43c728ed5d6415c8a","metric":"parameters","value":83000000.0,"unit":"parameters","observationPeriod":"2016-11-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:48a43c728ed5d6415c8a:training-compute","entityId":"epoch:48a43c728ed5d6415c8a","metric":"training-compute","value":1.2e+19,"unit":"FLOP","observationPeriod":"2016-11-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12,000 PFLOPs = 1.2 * 10^19 FLOPs\n\nhttps://github.com/amirgholami/ai_and_memory_wall"},{"id":"epoch:7e6c52250a1506e339b9:parameters","entityId":"epoch:7e6c52250a1506e339b9","metric":"parameters","value":564000000.0,"unit":"parameters","observationPeriod":"2016-11-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Finetunes 4 versions of VGG\nVGG 16: 138000000\nVGG 19: 144000000\nTotal 2*(138000000+144000000)=564000000\n"},{"id":"epoch:e72ca12951e3d4fcc45d:parameters","entityId":"epoch:e72ca12951e3d4fcc45d","metric":"parameters","value":37400000.0,"unit":"parameters","observationPeriod":"2016-11-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:e72ca12951e3d4fcc45d:training-compute","entityId":"epoch:e72ca12951e3d4fcc45d","metric":"training-compute","value":2.2e+21,"unit":"FLOP","observationPeriod":"2016-11-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"50 epochs * 50,000 images * 10.0 GFLOPSs * 12800 networks * 2 add-multiply * 3 backward pass \n= 1.9e6 PF = 22 pfs-days\n\nsource: https://openai.com/blog/ai-and-compute/"},{"id":"epoch:b063f7d41e78c7637064:parameters","entityId":"epoch:b063f7d41e78c7637064","metric":"parameters","value":54000000.0,"unit":"parameters","observationPeriod":"2016-11-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"54M (Table 2)"},{"id":"epoch:b063f7d41e78c7637064:training-compute","entityId":"epoch:b063f7d41e78c7637064","metric":"training-compute","value":1.05e+16,"unit":"FLOP","observationPeriod":"2016-11-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 54000000 parameters * 929000 tokens * 35 epochs = 1.053486e+16 FLOP"},{"id":"epoch:2885bcb0ac7f676ad7cf:parameters","entityId":"epoch:2885bcb0ac7f676ad7cf","metric":"parameters","value":2600000.0,"unit":"parameters","observationPeriod":"2016-11-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"There are two similar models described in sections \"Models details\"\ncitation from the paper about model for SQuAD \"The model has about 2.6 million parameters\"\ncitation about model for cloze test\n\"The model architecture used for this task is very similar to that for SQuAD (Section 4) with only a few small changes to adapt it to the cloze test. \"\n"},{"id":"epoch:2885bcb0ac7f676ad7cf:training-compute","entityId":"epoch:2885bcb0ac7f676ad7cf","metric":"training-compute","value":3.4686144e+18,"unit":"FLOP","observationPeriod":"2016-11-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"flops = (8) * (6691 * 10**9) * (60 * 3600) * 3 // 10\n(num gpu) * (peak flops) * (time in seconds) * (assumed utilization rate) =\n\ncitation from the section about cloze test experiments \"The entire training process takes roughly 60 hours on eight Titan X GPUs. The other hyper-parameters are identical to the model described in Section 4\" (section 4 is about SQuAD experiments and cloze test experiments require more compute and data).\nflops  6.691 TFLOPS from https://www.techpowerup.com/gpu-specs/geforce-gtx-titan-x.c2632"},{"id":"epoch:72f01c9a2ad8b494d266:parameters","entityId":"epoch:72f01c9a2ad8b494d266","metric":"parameters","value":51000000.0,"unit":"parameters","observationPeriod":"2016-11-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"51M (Table 3)"},{"id":"epoch:72f01c9a2ad8b494d266:training-compute","entityId":"epoch:72f01c9a2ad8b494d266","metric":"training-compute","value":2.13e+16,"unit":"FLOP","observationPeriod":"2016-11-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 51000000 parameters * 929000 tokens * 75 epochs = 2.132055e+16 FLOP"},{"id":"epoch:3aab67cbc272dcd5f369:parameters","entityId":"epoch:3aab67cbc272dcd5f369","metric":"parameters","value":22100000.0,"unit":"parameters","observationPeriod":"2016-11-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"VD-LSTM+REAL Large has 51M parameters. The parameter count of the Medium model is not reported, but they say it has 650 hidden units per layer, compared to 1500 for the Large model. Neglecting the rest of the architecture,\n51M * (650/1500) = 22.1M"},{"id":"epoch:8aea9f347bc0376bf8c2:parameters","entityId":"epoch:8aea9f347bc0376bf8c2","metric":"parameters","value":6800000.0,"unit":"parameters","observationPeriod":"2016-11-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"VD-LSTM+REAL Large has 51M parameters. The parameter count of the Small model is not reported, but they say it has 200 hidden units per layer, compared to 1500 for the Large model. Neglecting the rest of the architecture,\n51M * (200/1500) = 6.8M"},{"id":"epoch:777597845fbaa34a959b:parameters","entityId":"epoch:777597845fbaa34a959b","metric":"parameters","value":409536.0,"unit":"parameters","observationPeriod":"2016-10-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Three networks, each three layers. First takes in 459 inputs and outputs 12, second and third take in 459 + (12*17) = 663 inputs.\n\nNetwork 1: (459 * 150 + 150) + (150 * 150 + 150) + (150 * 150 + 150) + (150 * 12 + 12) = 116,112\nNetworks 2 and 3: (663 * 150 + 150) + (150 * 150 + 150) + (150 * 150 + 150) + (150 * 12 + 12) = 146,712\nTotal: 116,112 + (2 * 146,712) = 409,536"},{"id":"epoch:777597845fbaa34a959b:training-compute","entityId":"epoch:777597845fbaa34a959b","metric":"training-compute","value":1.822e+16,"unit":"FLOP","observationPeriod":"2016-10-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"120 epochs, dataset 5789 proteins. There are about 300 residues per protein (115,479 residues / 418 proteins) according to https://www.ncbi.nlm.nih.gov/pmc/articles/PMC22960/. \nFirst network gets 27 features per residue, second and third get 39.\nFLOPs from first: 6 * 116112 * (27 * 300 * 5789 * 120) = 3.92e15\nFLOPs from 2nd and 3rd: 2 *6 * 146712 * (39 * 300 * 5789 * 120) = 1.43e16\nTotal: 1.822E16"},{"id":"epoch:717f76fb5c57e1f108cd:parameters","entityId":"epoch:717f76fb5c57e1f108cd","metric":"parameters","value":22855952.0,"unit":"parameters","observationPeriod":"2016-10-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 3"},{"id":"epoch:717f76fb5c57e1f108cd:training-compute","entityId":"epoch:717f76fb5c57e1f108cd","metric":"training-compute","value":4.36e+20,"unit":"FLOP","observationPeriod":"2016-10-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"60 K80 GPUs * 30 days * 8.5 TFLOPS/GPU * 0.33 utilization  = 4.36e20\n\nAuthors of \"AI and Memory Wall\" (https://github.com/amirgholami/ai_and_memory_wall) estimated model's training compute as 450,000 PFLOP = 4.5*10^20 FLOP"},{"id":"epoch:254bc33523ba687ee60a:parameters","entityId":"epoch:254bc33523ba687ee60a","metric":"parameters","value":278000000.0,"unit":"parameters","observationPeriod":"2016-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 5 in 'Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer'\n\nhttps://arxiv.org/abs/1701.06538"},{"id":"epoch:254bc33523ba687ee60a:training-compute","entityId":"epoch:254bc33523ba687ee60a","metric":"training-compute","value":6.620000000001e+21,"unit":"FLOP","observationPeriod":"2016-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From AI and Compute:\n\"sqrt(10 * 100) factor added because production model used 2-3 orders of magnitude more data, but only 1 epoch rather than 10.\n96 K80 GPU\u2019s * 9 days * 8.5 TFLOPS * 0.33 utilization * sqrt(10 * 100)  \n= 6.9e6 PF = 79 pfs-days\"\nsource: https://openai.com/blog/ai-and-compute/\n\nhttps://www.wolframalpha.com/input?i=96+*+9+days+*+8.5+TFLOPS+*+0.33+*+sqrt%281000%29\n"},{"id":"epoch:d648aaf33dbac7ee8f35:parameters","entityId":"epoch:d648aaf33dbac7ee8f35","metric":"parameters","value":21000000.0,"unit":"parameters","observationPeriod":"2016-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d648aaf33dbac7ee8f35:training-compute","entityId":"epoch:d648aaf33dbac7ee8f35","metric":"training-compute","value":7490000000000000.0,"unit":"FLOP","observationPeriod":"2016-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 21000000 parameters * 929000 tokens * 64 epochs = 7.491456e+15 FLOP"},{"id":"epoch:e61a6b1465f3f8574a52:parameters","entityId":"epoch:e61a6b1465f3f8574a52","metric":"parameters","value":21000000.0,"unit":"parameters","observationPeriod":"2016-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"21M (Table 3)"},{"id":"epoch:e61a6b1465f3f8574a52:training-compute","entityId":"epoch:e61a6b1465f3f8574a52","metric":"training-compute","value":1.6128e+16,"unit":"FLOP","observationPeriod":"2016-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 21000000 parameters * 2000000 tokens * 64 epochs = 1.6128e+16 FLOP"},{"id":"epoch:4b82b8a83a43f1078f37:parameters","entityId":"epoch:4b82b8a83a43f1078f37","metric":"parameters","value":46000000.0,"unit":"parameters","observationPeriod":"2016-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:7cb22330b2bdc754fece:parameters","entityId":"epoch:7cb22330b2bdc754fece","metric":"parameters","value":20000000.0,"unit":"parameters","observationPeriod":"2016-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:7cb22330b2bdc754fece:training-compute","entityId":"epoch:7cb22330b2bdc754fece","metric":"training-compute","value":7100000000000000.0,"unit":"FLOP","observationPeriod":"2016-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*20000000*929000*64=7.13472e+15"},{"id":"epoch:f6e26c5713cef2482432:parameters","entityId":"epoch:f6e26c5713cef2482432","metric":"parameters","value":21000000.0,"unit":"parameters","observationPeriod":"2016-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:f6e26c5713cef2482432:training-compute","entityId":"epoch:f6e26c5713cef2482432","metric":"training-compute","value":1600000000000000.0,"unit":"FLOP","observationPeriod":"2016-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*21000000*2000000*64=1.6128e+16"},{"id":"epoch:3887e202f91aae9874ca:parameters","entityId":"epoch:3887e202f91aae9874ca","metric":"parameters","value":84000000.0,"unit":"parameters","observationPeriod":"2016-09-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"84M from Table 1."},{"id":"epoch:3887e202f91aae9874ca:training-compute","entityId":"epoch:3887e202f91aae9874ca","metric":"training-compute","value":1.008e+17,"unit":"FLOP","observationPeriod":"2016-09-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6ND = 6 FLOP/param/token * 84000000 parameters * 200000000 tokens = 1.008e+17 FLOP (Speculative confidence since the amount of epochs is unknown)"},{"id":"epoch:4385e7959a7986903b15:parameters","entityId":"epoch:4385e7959a7986903b15","metric":"parameters","value":10200000.0,"unit":"parameters","observationPeriod":"2016-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:9705e03d9ba211e3783b:training-compute","entityId":"epoch:9705e03d9ba211e3783b","metric":"training-compute","value":2.9741645e+19,"unit":"FLOP","observationPeriod":"2016-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"ResNet-200 takes about 3 weeks to train on 8 GPUs\". didn't specify which GPU\nupd: \ncommon GPU performance for 2016 is 6.83E+12 FLOPs/s (https://epoch.ai/blog/estimating-training-compute#forward-pass-compute-and-parameter-counts-of-common-layers) \nthen 6.83E+12*3*7*24*3600*8*0.3=2.9741645e+19 (Speculative)"},{"id":"epoch:318038fd47e8b0c5da90:parameters","entityId":"epoch:318038fd47e8b0c5da90","metric":"parameters","value":3172117056.0,"unit":"parameters","observationPeriod":"2016-09-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Large ensemble\nCNNs: 85000000 + 38000000 + 65000000 = 188000000 = 188M\nLSTMs: 40 / 140 input - 6*512 layers - 9000 / 27000 output\n4*(140+512)*512 + 5*4*(512+512)*512 + 4*(512+27000)*27000=2983117056\nLM: 1000*1000 = 1000000 + Embedding\nTotal: 1000000+2983117056+188000000=3172117056 (underestimate because it doesn't account for embeddings)\n"},{"id":"epoch:63646df6e5a6f548429b:parameters","entityId":"epoch:63646df6e5a6f548429b","metric":"parameters","value":16600000.0,"unit":"parameters","observationPeriod":"2016-09-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Largest model: TDNN-A: 16.6 million parameters (Table 2)"},{"id":"epoch:b52dc70b70cc0b25e79a:parameters","entityId":"epoch:b52dc70b70cc0b25e79a","metric":"parameters","value":34000000.0,"unit":"parameters","observationPeriod":"2016-08-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 3"},{"id":"epoch:009ea8ec0d421fc3dce7:parameters","entityId":"epoch:009ea8ec0d421fc3dce7","metric":"parameters","value":5480000.0,"unit":"parameters","observationPeriod":"2016-08-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"SOTA CIFAR-10 model was 5.48m params"},{"id":"epoch:198e3c31bd9fde79b17e:parameters","entityId":"epoch:198e3c31bd9fde79b17e","metric":"parameters","value":82130304.0,"unit":"parameters","observationPeriod":"2016-08-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"from appendix E: The convolutional neural network uses a 64\u00d7(5\u00d75)-64\u00d7(5\u00d75)-64\u00d7(5\u00d75)-512 architecture."},{"id":"epoch:198e3c31bd9fde79b17e:training-compute","entityId":"epoch:198e3c31bd9fde79b17e","metric":"training-compute","value":6.448896e+16,"unit":"FLOP","observationPeriod":"2016-08-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(peak FLOPs for GPU - 1244 GFLOPs) times (training time - 3600 * 48 second) * (0.3 assumed utilization rate) "},{"id":"epoch:da3a90e81fdb94f70d52:parameters","entityId":"epoch:da3a90e81fdb94f70d52","metric":"parameters","value":3700000.0,"unit":"parameters","observationPeriod":"2016-07-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" The total number of\nweights was increased to approximately 3.7M. "},{"id":"epoch:03f28a809cb64ef41fe6:parameters","entityId":"epoch:03f28a809cb64ef41fe6","metric":"parameters","value":32000000.0,"unit":"parameters","observationPeriod":"2016-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"To examine the effect of recurrence depth we train RHNs with fixed total parameters (32 M)\""},{"id":"epoch:03f28a809cb64ef41fe6:training-compute","entityId":"epoch:03f28a809cb64ef41fe6","metric":"training-compute","value":3570000000000000.0,"unit":"FLOP","observationPeriod":"2016-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 32000000 parameters * 929000 tokens * 20 epochs = 3.56736e+15 FLOP"},{"id":"epoch:2e3116509dd7e15ad0b2:parameters","entityId":"epoch:2e3116509dd7e15ad0b2","metric":"parameters","value":23000000.0,"unit":"parameters","observationPeriod":"2016-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2e3116509dd7e15ad0b2:training-compute","entityId":"epoch:2e3116509dd7e15ad0b2","metric":"training-compute","value":1.28e+17,"unit":"FLOP","observationPeriod":"2016-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*23000000*929000*1000=1.28202e+17"},{"id":"epoch:8c6f9910ffa3202b354c:parameters","entityId":"epoch:8c6f9910ffa3202b354c","metric":"parameters","value":1280000.0,"unit":"parameters","observationPeriod":"2016-07-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"If a graph has 10,000 nodes and the embeddings are 128-dimensional, then the total number of parameters would be:\nP=10,000\u00d7128=1,280,000"},{"id":"epoch:141ab822d413f7c14ae8:training-compute","entityId":"epoch:141ab822d413f7c14ae8","metric":"training-compute","value":7.1935776e+17,"unit":"FLOP","observationPeriod":"2016-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1,464  images in 2012 VOC (https://paperswithcode.com/dataset/pascal-voc)/\n9,963 images in 2007 VOC (https://www.tensorflow.org/datasets/catalog/voc)\n83K training images in MS COCO  (https://paperswithcode.com/dataset/coco)\n\nThey used a Nvidia K40 GPU and report training time/image in seconds (table 3)\n\nAssumed a 0.33 util rate\n\nSection 4.2 (MS COCO):\n\"Next we evaluate on the MS COCO dataset [ 13 ] that has 80 object categories. Our experiments involve the 80k train set, 40k val set, and 20k test-dev set. We set the learning rate as 0.001 for 90k iterations and 0.0001 for next 30k iterations, with an effective mini-batch size of 8.\"\n0.45s K40 training time per image (table 3)\nK40 has 5046000000000 FLOP/s\n\nTotal examples: 120000*8=960000 (12 epochs)\n5046000000000 FLOP/s * 960000 images * 0.45 s/image * 0.33 utilization = 719357760000000000 FLOP"},{"id":"epoch:29435812d8f5d1f8e747:parameters","entityId":"epoch:29435812d8f5d1f8e747","metric":"parameters","value":138000000.0,"unit":"parameters","observationPeriod":"2016-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Based on VGG16 architecture, actual parameter count is slightly larger since the model contains some small added layers which are not described in detail."},{"id":"epoch:24f833b36568005541d4:training-compute","entityId":"epoch:24f833b36568005541d4","metric":"training-compute","value":1.454112e+17,"unit":"FLOP","observationPeriod":"2016-05-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"12 hours of training for POS tagging\nGeForce GTX TITAN X GPU\n0.33 utilization rate\n"},{"id":"epoch:c9be72f58d277c37062c:parameters","entityId":"epoch:c9be72f58d277c37062c","metric":"parameters","value":138000000.0,"unit":"parameters","observationPeriod":"2016-05-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"VGG16 with some added layers which are not specified in detail."},{"id":"epoch:e038372211a14879fbec:parameters","entityId":"epoch:e038372211a14879fbec","metric":"parameters","value":8970000.0,"unit":"parameters","observationPeriod":"2016-04-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:2f1966e87dfd71bfeaeb:parameters","entityId":"epoch:2f1966e87dfd71bfeaeb","metric":"parameters","value":1700000.0,"unit":"parameters","observationPeriod":"2016-04-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Same parameter count as DQN"},{"id":"epoch:7e4e1de28b28d465d2b7:parameters","entityId":"epoch:7e4e1de28b28d465d2b7","metric":"parameters","value":37000000.0,"unit":"parameters","observationPeriod":"2016-03-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Parameter count is not explicitly stated, but they give details:\n\n\"The MLP we train on MNIST consists of 3 hidden layers of 4096 binary units (see Section 1) and a L2-SVM output layer\"\n\nApproximately 37m, based on 784 pixels * 4096 + 2 * 4096^2"},{"id":"epoch:50b0fd8d9791f25fd9e9:parameters","entityId":"epoch:50b0fd8d9791f25fd9e9","metric":"parameters","value":138000000.0,"unit":"parameters","observationPeriod":"2016-03-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Trains SVM on top of VGG-Face (VGG16 architecture). SVM parameters are not included in the estimate."},{"id":"epoch:4784d9c42a60c1b3b424:training-compute","entityId":"epoch:4784d9c42a60c1b3b424","metric":"training-compute","value":9.69408e+16,"unit":"FLOP","observationPeriod":"2016-03-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8 hours of training for NER\nGeForce GTX TITAN X GPU\n0.33 utilization rate\n"},{"id":"epoch:884bca6c3d5783755bb2:parameters","entityId":"epoch:884bca6c3d5783755bb2","metric":"parameters","value":1500000.0,"unit":"parameters","observationPeriod":"2016-03-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"approximately 1.5M parameters\nin total\""},{"id":"epoch:7df4c20c3788ca120b2b:parameters","entityId":"epoch:7df4c20c3788ca120b2b","metric":"parameters","value":1200000.0,"unit":"parameters","observationPeriod":"2016-02-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The paper says \"SqueezeNet achieves AlexNet-level accuracy on ImageNet with 50x fewer parameters.\"\n\nAlexNet has 60 million parameters."},{"id":"epoch:79d7174b174923928d15:parameters","entityId":"epoch:79d7174b174923928d15","metric":"parameters","value":56000000.0,"unit":"parameters","observationPeriod":"2016-02-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:d17a5ddb1e2df9261883:parameters","entityId":"epoch:d17a5ddb1e2df9261883","metric":"parameters","value":43000000.0,"unit":"parameters","observationPeriod":"2016-02-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The folks from Google strike again with Inception-v4, 43M parameters.\"\n\nhttps://towardsdatascience.com/illustrated-10-cnn-architectures-95d78ace614d"},{"id":"epoch:54fd7af4f271eb5d8ef1:parameters","entityId":"epoch:54fd7af4f271eb5d8ef1","metric":"parameters","value":1040000000.0,"unit":"parameters","observationPeriod":"2016-02-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.04B (Table 1)"},{"id":"epoch:54fd7af4f271eb5d8ef1:training-compute","entityId":"epoch:54fd7af4f271eb5d8ef1","metric":"training-compute","value":1.0220883e+20,"unit":"FLOP","observationPeriod":"2016-02-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"assuming (!) 10 days similarly to another model in the paper\n240*3600*5046000000000*32*0.3 = 4.1853542e+19\n\nassuming (!) 50 epochs similarly to another model in the paper\n6*1.04*10^9*0.8*10^9*50 = 2.496e+20\n\nsqrt(4.1853542e+19*2.496e+20) = 1.0220883e+20"},{"id":"epoch:10faca516e5f866b2565:parameters","entityId":"epoch:10faca516e5f866b2565","metric":"parameters","value":1040000000.0,"unit":"parameters","observationPeriod":"2016-02-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.04B (Table 1)"},{"id":"epoch:10faca516e5f866b2565:training-compute","entityId":"epoch:10faca516e5f866b2565","metric":"training-compute","value":8.7892439e+19,"unit":"FLOP","observationPeriod":"2016-02-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"504*3600*5046000000000*32*0.3 =8.7892439e+19 "},{"id":"epoch:b52941072e1d41f490ae:training-compute","entityId":"epoch:b52941072e1d41f490ae","metric":"training-compute","value":1.9e+21,"unit":"FLOP","observationPeriod":"2016-01-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"This number is pretty uncertain. I expect it to be right to around a factor of 3, at least compared to AlphaGo Fan.\n\nThe architecture used was pretty much the same as AlphaGo Fan, but it was \"trained for longer\" and had around 5.33x the number of convolutional layers of AlphaGo Fan (256/48 = 5.33). \n\nThe convolutional layers are the major contributor to the training compute, so I somewhat arbitrarily just multiply the compute for AlphaGo Fan by 5. Thus 3.8e20 * 5 = 1.9e21\n\nOtherwise there has been little said about this model specifically - I've mainly relied on the source for AlphaGo Zero and AlphaGo Fan, linked below\n\nAlphaGo Fan: https://www.nature.com/articles/nature16961\n\nAlphaGo Zero: https://www.nature.com/articles/nature24270.epdf?author_access_token=VJXbVjaSHxFoctQQ4p2k4tRgN0jAjWel9jnR3ZoTv0PVW4gB86EEpGqTRDtpIz-2rmo8-KG06gqVobU5NSCFeHILHcVFUeMsbvwS-lxjqQGg98faovwjxeTUgZAUMnRQ"},{"id":"epoch:2052a6346e5d7657d838:parameters","entityId":"epoch:2052a6346e5d7657d838","metric":"parameters","value":66000000.0,"unit":"parameters","observationPeriod":"2015-12-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"66M according to https://arxiv.org/pdf/1611.01462"},{"id":"epoch:2052a6346e5d7657d838:training-compute","entityId":"epoch:2052a6346e5d7657d838","metric":"training-compute","value":5886144000000000.0,"unit":"FLOP","observationPeriod":"2015-12-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 66000000 parameters * 929000 tokens * 16 epochs = 5.886144e+15 FLOP"},{"id":"epoch:5c94da8f78f5d0254203:parameters","entityId":"epoch:5c94da8f78f5d0254203","metric":"parameters","value":60200000.0,"unit":"parameters","observationPeriod":"2015-12-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Taken from https://arxiv.org/abs/1605.07146"},{"id":"epoch:5c94da8f78f5d0254203:training-compute","entityId":"epoch:5c94da8f78f5d0254203","metric":"training-compute","value":1.041408e+19,"unit":"FLOP","observationPeriod":"2015-12-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"11.3 *10^9 mult-adds per forward pass (Table 1)\n2 FLOPS/ mult-add\n3 for forward & backward pass\n1.2 * 10^6 examples in dataset\n128 epochs\n\n-> 1.041408 \u00d7 10^19 FLOP\n\nAuthors of \"AI and Memory Wall\" (https://github.com/amirgholami/ai_and_memory_wall) estimated model's training compute as 11,000 PFLOP = 1.1*10^19 FLOP"},{"id":"epoch:e2be9ac83c3943b939c1:parameters","entityId":"epoch:e2be9ac83c3943b939c1","metric":"parameters","value":1700000.0,"unit":"parameters","observationPeriod":"2015-12-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 6"},{"id":"epoch:08737f064f1922ba9f65:parameters","entityId":"epoch:08737f064f1922ba9f65","metric":"parameters","value":44500000.0,"unit":"parameters","observationPeriod":"2015-12-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Taken from https://arxiv.org/abs/1605.07146"},{"id":"epoch:08737f064f1922ba9f65:training-compute","entityId":"epoch:08737f064f1922ba9f65","metric":"training-compute","value":7.004e+18,"unit":"FLOP","observationPeriod":"2015-12-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Forward FLOP: 15200000000\n120 epochs\n1280000*120*3*15200000000=7.00416e+18\n"},{"id":"epoch:6659e18078054161429a:parameters","entityId":"epoch:6659e18078054161429a","metric":"parameters","value":38000000.0,"unit":"parameters","observationPeriod":"2015-12-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"All networks have 38 million parameters."},{"id":"epoch:6659e18078054161429a:training-compute","entityId":"epoch:6659e18078054161429a","metric":"training-compute","value":2.6e+19,"unit":"FLOP","observationPeriod":"2015-12-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1 timestep = (1280 hidden units)^2 * (7 RNN layers * 4 matrices for bidirectional + 2 DNN layers) * (2 for doubling parameters from 36M to 72M) = 98 MFLOP\n20 epochs * 12,000 hours * 3600 seconds/hour * 50 samples/sec * 98 MFLOP * 3 add-multiply * 2 backprop \n= 26,000 PF = 0.30 pfs-days\n\nSee also AI and Compute by Dario Amodei and OpenAI https://openai.com/research/ai-and-compute"},{"id":"epoch:ab99247fe891263a5eac:parameters","entityId":"epoch:ab99247fe891263a5eac","metric":"parameters","value":23626728.0,"unit":"parameters","observationPeriod":"2015-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 3 from Xception paper"},{"id":"epoch:ab99247fe891263a5eac:training-compute","entityId":"epoch:ab99247fe891263a5eac","metric":"training-compute","value":1e+20,"unit":"FLOP","observationPeriod":"2015-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Authors of \"AI and Memory Wall\" (https://github.com/amirgholami/ai_and_memory_wall) estimated model's training compute as 100,000 PFLOP = 1*10^20 FLOP"},{"id":"epoch:d8ba413bf204232c15d7:parameters","entityId":"epoch:d8ba413bf204232c15d7","metric":"parameters","value":2300000.0,"unit":"parameters","observationPeriod":"2015-11-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2.3M from Table 2."},{"id":"epoch:e7ddfd6de790f5f78d3f:parameters","entityId":"epoch:e7ddfd6de790f5f78d3f","metric":"parameters","value":5355584.0,"unit":"parameters","observationPeriod":"2015-11-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Parameters based on Figure 2:\nCNN layers: 4*4*16*6+4*4*32*16=9728\nLocal CNN: 3*3*48*32*11*11+3*3*64*49*9*9=3958848\nFC: 9*9*64*256+234*256=1387008\nTotal: 9728+3958848+1387008=5355584\n"},{"id":"epoch:8fcbf30faabda8b8c9a6:parameters","entityId":"epoch:8fcbf30faabda8b8c9a6","metric":"parameters","value":138000000.0,"unit":"parameters","observationPeriod":"2015-10-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Taken from VGG16 base model, ignoring minor architectural changes"},{"id":"epoch:8fcbf30faabda8b8c9a6:training-compute","entityId":"epoch:8fcbf30faabda8b8c9a6","metric":"training-compute","value":1.2311081250000001e+19,"unit":"FLOP","observationPeriod":"2015-10-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Base model: 12291000000000002000\nBase model inference FLOP: 15300000000\nFinetune for 7 epochs with 62500 training examples\n15300000000*3*62500*7=20081250000000000=2e16\n\nTotal compute: 12291000000000002000+20081250000000000=12311081250000002000"},{"id":"epoch:9947a60096ef7335307a:parameters","entityId":"epoch:9947a60096ef7335307a","metric":"parameters","value":8209984.0,"unit":"parameters","observationPeriod":"2015-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The input to the policy network is a 19\u2009\u00d7\u200919\u2009\u00d7\u200948 image stack consisting of 48 feature planes. The first hidden layer zero pads the input into a 23\u2009\u00d7\u200923 image, then convolves k filters of kernel size 5\u2009\u00d7\u20095 with stride 1 with the input image and applies a rectifier nonlinearity. Each of the subsequent hidden layers 2 to 12 zero pads the respective previous hidden layer into a 21\u2009\u00d7\u200921 image, then convolves k filters of kernel size 3\u2009\u00d7\u20093 with stride 1, again followed by a rectifier nonlinearity. The final layer convolves 1 filter of kernel size 1\u2009\u00d7\u20091 with stride 1, with a different bias for each position, and applies a softmax function. The match version of AlphaGo used k\u2009=\u2009192 filters; Fig. 2b and Extended Data Table 3 additionally show the results of training with k\u2009=\u2009128, 256 and 384 filters.\n\nThe input to the value network is also a 19\u2009\u00d7\u200919\u2009\u00d7\u200948 image stack, with an additional binary feature plane describing the current colour to play. Hidden layers 2 to 11 are identical to the policy network, hidden layer 12 is an additional convolution layer, hidden layer 13 convolves 1 filter of kernel size 1\u2009\u00d7\u20091 with stride 1, and hidden layer 14 is a fully connected linear layer with 256 rectifier units. The output layer is a fully connected linear layer with a single tanh unit."},{"id":"epoch:9947a60096ef7335307a:training-compute","entityId":"epoch:9947a60096ef7335307a","metric":"training-compute","value":3.8e+20,"unit":"FLOP","observationPeriod":"2015-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assume 0.3 utilisation rate, 1e13 GPU FLOP/s [single precision]. Trained in three stages using 50 GPUs over 3 weeks + 1 day + 1 week\n\nTraining compute = (50 GPUs)(29 days)(86400s/day)(0.3 utilisation rate)(1e13 FLOP/s) = 3.8e20 FLOPs\n"},{"id":"epoch:3519204ff0e0a111cf77:parameters","entityId":"epoch:3519204ff0e0a111cf77","metric":"parameters","value":19000000.0,"unit":"parameters","observationPeriod":"2015-08-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"19M (Table 3)"},{"id":"epoch:3519204ff0e0a111cf77:training-compute","entityId":"epoch:3519204ff0e0a111cf77","metric":"training-compute","value":2650000000000000.0,"unit":"FLOP","observationPeriod":"2015-08-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 19000000 parameters * 929000 tokens * 25 epochs = 2.64765e+15 FLOP"},{"id":"epoch:3e33333dca03ae421dc9:parameters","entityId":"epoch:3e33333dca03ae421dc9","metric":"parameters","value":5006000.0,"unit":"parameters","observationPeriod":"2015-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:3e33333dca03ae421dc9:training-compute","entityId":"epoch:3e33333dca03ae421dc9","metric":"training-compute","value":4.8098691e+17,"unit":"FLOP","observationPeriod":"2015-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Total compute: 1535436288 forward pass flops *3* 1000000 iterations * 128 batch size = 589607534592000000 FLOP = 5.89607534592e17 FLOP\n\nForward pass FLOPs based on CNN details in Table 1:\nConv11: 2*3*3*32*100*100*1=5760000\nConv12: 2*3*3*64*100*100*32=368640000\nConv21: 2*3*3*64*50*50*64=184320000\nConv22: 2*3*3*64*50*50*128=368640000\nConv32: 2*3*3*128*25*25*96=138240000\nConv33: 2*3*3*96*25*25*192=207360000\nConv41: 2*3*3*192*13*13*128=74760192\nConv42: 2*3*3*128*13*13*256=99680256\nConv51: 2*3*3*256*7*7*160=36126720\nConv52: 2*3*3*160*7*7*320=45158400\nFc6: 2*320*10548=6750720\nTotal forward pass FLOPs: 5760000+368640000+184320000+368640000+138240000+207360000+74760192+99680256+36126720+45158400+6750720=1535436288\n\n\n1682000000000 FLOP / GPU / sec * 1 GPU * 9 days * 24 hours / day * 3600 sec / hour * 0.3 [assumed utilization] = 3.9237696e+17 FLOP\n\nsqrt(5.89607534592e17 * 3.9237696e+17) = 4.8098691e+17"},{"id":"epoch:fc67fc7a09cb87640566:parameters","entityId":"epoch:fc67fc7a09cb87640566","metric":"parameters","value":5006000.0,"unit":"parameters","observationPeriod":"2015-07-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Taken from Table 1 of https://arxiv.org/abs/1508.01722 which uses the same architecture"},{"id":"epoch:73782472b5953476ccc3:parameters","entityId":"epoch:73782472b5953476ccc3","metric":"parameters","value":20000000.0,"unit":"parameters","observationPeriod":"2015-07-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"20M (Table 3)"},{"id":"epoch:73782472b5953476ccc3:training-compute","entityId":"epoch:73782472b5953476ccc3","metric":"training-compute","value":3340000000000000.0,"unit":"FLOP","observationPeriod":"2015-07-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 20000000 parameters * 929000 tokens * 30 epochs = 3.3444e+15 FLOP"},{"id":"epoch:6ce4de232b1fa39f3830:parameters","entityId":"epoch:6ce4de232b1fa39f3830","metric":"parameters","value":13600000.0,"unit":"parameters","observationPeriod":"2015-06-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The network contains 13.6 \u00b7 106 parameters\""},{"id":"epoch:7a8d4ec1b9bb7a5d6681:parameters","entityId":"epoch:7a8d4ec1b9bb7a5d6681","metric":"parameters","value":271684800.0,"unit":"parameters","observationPeriod":"2015-06-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Calculation based on figure 3 of the paper:\n7 * 7 * 3 * 64 + 3 * 3 * 64 * 192 + 1 * 1 * 192 * 128 + 3 * 3 * 128 * 256 + 1 * 1 * 256 * 256 + 3 * 3 * 256 * 512 + 4 * (1 * 1 * 512 * 256 + 3 * 3 * 256 * 512) + 1 * 1 * 512 * 512 + 3 * 3 * 512 * 1024 + 2 * (1 * 1 * 1024 * 512 + 3 * 3 * 512 * 1024) + 4 * (3 * 3 * 1024 * 1024) + 7 * 7 * 1024 * 4096 + 4096 * 7 * 7 * 30"},{"id":"epoch:3329d781825becd2a3a6:parameters","entityId":"epoch:3329d781825becd2a3a6","metric":"parameters","value":17408.0,"unit":"parameters","observationPeriod":"2015-06-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The paper trains two regression models, each using hardcoded SIFT / BRIEF features. \n68 keypoints (given in the paper) and 128 dimensional features per keypoint (assumption based on SIFT features)\n2*68*128=17408"},{"id":"epoch:4aacb751d1786c378822:parameters","entityId":"epoch:4aacb751d1786c378822","metric":"parameters","value":9106245.0,"unit":"parameters","observationPeriod":"2015-05-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The input layer takes either a single RGB frame (224x224x3) for the spatial stream or a stack of 10 optical flow frames (224x224x20) for the temporal stream.\nThe first convolutional layer has 96 filters of size 7x7 with stride 2.\nThis is followed by max pooling with size 3x3 and stride 2.\nThe second convolutional layer has 256 filters of size 5x5 with stride 2.\nAfter that is another max pooling layer (3x3, stride 2).\nThe third convolutional layer has 512 filters of size 3x3 with stride 1.\nThe fourth convolutional layer has 512 filters of size 3x3 with stride 1.\nThe fifth convolutional layer has 512 filters of size 3x3 with stride 1.\nNext is a max pooling layer (3x3, stride 2).\nThe fully-connected layers have 4096, 2048, and 101 neurons respectively.\n\n(7*7*20+1)*96 + (5*5*20+1)*256 + (3*3*20+1)*512 + (3*3*20+1)*512 + (3*3*20+1)*512 + 2*4096 + (4096+1)*2048 + (2048+1)*101 = 9106245"},{"id":"epoch:23951cd2ba14cc493e22:parameters","entityId":"epoch:23951cd2ba14cc493e22","metric":"parameters","value":37676160.0,"unit":"parameters","observationPeriod":"2015-05-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 1: 2*(1*3*3*64+64*3*3*64+64*3*3*128+128*3*3*128+128*3*3*256+256*3*3*256+256*3*3*512+512*3*3*512+512*3*3*1024+1024*3*3*1024)=37676160"},{"id":"epoch:23951cd2ba14cc493e22:training-compute","entityId":"epoch:23951cd2ba14cc493e22","metric":"training-compute","value":5.0832252e+16,"unit":"FLOP","observationPeriod":"2015-05-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"10*60*60*4706690000000*0.3=50832252000000000\n\u201ctraining time of only 10 hours on a NVidia Titan GPU (6 GB).\u201d\nUsing single precision FLOP: 4706.69 GFLOPS (https://www.gpuzoo.com/GPU-NVIDIA/GeForce_GTX_TITAN.html) "},{"id":"epoch:58a64ecb2d17ce27c9bf:parameters","entityId":"epoch:58a64ecb2d17ce27c9bf","metric":"parameters","value":6000000.0,"unit":"parameters","observationPeriod":"2015-05-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:a3ffe4bd3015c35bfdfc:parameters","entityId":"epoch:a3ffe4bd3015c35bfdfc","metric":"parameters","value":18413568.0,"unit":"parameters","observationPeriod":"2015-04-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3*40 inputs - 2L DNN (2048) - BLSTM (128*2) - 2L DNN (2048) - 3431 outputs\n3*40*2048 + 2048*2048 + 2*4*(2048+128)*128 + 256*2048 + 2048*2048 + 2048*3431=18413568\n"},{"id":"epoch:a3ffe4bd3015c35bfdfc:training-compute","entityId":"epoch:a3ffe4bd3015c35bfdfc","metric":"training-compute","value":1.9410191999999997e+17,"unit":"FLOP","observationPeriod":"2015-04-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"51 hours in wall clock time with a NVIDIA Tesla K20 GPU\"\nK20 FLOPs: 3524000000000\nCompute: 0.3*51*60*60*3524000000000=194101919999999970= 1.94e17\n"},{"id":"epoch:74f4a897cb30a1d0da80:parameters","entityId":"epoch:74f4a897cb30a1d0da80","metric":"parameters","value":8000000.0,"unit":"parameters","observationPeriod":"2015-03-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8M according to https://arxiv.org/pdf/1508.06615"},{"id":"epoch:74f4a897cb30a1d0da80:training-compute","entityId":"epoch:74f4a897cb30a1d0da80","metric":"training-compute","value":3.4153451e+16,"unit":"FLOP","observationPeriod":"2015-03-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"5046000000000 FLOP / sec/ GPU * 1 GPU * 48 hours [\"Likely\" confidence since 2 days of training refer to another dataset] * 3600 sec / hour * 0.3 [assumed utilization] = 2.6158464e+17 FLOP\n\nAssuming (!) 100 epochs (-> \"Speculative\" confidence\"):\n\n6 FLOP / parameter / token * 8000000 parameters * 929000 tokens * 100 epochs = 4.4592e+15 FLOP\n\nsqrt(2.6158464e+17*4.4592e+15 ) = 3.4153451e+16 FLOP\n\n________\nin the algorithmic progress report paper the estimation was 7.3 \u00d7 10^16 FLOP (hardware-based estimation assuming another Tesla K40 chip)\n"},{"id":"epoch:05923a966d81eeece817:parameters","entityId":"epoch:05923a966d81eeece817","metric":"parameters","value":2009999.9999999998,"unit":"parameters","observationPeriod":"2015-03-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:249d03e7324537e50385:parameters","entityId":"epoch:249d03e7324537e50385","metric":"parameters","value":205190.0,"unit":"parameters","observationPeriod":"2015-02-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1\n\nhttps://arxiv.org/abs/1503.00075"},{"id":"epoch:a98816e061cf1ce2517e:parameters","entityId":"epoch:a98816e061cf1ce2517e","metric":"parameters","value":1693362.0,"unit":"parameters","observationPeriod":"2015-02-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The input to the neural network consists of an 84x84x4 image produced by the preprocess-ing mapw. The first hidden layer convolves 32 filters of 8x8 with stride 4 with theinput image and applies a rectifier nonlinearity. The second hidden layer con-volves 64 filters of 4x4 with stride 2, again followed by a rectifier nonlinearity.This is followedby a thirdconvolutional layer thatconvolves 64 filtersof 3x3 withstride 1 followed by a rectifier. The final hidden layer is fully-connected and con-sists of 512 rectifier units. The output layer is a fully-connected linear layer with asingle output for each valid action. The number of valid actions varied between 4 and 18 on the games we considered.\"\n\nExample num params here: https://colab.research.google.com/drive/1Ty6SFYWd7EcKoxJohucL2OdiLR_3oXnI?usp=sharing"},{"id":"epoch:d6c7d43dd2194ba260bf:parameters","entityId":"epoch:d6c7d43dd2194ba260bf","metric":"parameters","value":33500.0,"unit":"parameters","observationPeriod":"2015-02-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:787858b685ff69a56935:parameters","entityId":"epoch:787858b685ff69a56935","metric":"parameters","value":87048800.0,"unit":"parameters","observationPeriod":"2015-02-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"I used the architecture in table 3\nI ignored biases, and assumed a SPP bin size of 256\n\n3*7*7*96+96*3*3*384+384*3*3*384*5+384*3*3*768+768*3*3*768*5+768*3*3*896+896*3*3*896*5+896*(7*7+3*3+2*2+1)*4096+4096*4096+4096*1000=330581792\n\n\n\n"},{"id":"epoch:787858b685ff69a56935:training-compute","entityId":"epoch:787858b685ff69a56935","metric":"training-compute","value":2.397403008e+19,"unit":"FLOP","observationPeriod":"2015-02-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"training C on eight K40 GPUs, takes about 3-4 weeks\"\n0.33 util rate\n(From Imagenet paper-data, Besiroglu et al., forthcoming) "},{"id":"epoch:86f312af2cb789cec1d9:parameters","entityId":"epoch:86f312af2cb789cec1d9","metric":"parameters","value":138000000.0,"unit":"parameters","observationPeriod":"2015-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Their largest model uses the VGG16 architecture"},{"id":"epoch:59823beb8dfed9d942ec:parameters","entityId":"epoch:59823beb8dfed9d942ec","metric":"parameters","value":26500000.0,"unit":"parameters","observationPeriod":"2014-12-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:0d689a1678ca02a55470:parameters","entityId":"epoch:0d689a1678ca02a55470","metric":"parameters","value":2370000.0,"unit":"parameters","observationPeriod":"2014-12-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"CIFAR-10 with c64-c64-c128-1000 architecture\n\n\"Our CNN architecture has three alternating stages of 5x5 convolution filters and 3x3 max pooling with stride of 2 that are followed by a fully connected layer of 1000 rectified linear hidden units (ReLU\u2019s). \"\n\n\"Training cost over 45 epochs. CIFAR-10 with c64-c64-c128-1000 architecture.\"\n\nCIFAR-10's input dimension is 32x32x3\n\nParameter:\n3*64*5*5+64*64*5*5+64*128*5*5+128*4*4*1000+1000*10=2370000\n\n\n\n"},{"id":"epoch:0d689a1678ca02a55470:training-compute","entityId":"epoch:0d689a1678ca02a55470","metric":"training-compute","value":624979800000000.0,"unit":"FLOP","observationPeriod":"2014-12-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From https://openai.com/blog/ai-and-compute/ Appendix\n\nless than 0.0007 pfs-days (86400*10^15*0.0007)\n\nManual estimate: \n3 (forward-backward adjustment) * 92589600 flop per forward pass * 50000 examples * 45 epochs = 624979800000000 (6.25e14)\n\n"},{"id":"epoch:348578040036e69ae67a:parameters","entityId":"epoch:348578040036e69ae67a","metric":"parameters","value":16100000.000000002,"unit":"parameters","observationPeriod":"2014-12-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:eacc347884e5333048af:parameters","entityId":"epoch:eacc347884e5333048af","metric":"parameters","value":27000000.0,"unit":"parameters","observationPeriod":"2014-12-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"27M weights in largest CIFAR-100 model"},{"id":"epoch:eacc347884e5333048af:training-compute","entityId":"epoch:eacc347884e5333048af","metric":"training-compute","value":1e+17,"unit":"FLOP","observationPeriod":"2014-12-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"For the 12M param model, training required \"18 hours on a GeForce GTX 780\". So would be somewhat larger for 27M.\n\n4 TFLOPS * 18 * 3600 * 0.4 = 1e17"},{"id":"epoch:c1baae40961dc434e024:parameters","entityId":"epoch:c1baae40961dc434e024","metric":"parameters","value":62000000000.0,"unit":"parameters","observationPeriod":"2014-12-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"62B from Table 2"},{"id":"epoch:c1baae40961dc434e024:training-compute","entityId":"epoch:c1baae40961dc434e024","metric":"training-compute","value":2.97600000001e+20,"unit":"FLOP","observationPeriod":"2014-12-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.wolframalpha.com/input?i=0.8+billion+*+62+billion+*+6+FLOP"},{"id":"epoch:4b2273d35ad59c8b748e:parameters","entityId":"epoch:4b2273d35ad59c8b748e","metric":"parameters","value":706048.0,"unit":"parameters","observationPeriod":"2014-11-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Architecture details in Figure 4\nConv1: 3*7*7*32=4704\nConv2: 32*5*5*48=38400Conv3: 48*3*3*64=27648\nConv4: 64*3*3*96=55296\nFc5: 5*2*96*500=480000\nFc6: 500*200=100000\nTotal: 4704+38400+27648+55296+480000+100000=706048\n"},{"id":"epoch:4b2273d35ad59c8b748e:training-compute","entityId":"epoch:4b2273d35ad59c8b748e","metric":"training-compute","value":1.0854e+16,"unit":"FLOP","observationPeriod":"2014-11-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training time: 3h\nAssumed FP32 GPU FLOPs in 2014: 3.35e+12Assumed utilization: 0.3\nTraining FLOP: 3*60*60*0.3*3.35e12=10854000000000000=1.08e16\n"},{"id":"epoch:e409d88ff37eb2dc904b:parameters","entityId":"epoch:e409d88ff37eb2dc904b","metric":"parameters","value":142552000.0,"unit":"parameters","observationPeriod":"2014-11-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1st model: CaffeNet fc6 feature extractor (4096-length vectors) -> LSTM with 1024 hidden units\n\n2nd model: CaffeNet fc6 feature extractor (4096-length vectors) -> 2 layer LSTM with 1000 hidden units\n\n3rd mode: Like the second, but has encoder and decoder LSTMs (both with 2 layers)\n\nAlexNet (close relative to CaffeNet) has 61M params.\n\nLSTM RNN number of parameters is given by L*(n*m + n^2 + n) where L:= Number of layers, n:= hidden units, m:= input vector length\n"},{"id":"epoch:f243365f1343b7b4582b:parameters","entityId":"epoch:f243365f1343b7b4582b","metric":"parameters","value":6797700.0,"unit":"parameters","observationPeriod":"2014-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Computed summing the parameters on table 1 of section 5"},{"id":"epoch:f243365f1343b7b4582b:training-compute","entityId":"epoch:f243365f1343b7b4582b","metric":"training-compute","value":1.51e+18,"unit":"FLOP","observationPeriod":"2014-09-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"AI and Compute  (https://openai.com/blog/ai-and-compute/) charts imply a value of 1.51e18 (value extracted using WebPlotDigitizer  https://automeris.io/WebPlotDigitizer/ ).\n\nBased on the paper, there are 1.5B multiply-adds per inference, and 1.2M images in the training set, but an unknown number of epochs. They decrease the learning rate by 4% every 8 epochs, so there are likely many. If the figure from AI and Compute is taken as true, there were likely 140 epochs"},{"id":"epoch:b8a1a7cb34b7daaeb270:parameters","entityId":"epoch:b8a1a7cb34b7daaeb270","metric":"parameters","value":5000000.0,"unit":"parameters","observationPeriod":"2014-09-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Estimate from table 2 of https://arxiv.org/abs/1609.07843\n\nThe authors of the linked paper draw on estimates from table 3 of https://arxiv.org/pdf/1508.06615.pdf"},{"id":"epoch:b8a1a7cb34b7daaeb270:training-compute","entityId":"epoch:b8a1a7cb34b7daaeb270","metric":"training-compute","value":4.4e+16,"unit":"FLOP","observationPeriod":"2014-09-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"40h, 1 GPU, 1028e9 Peak FLOP/s, 30%\n\n1028000000000 FLOP/s/GPU * 1GPU * 40 hours * 3600 s/hour * 0.3 [assumed utilization] = 4.44096e+16 FLOP"},{"id":"epoch:cac7b5478f714b05e606:parameters","entityId":"epoch:cac7b5478f714b05e606","metric":"parameters","value":1920000000.0,"unit":"parameters","observationPeriod":"2014-09-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The resulting LSTM has 384M parameters of which 64M are pure recurrent connections (32M for the \u201cencoder\u201d LSTM and 32M\nfor the \u201cdecoder\u201d LSTM).\nThe paper uses an ensemble of 5 LSTMs."},{"id":"epoch:cac7b5478f714b05e606:training-compute","entityId":"epoch:cac7b5478f714b05e606","metric":"training-compute","value":5.6e+19,"unit":"FLOP","observationPeriod":"2014-09-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"384E+6 parameters * 2 FLOP/parameter * (348E+6 + 304E+6 points per epoch) * 7.5 epochs * 3 FLOP/point ~= 1.126656e+19 FLOP\nTimes 5 independent models in ensemble => 5.6E+19 FLOP\n\nIf we assume NVIDIA K40 (in use at the time): 10 days * 24 * 60 * 60 seconds/day * 8 GPUs * 33% * 5e12 FLOP/s * 5 models in ensemble ~= 5.7E+19 FLOP\n\nAuthors of \"AI and Memory Wall\" estimated model's training compute as 11,000 PFLOPS = 1.1*10^19 FLOPS\n(https://github.com/amirgholami/ai_and_memory_wall)"},{"id":"epoch:84cf7a7c495d6ae5626d:parameters","entityId":"epoch:84cf7a7c495d6ae5626d","metric":"parameters","value":66000000.0,"unit":"parameters","observationPeriod":"2014-09-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The large LSTM has 1500 units per layer and its parameters are initialized uniformly in [\u22120.04, 0.04]\"\n\n66M according to Table 3 of https://arxiv.org/pdf/1611.01462"},{"id":"epoch:84cf7a7c495d6ae5626d:training-compute","entityId":"epoch:84cf7a7c495d6ae5626d","metric":"training-compute","value":4.2966069e+16,"unit":"FLOP","observationPeriod":"2014-09-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3520000000000 FLOP / s/ GPU [tesla k20c assumed, the paper says just \"Nvidia k20\"] * 1 GPU * 24 hours [could be less, they report \"the entire day\"] * 3600 s / hour * 0.3 [assumed utilization] = 9.12384e+16 FLOP \n\n6 FLOP / parameter / token * 66000000 parameters * 929000 tokens * 55 epochs = 2.023362e+16 FLOP\n\nsqrt(2.023362e+16*9.12384e+16) = 4.2966069e+16 "},{"id":"epoch:df736489086af36e8b4e:parameters","entityId":"epoch:df736489086af36e8b4e","metric":"parameters","value":138000000.0,"unit":"parameters","observationPeriod":"2014-09-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Source: Table 2\nhttps://arxiv.org/abs/1409.1556"},{"id":"epoch:df736489086af36e8b4e:training-compute","entityId":"epoch:df736489086af36e8b4e","metric":"training-compute","value":1.2291e+19,"unit":"FLOP","observationPeriod":"2014-09-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"3 weeks * 4 Titan Black GPUs * 0.30 utilization\n\nSection 3.3: \"On a system equipped with four NVIDIA Titan Black GPUs, training a single net took 2\u20133 weeks depending on the architecture.\"\n\nTitan Black performance: 5.645 TFLOPS (assuming FP32)\n\nhttps://www.wolframalpha.com/input?i=5.645+TFLOPS+*+3+weeks+*+4+*+0.3\n\n\n"},{"id":"epoch:4e8cbb1e375262b32016:parameters","entityId":"epoch:4e8cbb1e375262b32016","metric":"parameters","value":144000000.0,"unit":"parameters","observationPeriod":"2014-09-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Source: Table 2\nhttps://arxiv.org/abs/1409.1556"},{"id":"epoch:4e8cbb1e375262b32016:training-compute","entityId":"epoch:4e8cbb1e375262b32016","metric":"training-compute","value":1.1e+19,"unit":"FLOP","observationPeriod":"2014-09-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Authors of \"AI and Memory Wall\" (https://github.com/amirgholami/ai_and_memory_wall) estimated model's training compute as 11,000 PFLOP = 1.1*10^19 FLOP"},{"id":"epoch:6cc235593bdedecf088e:training-compute","entityId":"epoch:6cc235593bdedecf088e","metric":"training-compute","value":1.5552e+18,"unit":"FLOP","observationPeriod":"2014-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From https://openai.com/blog/ai-and-compute/ Appendix.\n\n0.018 pfs-days\n(86400*10^15*0.018)\n\n252 hours in a Quadro K-6000 GPU (assumed utilization: 0.33)\n\n5196000000000 FLOP/s *252 hours * 3600 second/hour * 0.33 utilization = 1555200000000000000 FLOP"},{"id":"epoch:bb110c7637e92374b543:parameters","entityId":"epoch:bb110c7637e92374b543","metric":"parameters","value":313856.0,"unit":"parameters","observationPeriod":"2014-08-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our final detector contains 1,226 deep quadratic trees\"\n\"(depth of eight in this paper)\"\nParameters of the trained decision trees: 1226*(2^8)=313856\n"},{"id":"epoch:247403985cc7db1edd41:parameters","entityId":"epoch:247403985cc7db1edd41","metric":"parameters","value":6144.0,"unit":"parameters","observationPeriod":"2014-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The paper trains a decision tree ensemble.\n\u201cwe choose 2048 as the number of weak classifiers contained in the soft cascade. As each weak classifier is a depth-2 decision tree, it takes only two comparing operations to apply a weak classifier, which is quite fast.\u201d\nParameters: 2048*(2^2-1)=6144\n"},{"id":"epoch:247403985cc7db1edd41:training-compute","entityId":"epoch:247403985cc7db1edd41","metric":"training-compute","value":76380000000000.0,"unit":"FLOP","observationPeriod":"2014-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training compute: 10.2 minutes * 4 * 3.9* 10^9 cycles per second * 16 FLOP per cycle * 0.5 = 7.638*10^13 FLOPs\n\u201con a PC with Intel Core i7-3770 CPU and 16GB RAM\u201d\nIvy Bridge, 4 cores, 3.9GHz, 16 FP32 FLOP/cycle (https://en.wikipedia.org/wiki/Floating_point_operations_per_second ) \n\u201cand 10.2 mins for multi-scale version.\u201d"},{"id":"epoch:dbb1ce69859f6fd4d9b3:training-compute","entityId":"epoch:dbb1ce69859f6fd4d9b3","metric":"training-compute","value":6.9e+16,"unit":"FLOP","observationPeriod":"2014-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Each three-player agent was trained for about 12 billion episodes, requiring about 48 hours of training time [...] on a modern computer without using parallelization\"\n\nAssume an Intel i7 so 400e9 FLOP/s.\n6.9e16 = 400e9*60*60*48"},{"id":"epoch:2a14b45be254466a1728:parameters","entityId":"epoch:2a14b45be254466a1728","metric":"parameters","value":126125568.0,"unit":"parameters","observationPeriod":"2014-06-23","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Using shorthand notation, the full [single frame] architecture is C(96, 11, 3)-N-P-C(256, 5, 1)-N-P-C(384, 3, 1)-C(384, 3, 1)-C(256, 3, 1)-P-FC(4096)-FC(4096), where C(d, f, s) indicates a convolutional layer with d filters of spatial size f \u00d7f, applied to the input with stride s\"\n\nTwo such single-frame architectures are concatenated as shown in figure 2\n\n\"Since the input is only of half the\nspatial size as the full-frame models, we take out the last\npooling layer to ensure that both streams still terminate in a\nlayer of size 7\u00d77\u00d7256. \"\n\nWe assume the input are T=10 frames with C=3 color channels each\n\n2*(256*(10*3*5*5+1) + 384*(256*3*3+1) + 384*(384*3*3+1) + 256*(384*3*3+1)) + (2*7*7*256 + 1)*4096 + (4096+1)*4096\n\n\n\n"},{"id":"epoch:0c11c9c0ca4c5768469a:parameters","entityId":"epoch:0c11c9c0ca4c5768469a","metric":"parameters","value":26500000.0,"unit":"parameters","observationPeriod":"2014-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The network had five levels of bidirectional LSTM hidden layers, with 500 cells in each layer, giving a total of \u223c 26.5M weights.\""},{"id":"epoch:b9b7831a71a3ef061891:parameters","entityId":"epoch:b9b7831a71a3ef061891","metric":"parameters","value":144496000.0,"unit":"parameters","observationPeriod":"2014-06-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Model contains a word embedding. a matrix combining two word embeddings, and image embedding (built upon a pretrained RCNN image model.\nWord embedding: 400000 * 200 =80000000 (\"Here, We is a d \u00d7 400, 000 matrix that encodes a 1-of-k vector into a d-dimensional word vector representation (we use d = 200).\"\nEmbedding dimension: 1000 (\"The size of the embedded space is cross-validated, and we found that values of approximately 1000 generally work well.\"\nWord combination matrix: 400* 1000=400000\nImage embedding: 4096*1000=4096000 (\"We use the Caffe [41] implementation of the ImageNet Detection RCNN model [27] to detect objects in all images. On our machine with a Tesla K40 GPU, the RCNN processes one image in approximately 25 seconds. We discard the predictions for 200 ImageNet detection classes and only keep the 4096-D activations\")\nCNN: 60,000,000 \"The CNN architecture is identical to the one described in Girhsick et al. [26]. It contains approximately 60 million parameters\"\nTotal parameters: 4096000+80000000+400000+60000000=144,496,000"},{"id":"epoch:fe7908d69393d9977497:training-compute","entityId":"epoch:fe7908d69393d9977497","metric":"training-compute","value":3.411072e+18,"unit":"FLOP","observationPeriod":"2014-06-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"All networks in this paper can be\ntrained on a single GeForce GTX Titan GPU (6 GB memory) within two to four weeks.\"\n4.7e12 FLOP/s * 4* 7*24*60*60 seconds * 0.3 utilisation"},{"id":"epoch:dec08c8d37856be390cd:training-compute","entityId":"epoch:dec08c8d37856be390cd","metric":"training-compute","value":5.184e+17,"unit":"FLOP","observationPeriod":"2014-06-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From https://openai.com/blog/ai-and-compute/ Appendix\n\n\"Less than 0.006 pfs-days\"\n(86400*10^15*0.006)\n\nSeems extremely speculative, unless someone at OpenAI privately corresponded with the authors. There is no information about compute or training in the GANs paper."},{"id":"epoch:a593f9849bd1bd40e667:parameters","entityId":"epoch:a593f9849bd1bd40e667","metric":"parameters","value":47795232.0,"unit":"parameters","observationPeriod":"2014-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The best architecture that we found uses three convolutional layers each followed by a max-pooling layer. The convolutional layers have 96, 128 and 256 filters respectively. Each convolutional layer has a 5 \u00d7 5 receptive field applied with a stride of 1 pixel. Each max pooling layer pools 3 \u00d7 3 regions at strides of 2 pixels. The convolutional layers are followed by two fully connected hidden layers having 2048 units each.\"\nInputs: 32 x 32 x 3\nconv_1: 96 x 5 x 5 x 3 = 7,200\nconv_2: 128 x 5 x 5 x 96 = 307,200\nconv_3: 256 x 5 x 5 x 128 = 819,200\n(output shape after CNN 3 will be: 20 x 20 x 256)\n\nmax_pool has no learnable parameters but further reduces output shape to 9 x 9 x 256\n\nFFN_1: (9 x 9 x 256) x 2048 = 42,467,328\nFFN_2: 2048 x 2048 = 4,194,304\n\nSo total number of parameters appear to be 47,795,232"},{"id":"epoch:19c449146516e6992a58:parameters","entityId":"epoch:19c449146516e6992a58","metric":"parameters","value":13040.0,"unit":"parameters","observationPeriod":"2014-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"D=25 \"For recursive neural models, the dimension of word vector is set to 25, and f = tanh is used as the nonlinearity function. We employ 10 composition matrices in AdaRNN.\"\nComposition matrices: \"W \u2208 R D\u00d72D is the composition matrix, and b is the bias vector.\"\nC=10 \"We employ 10 composition matrices in AdaRNN.\"\nCombination matrix: \"S \u2208 R C\u00d7(2D+|e|) is the matrix used to determine which composition function we use, vl , vr are the left and right child vectors, and e are external feature vector. In this work, e is a one-hot binary feature vector which indicates what the dependency type is.\"\n|e| > 4: (see Figure 2)\nWeights: 10 * 25 * 50 + 10 * (50+4) =13040 (ignoring embedding to the 25 dimension embedding space)"},{"id":"epoch:08af608dfb9e451ca1a7:parameters","entityId":"epoch:08af608dfb9e451ca1a7","metric":"parameters","value":32000000.0,"unit":"parameters","observationPeriod":"2014-05-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" 75000*400+5000*400=32000000\n\"We learn the word vectors and paragraph vectors using 75,000 training documents\"\n\"In PV-DM, the learned vector representations have 400 dimensions for both words and documents\"\nParagraph embedding of dimension number of paragraphs * embedding size\nWord embedding of dimension |V|*embedding size\nAssuming vocabulary of 5000 since results are compared directly to Maas et. al., 2011"},{"id":"epoch:b49e6c858b8b9bf8da07:parameters","entityId":"epoch:b49e6c858b8b9bf8da07","metric":"parameters","value":239712.0,"unit":"parameters","observationPeriod":"2014-03-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The ANN consists of three layers (Fig. 3): a substrate layer inwhich information from the game screen (raw pixels, objects, ornoise) is given as input to the network; a processing layer whichadds a nonlinear internal representation; and a nonlinear outputlayer from which actions are read and conveyed to the Atari em-ulator. Both the input and output layers are fully connected tothe processing layer. The substrate dimensionality of the inputand processinglayers is 810 in the case of the object repre-sentation and 1621 for the pixel and noise representations.3The output layer consists of a 33 substrate mirroring the ninepossible directions of the Atari 2600 joystick and a single noderepresenting thefire button\""},{"id":"epoch:43c9b860b18ef1c33d5f:parameters","entityId":"epoch:43c9b860b18ef1c33d5f","metric":"parameters","value":10000.0,"unit":"parameters","observationPeriod":"2014-02-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:c0a2f004a4a8b46951fb:parameters","entityId":"epoch:c0a2f004a4a8b46951fb","metric":"parameters","value":1021535.0,"unit":"parameters","observationPeriod":"2014-02-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming 1000 input features, 35 classes and 3 hidden layers of 500 units each"},{"id":"epoch:10d3de97eae565375b83:parameters","entityId":"epoch:10d3de97eae565375b83","metric":"parameters","value":120000000.0,"unit":"parameters","observationPeriod":"2014-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"400k vocab * 300 vector dimensions"},{"id":"epoch:f9062e1f117d484c86ab:parameters","entityId":"epoch:f9062e1f117d484c86ab","metric":"parameters","value":120000000.0,"unit":"parameters","observationPeriod":"2014-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"400k vocab * 300 vector dimensions"},{"id":"epoch:61e5b87b841b0493492f:parameters","entityId":"epoch:61e5b87b841b0493492f","metric":"parameters","value":144000000.0,"unit":"parameters","observationPeriod":"2013-12-21","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"144M (Table 4)"},{"id":"epoch:ba250ecd1b66281c4775:parameters","entityId":"epoch:ba250ecd1b66281c4775","metric":"parameters","value":784000.0,"unit":"parameters","observationPeriod":"2013-12-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We trained generative models (decoders) and corresponding encoders\n(a.k.a. recognition models) having 500 hidden units in case of MNIST\"\n\n784*500*2=784000 (Ignoring latent dimension)"},{"id":"epoch:ba250ecd1b66281c4775:training-compute","entityId":"epoch:ba250ecd1b66281c4775","metric":"training-compute","value":475200000000000.0,"unit":"FLOP","observationPeriod":"2013-12-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From https://openai.com/blog/ai-and-compute/ Appendix\n\n\"less than 0.0000055 pfs-days\"\n(86400*10^15*0.0000055)\n\nFigure 2 shows evaluations with 10^8 training samples\n\n6*784000*100000000=470400000000000"},{"id":"epoch:14bf22551770cf46b69c:parameters","entityId":"epoch:14bf22551770cf46b69c","metric":"parameters","value":6160000.0,"unit":"parameters","observationPeriod":"2013-12-20","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:7f1656b6a9283fc7e98b:parameters","entityId":"epoch:7f1656b6a9283fc7e98b","metric":"parameters","value":836096.0,"unit":"parameters","observationPeriod":"2013-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The input to the neural network consists is an 84 \u00d7 84 \u00d7 4 image produced by \u03c6. The first hidden layer convolves 16 8 \u00d7 8 filters with stride 4 with the input image and applies a rectifier nonlinearity [10, 18]. The second hidden layer convolves 32 4 \u00d7 4 filters with stride 2, again followed by a rectifier nonlinearity. The final hidden layer is fully-connected and consists of 256 rectifier units. The output layer is a fully connected linear layer with a single output for each valid action. The number of valid actions varied between 4 and 18 on the games we considered.\"\n\nParameter: 4*16*8*8+16*32*4*4+10*10*32*256+18*256=836096"},{"id":"epoch:7f1656b6a9283fc7e98b:training-compute","entityId":"epoch:7f1656b6a9283fc7e98b","metric":"training-compute","value":2846883840000000.0,"unit":"FLOP","observationPeriod":"2013-12-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Network is 84x84x4 input, 16, 8x8, stride 4, 32 4x4 stride 2, 256 fully connected\nFirst layer: 20*20*4*16*8*8 = 1638400\nSecond layer: 9*9*16*32*4*4 = 663552\nThird layer: 9*9*32*256 = 663552\nTotal ~ 2965504\n2965504 * 5M updates * 32 batch size * 2 multiply-add * 3 backward pass\n= 2965504*50000*100*32*6 = 2846883840000000\n\n"},{"id":"epoch:946ad0c474552fe0ac29:parameters","entityId":"epoch:946ad0c474552fe0ac29","metric":"parameters","value":6000000.0,"unit":"parameters","observationPeriod":"2013-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:36a4487d2edbcee2be05:parameters","entityId":"epoch:36a4487d2edbcee2be05","metric":"parameters","value":20000000000.0,"unit":"parameters","observationPeriod":"2013-12-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"20B from Table 1"},{"id":"epoch:2958be60992099d90276:parameters","entityId":"epoch:2958be60992099d90276","metric":"parameters","value":29900000.0,"unit":"parameters","observationPeriod":"2013-12-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The DBLSTM network had five bidirectional hidden levels, with 500 LSTM cells in each of the forward and backward\nlayers, and a size 3385 softmax output layer, giving a total of\n29.9M weights.\""},{"id":"epoch:a44a056467fef2b9d957:parameters","entityId":"epoch:a44a056467fef2b9d957","metric":"parameters","value":942000000.0,"unit":"parameters","observationPeriod":"2013-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Based on the TransE architecture, the authors give a formula for how the model size scales with the dimensionality of the dataset. The model scale is proportional to: k*(n_e+n_r) where k is the embeddings dimension, n_e is the number of entities, and n_r is the number of relationships.\n\nThey studied using the TransE model for two datasets: FB15k and FB1M. The FB15k model has 810000 parameters.\n\nFB15k has 14951 entities and 1345 relationships. FB1M has 1000000 entities and 23382 relationships. Therefore, the FB1M model will be bigger than the FB15k model by a factor of (23382e6)/(14951*1345) => N = 8.1e5 * (23382e6)/(14951*1345) = 942e6."},{"id":"epoch:a44a056467fef2b9d957:training-compute","entityId":"epoch:a44a056467fef2b9d957","metric":"training-compute","value":1.340928e+18,"unit":"FLOP","observationPeriod":"2013-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8 GPUs (they don't specify which, so I used the average for FP32 for 2017 from the write-up table)\n8 hours \n0.33 util rate"},{"id":"epoch:4baf3aedca4a24489855:training-compute","entityId":"epoch:4baf3aedca4a24489855","metric":"training-compute","value":5.32e+17,"unit":"FLOP","observationPeriod":"2013-11-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1 GPU * 12 days * 1.54 TFLOPS/GTX 580 * 0.33 utilization \n= 532 PF = 0.0062 pfs-days\n\nSource: https://openai.com/blog/ai-and-compute\n\n\"We stopped training after 70 epochs,\nwhich took around 12 days on a single GTX580 GPU\""},{"id":"epoch:82bb47558406835c8ec3:parameters","entityId":"epoch:82bb47558406835c8ec3","metric":"parameters","value":69003872.0,"unit":"parameters","observationPeriod":"2013-11-11","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Computed from architecture description in Caffee\n\nhttps://nbviewer.jupyter.org/github/BVLC/caffe/blob/master/examples/detection.ipynb"},{"id":"epoch:2602f44321318f038fee:parameters","entityId":"epoch:2602f44321318f038fee","metric":"parameters","value":207600000.0,"unit":"parameters","observationPeriod":"2013-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We discarded from the vocabulary all words that occurred less than 5 times in the training data, which resulted in a vocabulary of size 692K [...] Starting with the same news data as in the previous experiments, we first constructed the phrase based training corpus and then we trained several Skip-gram models using different hyperparameters. As before, we used vector dimensionality 300 and context size 5.\""},{"id":"epoch:0fae6c7d89c0f9da6b8a:parameters","entityId":"epoch:0fae6c7d89c0f9da6b8a","metric":"parameters","value":692000000.0,"unit":"parameters","observationPeriod":"2013-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"We discarded from the vocabulary all words that occurred less than 5 times in the training data, which resulted in a vocabulary of size 692K"},{"id":"epoch:0fae6c7d89c0f9da6b8a:training-compute","entityId":"epoch:0fae6c7d89c0f9da6b8a","metric":"training-compute","value":3.888e+16,"unit":"FLOP","observationPeriod":"2013-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From https://openai.com/blog/ai-and-compute/ Appendix.\n\n\"less than 0.00045 pfs days\"\n(86400*10^15*0.00045)"},{"id":"epoch:bd45a8049a01db40827c:training-compute","entityId":"epoch:bd45a8049a01db40827c","metric":"training-compute","value":1.422e+16,"unit":"FLOP","observationPeriod":"2013-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The RNTN would usually achieve its best performance on the dev set after training for 3 - 5 hours.\"\n"},{"id":"epoch:495c80b2eafedced26ab:training-compute","entityId":"epoch:495c80b2eafedced26ab","metric":"training-compute","value":9331200000000000.0,"unit":"FLOP","observationPeriod":"2013-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The training of an RCTM takes about 15 hours on 3 multicore CPUs\"\nGiven the publication year, a rough estimate for the CPU performance is 16 FP32 per cycle, 4 cores, clock speed 4GHz, utilization of 0.3.\n15*60*60*3*4*12*4000000000*0.3=9331200000000000=9.33e15"},{"id":"epoch:e0a110107a98740518a8:parameters","entityId":"epoch:e0a110107a98740518a8","metric":"parameters","value":37230.0,"unit":"parameters","observationPeriod":"2013-09-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Sum numbers of weights in Table 1.b"},{"id":"epoch:e0a110107a98740518a8:training-compute","entityId":"epoch:e0a110107a98740518a8","metric":"training-compute","value":1.37e+17,"unit":"FLOP","observationPeriod":"2013-09-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Training each network requires one day of computation with an optimized GPU implementation\"\n\nAssuming 1.58E+12 FLOP/second on FP32 (from the table in the Estimating compute post), we get\n\n3600*24*1.58E+12 = 1.37E+17 FLOP"},{"id":"epoch:271819674708f308b1c8:parameters","entityId":"epoch:271819674708f308b1c8","metric":"parameters","value":54000000.0,"unit":"parameters","observationPeriod":"2013-08-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"the word-level network had 10,000 inputs and outputs and around\n54M weights\""},{"id":"epoch:271819674708f308b1c8:training-compute","entityId":"epoch:271819674708f308b1c8","metric":"training-compute","value":4210000000000000.0,"unit":"FLOP","observationPeriod":"2013-08-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 54000000 parameters * 929000 tokens * 14 epochs = 4.213944e+15 FLOP"},{"id":"epoch:710716e9ef61c94c0013:parameters","entityId":"epoch:710716e9ef61c94c0013","metric":"parameters","value":51609600.0,"unit":"parameters","observationPeriod":"2013-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 1 details architecture, weights are shared between scales\n\u201cThe network is then applied to each 3-dimension input map Xs. This input is transformed into a 16-dimension feature map, using a bank of 16 filters, 10 connected to the Y channel, the 6 others connected to the U and V channels. The second layer transforms this 16-dimension feature map into a 64-dimension feature map, each map being produced by a combination of 8 randomly selected feature maps from the previous layer. Finally the 64-dimension feature map is transformed into a 256-dimension feature map, each map being produced by a combination of 32 randomly selected feature maps from the previous layer.\u201d\nL1: 1*7*7*10+2*7*7*6=1078\nL2: 8*64*7*7=25088\nFC head is split for the three networks (assumption)\nFC: 32*80*60*256+32*40*30*256+32*20*15*256=51609600\n\nThis is a very different estimate to the 0.5M trainable parameters mentioned in the paper(!) \"The convolutional network has roughly 0.5 million trainable parameters\""},{"id":"epoch:710716e9ef61c94c0013:training-compute","entityId":"epoch:710716e9ef61c94c0013","metric":"training-compute","value":2.3774688e+17,"unit":"FLOP","observationPeriod":"2013-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Forward FLOP\nSplit for the three sizes\nF1:\nL1: 2*1*7*7*10*320*240+2*2*7*7*6*320*240=165580800\nL2: 2*8*64*7*7*160*120=963379200\nFC: 2*32*80*60*256=78643200\nF2:\nL1: 2*1*7*7*10*160*120+2*2*7*7*6*160*120=41395200\nL2: 2*8*64*7*7*80*60=240844800\nFC: 2*32*40*30*256=19660800\nF3:\nL1: 2*1*7*7*10*80*60+2*2*7*7*6*80*60=10348800\nL2: 2*8*64*7*7*40*30=60211200\nFC: 2*32*20*15*256=4915200\nTotal: 165580800+963379200+78643200+41395200+240844800+19660800+10348800+60211200+4915200=1584979200\n\nTraining compute: 1584979200*3*50000000=237746880000000000=2.4e17\n"},{"id":"epoch:8b1b492d3999ec3d8fe5:training-compute","entityId":"epoch:8b1b492d3999ec3d8fe5","metric":"training-compute","value":90842400000000.0,"unit":"FLOP","observationPeriod":"2013-06-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"They use a Intel Xeon E5-2470 Processor for 2 hours. This can do 12,617 MOps/Sec (average test results, assuming they achieved a similar utilization)\nhttps://www.cpubenchmark.net/cpu.php?cpu=Intel+Xeon+E5-2470+%40+2.30GHz&id=2003\n\n12617000000*2*60*60=90842400000000"},{"id":"epoch:2cb92b2ec3fcce530bb9:parameters","entityId":"epoch:2cb92b2ec3fcce530bb9","metric":"parameters","value":206899200.0,"unit":"parameters","observationPeriod":"2013-05-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The input for the DNN is eleven contiguous frames of 40-dimensional log-filterbank features. The DNN consists of four hidden layers each with 2560 nodes\"\nNetwork structure: 3 multilingual shared layers, 1 language specific hidden layer + output layer (Figure 2)\nLanguage specific layer output sizes: 1600, 3300, 2900, 5700, 3500, 5500, 6200, 4700, 5100, 4900, 3700 (Table 1)\nShared: 11*40*2560+2560*2560+2560*2560=14233600\nLanguage heads: 11*2560*2560+2560*1600+2560*3300+2560*2900+2560*5700+2560*3500+2560*5500+2560*6200+2560*4700+2560*5100+2560*4900+2560*3700=192665600\nTotal: 14233600+192665600=206899200=2e8"},{"id":"epoch:27379bc7c6dd47f14eba:parameters","entityId":"epoch:27379bc7c6dd47f14eba","metric":"parameters","value":101706240.0,"unit":"parameters","observationPeriod":"2013-05-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The overall input dimensionality is 1040,\"\n\"All layers of our networks have 2560 hidden units \"\n\"used to generate 7969 context-dependent tied acoustic states\"\nLargest model: 12 hidden layers (Fig 4)\nParameters: 1040*2560+12*2560*2560+2560*7969=101706240\n"},{"id":"epoch:27379bc7c6dd47f14eba:training-compute","entityId":"epoch:27379bc7c6dd47f14eba","metric":"training-compute","value":1.2773376e+17,"unit":"FLOP","observationPeriod":"2013-05-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"across 4 machines using up to 4 CPUs each\"\nCPU model not specified, I assumed a Sandy Bridge with 16 FLOP/cycle and 3.3GhZ based on the publication year (4*16*3300000000=211200000000 FLOP/s per machine)\nCompute: 4*211200000000*168*60*60*0.3 = 1.53e17\n\nAlternatively, the training set is \"several hundred hours of speech\", with inputs consisting of 26 frames, each frame is 10ms apart.\nIf we assume 400h of training data, 400h/10ms/26= 5,538,461 inputs\n6 * 101706240 * 5,538,461 = 3.38e15 FLOPs per epoch. Number of epochs unstated."},{"id":"epoch:262d094670f102b0038a:parameters","entityId":"epoch:262d094670f102b0038a","metric":"parameters","value":43000000.0,"unit":"parameters","observationPeriod":"2013-03-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:57373bc5a8b64bd072c1:training-compute","entityId":"epoch:57373bc5a8b64bd072c1","metric":"training-compute","value":2.612736e+18,"unit":"FLOP","observationPeriod":"2013-01-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Trained for 14 days on 180 CPU cores (Table 6)\nRoughly estimating the performance of CPUs in a HPC around 2013: 16 FP32 operations per cycle, 2.5GHz, 0.3 utilization\nTime: 14*24*60*60=1209600s\nFLOPs: 0.3*180*16*2500000000=2160000000000\nTraining compute: 1209600s * 2160000000000 = 2612736000000000000 = 2.61e18\nhttps://www.wolframalpha.com/input?i=16+FLOP+*+2.5+GHz+*+180+*+14+days+*+0.3"},{"id":"epoch:13f41055635fb7a76cef:parameters","entityId":"epoch:13f41055635fb7a76cef","metric":"parameters","value":195600.0,"unit":"parameters","observationPeriod":"2012-12-14","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"It uses 400 hidden units (selected via hyperparameter tuning)\nThe input size is 88 (corresponding to the 88 piano pitches)\nIt uses rectified linear units, so no activation function parameters\nSo the number of parameters would be:\nInput to hidden weights: 88 * 400 = 35,200\nHidden to hidden weights: 400 * 400 = 160,000\nBiases: 400\nTotal: ~195,600 parameters\nAbove estimate is by Claude 2. Should be checked manually."},{"id":"epoch:cb56b57edd4d51aeca34:parameters","entityId":"epoch:cb56b57edd4d51aeca34","metric":"parameters","value":2049999.9999999998,"unit":"parameters","observationPeriod":"2012-12-04","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:395611ef167d35db10cf:parameters","entityId":"epoch:395611ef167d35db10cf","metric":"parameters","value":47185920.0,"unit":"parameters","observationPeriod":"2012-12-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We used a deep network with five layers: four hidden layer with sigmoidal activations and 2560 nodes each, and a softmax output layer with 8192 nodes.\"\n\"The network was fully-connected layer-to-layer, for a total of approximately 42 million model parameters.\"\n2560*2560*4+2560*8192=47185920"},{"id":"epoch:395611ef167d35db10cf:training-compute","entityId":"epoch:395611ef167d35db10cf","metric":"training-compute","value":3.114e+17,"unit":"FLOP","observationPeriod":"2012-12-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.wolframalpha.com/input?i=6+FLOP+*+47185920+*+1.1+billion\nNumber of epochs unknown but most likely 1 and probably under 30.\nWe could narrow down the uncertainty further if we knew something about the hardware."},{"id":"epoch:55b6d46baf541acea0eb:parameters","entityId":"epoch:55b6d46baf541acea0eb","metric":"parameters","value":1700000000.0,"unit":"parameters","observationPeriod":"2012-12-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"we used Downpour SGD to train the 1.7 billion parameter image model\""},{"id":"epoch:cdde06923bf21b9cff67:parameters","entityId":"epoch:cdde06923bf21b9cff67","metric":"parameters","value":218896.0,"unit":"parameters","observationPeriod":"2012-12-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1 shows architecture\n1st layer CNN: 48*4*4=768\n2nd layer CNN: 48*48*5*5=57600\n3rd layer CNN: 48*48*4*4=36864\n4th layer CNN: 48*48*4*4=36864\n1st layer FC: 48*3*3*200=86400\n2nd layer FC: 200*2=400\nTotal parameters: 768+57600+36864+36864+86400+400=218896"},{"id":"epoch:cdde06923bf21b9cff67:training-compute","entityId":"epoch:cdde06923bf21b9cff67","metric":"training-compute","value":4.78e+17,"unit":"FLOP","observationPeriod":"2012-12-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"This amounts to 3 million training examples in total, in which both classes are equally represented. [...] We take advantage of this property, and synthetically augment the training set at the beginning of each epoch by randomly mirroring each training instance, and/or rotating it by \u00b190\u25e6\"\nAssuming 3 x training examples per epoch due to data augmentation\n\nCounting estimate:\n1st layer CNN: 2*92*92*48*4*4=13000704\n2nd layer CNN: 2*42*42*48*48*5*5=203212800\n3rd layer CNN: 2*18*18*48*48*4*4=23887872\n4th layer CNN: 2*6*6*4*4*48*48=2654208\n1st layer FC: 2*48*3*3*200=172800\n2nd layer FC: 2*200*2=800\nTotal forward flop: 13000704+203212800+23887872+2654208+172800+800=242929184\nTraining compute: 242929184*3*9000000*30=196772639040000000=1.9e17\n\nGPU hour estimate\n340 minutes per epoch, 30 epochs\n4 GTX 580 with 1580000000000 FLOPs\nAssumed utilization of 0.3\n30*340*60*1580000000000*4*0.3=1160352000000000000=1.16e18\n\nGeometric mean: sqrt(196772639040000000*1160352000000000000)=4.78e17\n"},{"id":"epoch:af89bb791ca2b7b5d038:parameters","entityId":"epoch:af89bb791ca2b7b5d038","metric":"parameters","value":9000000.0,"unit":"parameters","observationPeriod":"2012-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:3e49e54f8ed72a0a7d65:parameters","entityId":"epoch:3e49e54f8ed72a0a7d65","metric":"parameters","value":60000000.0,"unit":"parameters","observationPeriod":"2012-09-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Our neural network architecture has 60 million parameters.\""},{"id":"epoch:3e49e54f8ed72a0a7d65:training-compute","entityId":"epoch:3e49e54f8ed72a0a7d65","metric":"training-compute","value":4.7e+17,"unit":"FLOP","observationPeriod":"2012-09-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.2M images * 90 epochs * 0.75 GFLOP * (2 add-multiply) * (3 backward pass) \n= 470 PF = 0.0054 pfs-days\n\nSource: https://openai.com/blog/ai-and-compute/\n\nHardware method:\n2 GTX 580 3GB GPUs for \"between five and six days\". Assuming 5.5 days and 32-bit training:\n1.581 TFLOPS * 5.5 days * 2 = 1.5e18 FLOP\nComparing to the operation counting method, this implies around 31% MFU.\n\nAuthors of \"AI and Memory Wall\" (https://github.com/amirgholami/ai_and_memory_wall) estimated model's training compute as 460 PFLOP = 4.6*10^17 FLOP"},{"id":"epoch:72b3d8c9f7bec57f1863:parameters","entityId":"epoch:72b3d8c9f7bec57f1863","metric":"parameters","value":102720000.0,"unit":"parameters","observationPeriod":"2012-09-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Multiple models were trained, the largest on transcribed French podcast data.\n\"We trained an LSTM LM using 300 hidden nodes and 27 M running words of indomain training data\"\n\"Corpus sizes in number of running words; the vocabulary size of the Treebank corpus is 10 K, for Quaero French it is 170 K\"\nEmbedding and unembedding: 2*170000*300=102000000\nLSTM: 4*600*300=720000\nTotal: 102000000+720000=102720000=1.03e8\n(Assuming the embedding dimension is the same as the LSTM layer)"},{"id":"epoch:72b3d8c9f7bec57f1863:training-compute","entityId":"epoch:72b3d8c9f7bec57f1863","metric":"training-compute","value":1.66e+16,"unit":"FLOP","observationPeriod":"2012-09-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"FLOP per input for LSTM layer is 4*2*(M+N)*M, for N inputs and M outputs.\n\nEmbedding FLOPs: 2 * 170000 * 300 = 102,000,000\nLSTM FLOPs: 4 * 2 * (300 + 300) * 300 = 1,440,000\nUnembedding FLOPs: 2 * 170000 * 300 = 102,000,000\nTotal: 205,440,000 FLOPs per word per forward pass\nFor 27M training input words and including backward passes: 27M * 3 * 205,440,000 = 1.66e16\n\nHowever, it sounds like they're doing something with a secondary acoustic model, so this may be an underestimate."},{"id":"epoch:7fadecbb69d1d0ab3df2:parameters","entityId":"epoch:7fadecbb69d1d0ab3df2","metric":"parameters","value":12000000.0,"unit":"parameters","observationPeriod":"2012-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:fa59fadcaecbe3f5d86d:parameters","entityId":"epoch:fa59fadcaecbe3f5d86d","metric":"parameters","value":3510255.0,"unit":"parameters","observationPeriod":"2012-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We represent a word as both a continuous vector and a matrix of parameters. We initialize all word vectors x \u2208 Rn with pre-trained 50-dimensional word vectors from the unsupervised model of Collobert and Weston (2008). [...] Every word is also associated with a matrix X.  [...] If the vectors have dimensionality n, then each word\u2019s matrix has dimensionality X \u2208 Rn\u00d7n.\"\n\n\"We propose the following combination function which is input dependent:\np = fA,B(a, b) = f(Ba, Ab) = g(W x (Ba Ab)) ,(2)\nwhere A, B are matrices for single words, the global W \u2208 Rn\u00d72n is a matrix that maps both transformed words back into the same n-dimensional space.\"\n\n\"For computing nonterminal phrase matrices, we define the function\nP = fM(A, B) = WM\u0014A, B\u0015, (3)\nwhere WM \u2208 Rn\u00d72n, so P \u2208 Rn\u00d7n just like each input matrix.\"\n\n\"If every word is represented by an n-dimensional vector and additionally by an n \u00d7 n matrix, the dimensionality of the whole model may become too large with commonly used vector sizes of n = 100. In order to reduce the number of parameters, we represent word matrices by the following low-rank plus diagonal approximation: A = UV + diag(a), (5)\nwhere U \u2208 Rn\u00d7r, V \u2208 Rr\u00d7n, a \u2208 Rnand we set the rank for all experiments to r = 3.\"\n\n\"We train these representations by adding on top of each parent node a simple softmax classifier\nto predict a class distribution over, e.g., sentiment or relationship classes: d(p) = softmax(Wlabelp). If there are K labels, then d \u2208 RK is a K-dimensional multinomial distribution\"\n\nIn total there are V*(n+n*r + r*n) + n*2n + n*2n + (n+1)*k parameters, where n is the vector dimension, r is the low-rank decomposition dimension, V is the vocabulary size and k is the number of classes.\n\nIn the experiments we have that n=50, r=3, k=? and V=?. I'm guesstimating k=5 and V=10k."},{"id":"epoch:05fce1017812ee16dc8c:parameters","entityId":"epoch:05fce1017812ee16dc8c","metric":"parameters","value":1000000000.0,"unit":"parameters","observationPeriod":"2012-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"To answer this, we train a 9-layered locally connected sparse autoencoder with pooling and local contrast normalization on a large dataset of images (the model has 1 billion connections, the dataset has 10 million 200x200 pixel images downloaded from the Internet)\""},{"id":"epoch:05fce1017812ee16dc8c:training-compute","entityId":"epoch:05fce1017812ee16dc8c","metric":"training-compute","value":6e+17,"unit":"FLOP","observationPeriod":"2012-07-12","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming 1 epoch, 10 million images and 1 billion parameters, 6*N*D = 6*10^17 FLOP"},{"id":"epoch:d0875820c4200140e6fa:parameters","entityId":"epoch:d0875820c4200140e6fa","metric":"parameters","value":2000000.0,"unit":"parameters","observationPeriod":"2012-06-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Vocabulary size (V) = 10,000 words\nContext size (c) = 2 words\nFeature dimension (d) = 100\nPlugging this into the parameter count:\n\nContext word features: 10,000 x 100 = 1,000,000\nTarget word features: 10,000 x 100 = 1,000,000\nContext weights: 2 x 100 x 100 = 20,000\nTotal parameters: 1,000,000 + 1,000,000 + 20,000 = 2,020,000\nSo for the LBL model configuration used in the experiments in this paper, with a 10K word vocabulary, 2-word context, and 100D feature vectors, the total number of parameters is approximately 2 million."},{"id":"epoch:d0875820c4200140e6fa:training-compute","entityId":"epoch:d0875820c4200140e6fa","metric":"training-compute","value":501999999999999.94,"unit":"FLOP","observationPeriod":"2012-06-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6 FLOP / parameter / token * 2000000 parameters ['Likely' confidence] * 929000 tokens * 45 epochs [\"Likely\" confidence] = 5.0166e+14 FLOP"},{"id":"epoch:75d041cff09f079bd928:training-compute","entityId":"epoch:75d041cff09f079bd928","metric":"training-compute","value":2.731968e+17,"unit":"FLOP","observationPeriod":"2012-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"a single NVIDIA GTX 580 GPU... Training on ImageNet takes\nroughly four days with dropout and two days without.\"\n1.581 TFLOP/s * 4 day * 86400 s/day * 0.5 utilization"},{"id":"epoch:6fae4e450a83c461396e:training-compute","entityId":"epoch:6fae4e450a83c461396e","metric":"training-compute","value":4268700000000000.0,"unit":"FLOP","observationPeriod":"2012-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"a single NVIDIA GTX 580 GPU. Training on CIFAR-10 takes roughly 90 minutes\" p17\n1.581 TFLOP/s * 90 min * 60 s/min * 0.5 utilization"},{"id":"epoch:3c433c67bebfaa24b1e4:parameters","entityId":"epoch:3c433c67bebfaa24b1e4","metric":"parameters","value":5594010.0,"unit":"parameters","observationPeriod":"2012-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"We show results for 4 nets (784-800-800-10, 784-1200-1200-10, 784-2000-2000-10, 784-1200-1200-1200-10)\n\n784*2000+2000*2000+10*2000+6010=5594010"},{"id":"epoch:3c433c67bebfaa24b1e4:training-compute","entityId":"epoch:3c433c67bebfaa24b1e4","metric":"training-compute","value":6039370800000000.0,"unit":"FLOP","observationPeriod":"2012-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Num mul-add / forward pass\n2 FLOPs / mult-add\n3 total mult-add / fp mult-add\n3000 epochs\n60000 training samples"},{"id":"epoch:36e6935d60b6ec96ffd2:parameters","entityId":"epoch:36e6935d60b6ec96ffd2","metric":"parameters","value":48840185.0,"unit":"parameters","observationPeriod":"2012-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The input to the net is 21 adjacent frames with an advance of 10ms per frame. The neural net has 4 fully-connected hidden layers of 4000 units per layer and 185 \u201csoftmax\u201d output units that are subsequently merged into the 39 distinct classes used for the benchmark."},{"id":"epoch:830d937c53b74c54d267:parameters","entityId":"epoch:830d937c53b74c54d267","metric":"parameters","value":2653700.0,"unit":"parameters","observationPeriod":"2012-02-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We train five DNN columns per normalization, resulting in a total of 35 columns for the entire MCDNN.\n[Each DNN has an architecture] 1x29x29-20C4-MP2-40C5-MP3-150N-10N DNN\"\n\nParameter calculation: 20*4*4+20*40*5*5+150*40*3*3+10*150=75820\n35*75820=2653700"},{"id":"epoch:830d937c53b74c54d267:training-compute","entityId":"epoch:830d937c53b74c54d267","metric":"training-compute","value":1.57e+16,"unit":"FLOP","observationPeriod":"2012-02-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Num of multiply-adds per forward pass\n2 FLOPs/mult-add\n3 (fp+bp FLOPs / fp FLOPs)\n800 epochs\n60.000 training size\n35 networks\n\n\"Training a DNN takes almost 14 hours and after 500 training epochs little additional improvement is observed\"\n\n500*60000*3*174147400=1.57e+16"},{"id":"epoch:441f4f3cbc90b61f3487:parameters","entityId":"epoch:441f4f3cbc90b61f3487","metric":"parameters","value":5000000.0,"unit":"parameters","observationPeriod":"2011-11-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The capacity of our network architectures lies mainly in the word lookup table, which contains 50 \u00d7 100,000 parameters to train. [...] most of the trainable parameters are located in the lookup tables.\""},{"id":"epoch:302fa58f1c2d3b4c0c3e:parameters","entityId":"epoch:302fa58f1c2d3b4c0c3e","metric":"parameters","value":15260.0,"unit":"parameters","observationPeriod":"2011-11-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Did not take into account initial image feature extraction, only novel stuff.\n\n1. Perform PCA on the feature matrices from both domains. Learnable parameters are projection matrices.\n= 800 (# features) x 200 (reduced dimension) x 2 (once per subdomain)\n\n2. Perform partial least squares regression. Learnable parameters are\n\nMatrix P with dimensions 200 (# features) x 30 (dimension of latent space)\nMatrix Q with dimensions 1 (# responses) x 30 (dimension of latent space)\nProjection matrix of X onto latent space:  200 (# features) x 30 (dimension of latent space)\nProjection matrix of Y onto latent space:  1 (# responses) x 30 (dimension of latent space)\n"},{"id":"epoch:951bc3b026abce7a5811:parameters","entityId":"epoch:951bc3b026abce7a5811","metric":"parameters","value":1388800.0,"unit":"parameters","observationPeriod":"2011-10-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Parameters\nL1: 3*3*3*100=2700\nL3: 100*4*4*150=240000\nL5: 150*3*3*250=337500\nL7: 250*4*4*200=800000\nL8: 200*43=8600\nTotal: 2700+240000+337500+800000+8600=1388800\n"},{"id":"epoch:951bc3b026abce7a5811:training-compute","entityId":"epoch:951bc3b026abce7a5811","metric":"training-compute","value":991981425600000.0,"unit":"FLOP","observationPeriod":"2011-10-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training FLOP\nL1: 3*3*3*46*46*100=5713200\nL3: 100*4*4*20*20*150=96000000\nL5: 150*3*3*8*8*250=21600000\nL7: 250*4*4*200=800000\nL8: 200*43=8600\nTotal: 2*(5713200+96000000+21600000+800000+8600)=248243600\nTraining Compute: 248243600*3*26640*50=991981425600000=9.9e14"},{"id":"epoch:9fbececd5f7f90dfe6b9:parameters","entityId":"epoch:9fbececd5f7f90dfe6b9","metric":"parameters","value":128420.0,"unit":"parameters","observationPeriod":"2011-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1st CNN layer: 20*4*4=320\n2nd CNN layer: 20*40*9*9=64800\n1st FC layer: 40*3*3*150=54000\n2nd FC layer: 150*62=9300\nTotal : 320+64800+54000+9300=128420\n"},{"id":"epoch:9fbececd5f7f90dfe6b9:training-compute","entityId":"epoch:9fbececd5f7f90dfe6b9","metric":"training-compute","value":2.6e+16,"unit":"FLOP","observationPeriod":"2011-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"GPU hour estimate:\n30 epochs, 6 hours per network, 7 networks total\n7*6*60*60*(2*1580000000000+2*1345000000000)*0.3=2.65e17\n\nCounting estimate:\n800000 training examples, 30 epochs\n1st CNN layer: 2*20*26*26*4*4=432640\n2nd CNN layer: 2*20*6*6*9*9*40=4665600\n1st FC layer: 2*40*3*3*150=108000\n2nd FC layer: 2*150*62=18600\nTotal: 432640+4665600+108000+18600=5224840\n\n5224840*3*800000*30*7=2633319360000000=2.6e15\n\nGeometric mean: sqrt(2.65e17*2.6e15)=2.6e16\n"},{"id":"epoch:9d8c029b104589600617:parameters","entityId":"epoch:9d8c029b104589600617","metric":"parameters","value":120620.0,"unit":"parameters","observationPeriod":"2011-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1st CNN layer: 20*4*4=320\n2nd CNN layer: 20*40*9*9=64800\n1st FC layer: 40*3*3*150=54000\n2nd FC layer: 150*10=1500\nTotal: 320+64800+54000+1500=120620"},{"id":"epoch:9d8c029b104589600617:training-compute","entityId":"epoch:9d8c029b104589600617","metric":"training-compute","value":5.2e+16,"unit":"FLOP","observationPeriod":"2011-09-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"GPU hour estimate:\n800 epochs, 14 hours per network, 7 networks total\n7*14*60*60*(2*1580000000000+2*1345000000000)*0.3=6.19e17\n\nCounting estimate:\n50000 examples, 800 epochs\n1st CNN layer: 2*20*26*26*4*4=432640\n2nd CNN layer: 2*20*6*6*9*9*40=4665600\n1st FC layer: 2*40*3*3*150=108000\n2nd FC layer: 2*150*10=3000\nTotal: 432640+4665600+108000+3000=5209240\n5209240*3*50000*800*7=4375761600000000=4.4e15\n\nGeometric mean: sqrt(4.4e15*6.19e17)=5.2e16"},{"id":"epoch:d66d97af5346024de338:parameters","entityId":"epoch:d66d97af5346024de338","metric":"parameters","value":13125.0,"unit":"parameters","observationPeriod":"2011-08-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"It's a bayes net, parameters are probabilty tables for probability that X happens in direction i given that we go in direction i. There are 25 directions."},{"id":"epoch:f776b7ae1072c29e4593:parameters","entityId":"epoch:f776b7ae1072c29e4593","metric":"parameters","value":4878300.0,"unit":"parameters","observationPeriod":"2011-07-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"L1: 6*300*6*6=64800\nL3: 500*600*4*4=4800000\nL5: (108/4)*500=13500\nTotal: 64800+4800000+13500=4878300"},{"id":"epoch:f776b7ae1072c29e4593:training-compute","entityId":"epoch:f776b7ae1072c29e4593","metric":"training-compute","value":2.57985e+16,"unit":"FLOP","observationPeriod":"2011-07-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"7 epochs, 35 min per epoch\n2 GTX 480 + 2 GTX 580\n580 FLOP: 1580000000000\n480 FLOP: 1345000000000\nFLOP: 7*35*60*(2*1580000000000+2*1345000000000)*0.3=25798500000000000=2.6e16\n"},{"id":"epoch:79cb5c69783f02aa3e2a:parameters","entityId":"epoch:79cb5c69783f02aa3e2a","metric":"parameters","value":255000.0,"unit":"parameters","observationPeriod":"2011-06-19","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We build a fixed dictionary of the 5,000 most frequent tokens\"\n\"For all word vector models, we use 50-dimensional vectors\"\nParameters: 5000*50 + 5000=255000\n"},{"id":"epoch:505d6f10d591a5e755af:parameters","entityId":"epoch:505d6f10d591a5e755af","metric":"parameters","value":139808256.0,"unit":"parameters","observationPeriod":"2011-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2*(3072*8192+8192*4096+4096*2048+2048*1024+1024*512+512*256+256*128+128*64+64*28)=139808256\n\"n each autoencoder, the hidden layers halve in size until they reach the desired size, except that we use 28 instead of 32\""},{"id":"epoch:505d6f10d591a5e755af:training-compute","entityId":"epoch:505d6f10d591a5e755af","metric":"training-compute","value":3.672864e+16,"unit":"FLOP","observationPeriod":"2011-04-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"48*60*60*708500000000*0.3=36728640000000000=3.7e16\nGTX 285 with 708.5 GFLOP/s"},{"id":"epoch:33f4fb1d5b1bf0f678b9:parameters","entityId":"epoch:33f4fb1d5b1bf0f678b9","metric":"parameters","value":22160000.0,"unit":"parameters","observationPeriod":"2010-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"RNN 400/10 indicates that the model has 400 units in the hidden layer and merges words that occur less than 10 times in the training set.\n\nThe model is trained on a 6.4M word subset of the NYT subset of English Gigaword. In this colab notebook, I estimate the vocabulary after merging infrequent words using english word frequency data at around 27.5k: https://colab.research.google.com/drive/1K5qH0EqXtFwTLESNtp4oelCM28GpGXt6?usp=sharing\n\nThen RNN 400/10 has:\n(27.5k + 400) * 400 + (400 * 27.5k) = 22.16M parameters\n\nThe KN5 LM is effectively a lookup table over 5-grams in it's training data. In one sense, this means it has approximately as many parameters as unique 5-grams in the 37M words of training data. However, these aren't parameters in the same sense as neural networks, as they are accessed as a lookup table, so on each forward pass only one parameter is active. I choose not to include these parameters in our count."},{"id":"epoch:33f4fb1d5b1bf0f678b9:training-compute","entityId":"epoch:33f4fb1d5b1bf0f678b9","metric":"training-compute","value":1.7e+16,"unit":"FLOP","observationPeriod":"2010-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Convergence is usually achieved after 10-20 epochs.\"\n\n6 * 22,160,000 * 20 * 6.4M = 1.70e16"},{"id":"epoch:cdb5b14da7c422e7457a:parameters","entityId":"epoch:cdb5b14da7c422e7457a","metric":"parameters","value":19250000.0,"unit":"parameters","observationPeriod":"2010-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The acoustic HMMs are based on cross-word tied-states triphones trained discriminatively using MPE criteria. Feature extraction use 13 Mel-PLP\u2019s features with deltas, double and triple deltas reduced by HLDA to 39-dimension feature vector\"\n\nRNN 500/10 indicates that the model has a 500 units in the hidden layer and merges words that occur less than 10 times in the training set.\n\nSection 4.2 in this paper appears to indicate that NIST RT05 has a vocabulary of 50k words: https://www.fit.vut.cz/person/imikolov/public/rnnlm/char.pdf\n\nUsing this dataset of english word frequencies ...\nhttps://www.kaggle.com/datasets/rtatman/english-word-frequency\n... we can estimate that the number of words in the final vocabulary would have been around 19k when merging with threshold 5:\nhttps://colab.research.google.com/drive/1K5qH0EqXtFwTLESNtp4oelCM28GpGXt6?usp=sharing\n\nThus the model in question would have:\n\n(500 + 19k) * 500 + 500 * 19k = 19250000 parameters\n\nThe RT09 LM they interpolate with appears to be a 4-gram model. In one sense, n-gram models have parameters roughly equal to the number of unique n-grams in the training data. For comparison, the Google Web 1T 5-gram corpus has 1.3B unique 4-grams: https://catalog.ldc.upenn.edu/LDC2006T13 \nHowever, these aren't parameters in the same sense as neural networks as they are accessed as a lookup table, so on each forward pass only one parameter is active. I choose not to include these parameters in our count."},{"id":"epoch:cdb5b14da7c422e7457a:training-compute","entityId":"epoch:cdb5b14da7c422e7457a","metric":"training-compute","value":1.25e+16,"unit":"FLOP","observationPeriod":"2010-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Convergence is usually achieved after 10-20 epochs.\"\n\nAssuming a backward-forward ratio of 2:1, since this is a shallow network\n\n6 * 19250000 * 5.4M * 20 = 1.25e16"},{"id":"epoch:01a143b2abfe135c9962:parameters","entityId":"epoch:01a143b2abfe135c9962","metric":"parameters","value":70265000.0,"unit":"parameters","observationPeriod":"2010-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"This database entry refers to the 3xRNN rows in Table 2 (static and dynamic likely use the same model ensemble, but allow the model weights to update once when testing the dynamic version).\n\nI assume the 3xRNN represents interpolation between the three largest models shown explicitly (RNN 250/5, RNN 250/2, and RNN 400/10). This seems likely, since smaller models do considerably worse on their own.\n\nIn the following colab notebook, I estimate vocabulary sizes for the NYT Gigaword data at around 54.4k, 41.4k, and 27.6k for merge thresholds of 2, 5, and 10, respectively: https://colab.research.google.com/drive/1K5qH0EqXtFwTLESNtp4oelCM28GpGXt6#scrollTo=tedUkbgklNJ3\n\nSo the total number of parameters in each constituent model is:\n- RNN 250/2: (250 + 54.4k) * 250 + (250 * 54.4k) = 27,262,500\n- RNN 250/5: (250 + 41.4k) * 250 + (250 * 41.4k) = 20,762,500\n- RNN 400/10: (400 + 27.6k) * 400 + (400 * 27.6k) = 22,240,000\n\nIn total: 70,265,000 parameters"},{"id":"epoch:01a143b2abfe135c9962:training-compute","entityId":"epoch:01a143b2abfe135c9962","metric":"training-compute","value":5.396e+16,"unit":"FLOP","observationPeriod":"2010-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Convergence is usually achieved after 10-20 epochs.\"\nTraining was done over a 6.4M subset of the NYT section of English Gigaword.\n\n6 * 70,265,000 * 20 * 6.4M = 5.396e16"},{"id":"epoch:4115fb5b28203b6087b6:parameters","entityId":"epoch:4115fb5b28203b6087b6","metric":"parameters","value":77039000.0,"unit":"parameters","observationPeriod":"2010-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The acoustic HMMs are based on cross-word tied-states triphones trained discriminatively using MPE criteria. Feature extraction use 13 Mel-PLP\u2019s features with deltas, double and triple deltas reduced by HLDA to 39-dimension feature vector\"\n\nRNN 1000/5 indicates that the model has a 1000 units in the hidden layer and merges words that occur less than 5 times in the training set.\n\nSection 4.2 in this paper appears to indicate that NIST RT05 has a vocabulary of 50k words: https://www.fit.vut.cz/person/imikolov/public/rnnlm/char.pdf\n\nUsing this dataset of english word frequencies ...\nhttps://www.kaggle.com/datasets/rtatman/english-word-frequency\n... we can estimate that the number of words in the final vocabulary would have been around 38k when merging with threshold 5:\nhttps://colab.research.google.com/drive/1K5qH0EqXtFwTLESNtp4oelCM28GpGXt6?usp=sharing\n\nThus the model in question would have:\n\n(39 + 1000 + 38k) * 1000 + 1000 * 38k = 77,039,000 parameters\n\nThe RT09 LM they interpolate with appears to be a 4-gram model. In one sense, n-gram models have parameters roughly equal to the number of unique n-grams in the training data. However, these aren't parameters in the same sense as neural networks as they are accessed as a lookup table, so on each forward pass only one parameter is active. I choose not to include these parameters in our count."},{"id":"epoch:4115fb5b28203b6087b6:training-compute","entityId":"epoch:4115fb5b28203b6087b6","metric":"training-compute","value":5e+16,"unit":"FLOP","observationPeriod":"2010-09-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Convergence is usually achieved after 10-20 epochs.\"\n\nAssuming a backward-forward ratio of 2:1, since this is a shallow network\n\n6 * 77039000 * 5.4M * 20 = 5.00e16"},{"id":"epoch:b49596778c74b372b7bd:parameters","entityId":"epoch:b49596778c74b372b7bd","metric":"parameters","value":294912.0,"unit":"parameters","observationPeriod":"2010-09-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"C1: 3*16*16*16=12288\nC3: 16*128*6*6=73728\nF5: 128*4*4*102=208896\nTotal: 12288+73728+208896=294912\n"},{"id":"epoch:b49596778c74b372b7bd:training-compute","entityId":"epoch:b49596778c74b372b7bd","metric":"training-compute","value":1221124128768000.0,"unit":"FLOP","observationPeriod":"2010-09-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Forward FLOP\nC1: 2*3*16*16*16*125*125=384000000\nC3: 2*16*128*6*6*20*20=58982400\nF5: 2*128*4*4*102=417792\nTotal: 384000000+58982400+417792=443400192\n\n300 epochs\n30*102=3060 training examples\n\n443400192*3*3060*300=1221124128768000=1.2e15\n"},{"id":"epoch:b9f7f445c0f7f29a85e6:parameters","entityId":"epoch:b9f7f445c0f7f29a85e6","metric":"parameters","value":268664.0,"unit":"parameters","observationPeriod":"2010-09-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 1\nC1: 2*16*5*5=800\nC3: 16*64*6*6=36864\nF5: 64*6*6*100=230400\nF6: 100*6=600\nTotal: 800+36864+230400+600=268664"},{"id":"epoch:b9f7f445c0f7f29a85e6:training-compute","entityId":"epoch:b9f7f445c0f7f29a85e6","metric":"training-compute","value":1456277227200000.0,"unit":"FLOP","observationPeriod":"2010-09-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Forward FLOP\nC1: 2*2*16*5*5*92*92=13542400\nC3: 2*16*16*6*6*18*18=5971968\nF5: 2*64*6*6*100=460800\nF6: 2*100*6=1200\nTotal: 13542400+5971968+460800+1200=19976368\n\n1000 epochs, 24300 training examples\n\n19976368*3*24300*1000=1456277227200000=1.5e15\n"},{"id":"epoch:fa8a1743029cd23ce36e:parameters","entityId":"epoch:fa8a1743029cd23ce36e","metric":"parameters","value":16210006.0,"unit":"parameters","observationPeriod":"2010-06-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The stereo-pair images are subsampled from their original resolution of 108 \u00d7 108 \u00d7 2 to 32 \u00d7 32 \u00d7 2 to speed up experiments [...]  the architecture\nwith the best results have 4000 units in the first layer\nand 2000 in the second [...] there are 58,320 test\ncases (9,720 cases per class) \"\n\nSo the architecture has (32*32*2+1)x4000 + (4000+1)*2000 + (2000+1)*58,320/9,720 parameters"},{"id":"epoch:acd6ff85ad90fe9d5dde:training-compute","entityId":"epoch:acd6ff85ad90fe9d5dde","metric":"training-compute","value":1080000000000000.0,"unit":"FLOP","observationPeriod":"2010-06-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\u201cAll experiments are performed on a standard computer with a 3Ghz CPU\u201d\n\u201cIt takes 25 hours (about 25 iCCCP iterations) to train an object class with two mixture templates.\u201d\n\nIndividual models for some of the classes.\n\nAssuming a Core 2 with 8 FLOP/cycle was used with a utilization of 0.5:\n3000000000*8*25*60*60*0.5=1080000000000000"},{"id":"epoch:06ebb3327828d3f534dc:parameters","entityId":"epoch:06ebb3327828d3f534dc","metric":"parameters","value":7082000.0,"unit":"parameters","observationPeriod":"2010-05-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"pg250 of the paper, section 2.3: \n\"We optimized feedforward neural networks with one to\nfive hidden layers, with one thousand hidden units per\nlayer\"\n\nInput is a flattened 32x32 image, which corresponds to an input vector of length 3072\n\nOutput is a number from 0-9, so 10 neurons\n\nNo. of params: 3072*1000 + 4*1000*1000 + 1000*10 = 7,082,000\n"},{"id":"epoch:06ebb3327828d3f534dc:training-compute","entityId":"epoch:06ebb3327828d3f534dc","metric":"training-compute","value":350000000000000.0,"unit":"FLOP","observationPeriod":"2010-05-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Roughly two times the number of parameters for ops per forward pass. \n\nSo 2*7082000 params*3.5*140 epochs * 50k training images = 3.5e14"},{"id":"epoch:62ce646b920d980e9660:parameters","entityId":"epoch:62ce646b920d980e9660","metric":"parameters","value":12110000.0,"unit":"parameters","observationPeriod":"2010-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 1"},{"id":"epoch:62ce646b920d980e9660:training-compute","entityId":"epoch:62ce646b920d980e9660","metric":"training-compute","value":130788000000000.0,"unit":"FLOP","observationPeriod":"2010-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Networks with up to 12 million weights can successfully be trained by plain gradient descent to achieve test errors below 1% after 20-30 epochs in less than 2 hours of training.\"\n\n60k images in each MNIST epoch.\n\nArchitecture-based estimate: 6 * 12.11M * 60k * 30 = 1.31e14\n\nWe can also get a rough hardware estimate. The authors use single precision, GTX280 gets 6.221e11 FLOPs in single precision. Training 30 epochs takes less than 2 hours, but on each epoch the training set is augmented in online fashion, which takes 93 seconds.\n(2*3600 - 93*30) * 6.221e11 * 0.3 = 8.23e14\n\nThe architecture approach seems less uncertain."},{"id":"epoch:422cacdf1478ebf21283:parameters","entityId":"epoch:422cacdf1478ebf21283","metric":"parameters","value":1025.0,"unit":"parameters","observationPeriod":"2010-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Somewhat low confidence, but it seems like the number of learnable parameters is the size of the codebook, plus parameters in the SVM used for classification. Since it is a linear SVM, there will be one parameter per input feature, plus a single bias term.\n\nSo in total, 512 learnable codebook values, plus 513 SVM parameters = 1025 parameters"},{"id":"epoch:dce3f8af9fd0463de2b7:parameters","entityId":"epoch:dce3f8af9fd0463de2b7","metric":"parameters","value":13729792.0,"unit":"parameters","observationPeriod":"2009-11-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Locally connected: 128*128*4*4*4*7 + 64*64*8*4*4*4 + 32*32*16*4*4*8 + 16*16*32*4*4*16=13631488\nClassification head: 16*16*32*12=98304\nTotal: 13631488+98304=13729792\n\"five regular layers with the dimensions 256\u00d7256, 128\u00d7128, . . ., 16\u00d716.\"\n\"size of the receptive field to be 4 \u00d7 4 neurons\""},{"id":"epoch:dce3f8af9fd0463de2b7:training-compute","entityId":"epoch:dce3f8af9fd0463de2b7","metric":"training-compute","value":3295150080000000.0,"unit":"FLOP","observationPeriod":"2009-11-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2*13729792*3*40000*1000=3295150080000000"},{"id":"epoch:dfdd47a7ee380f6b739f:parameters","entityId":"epoch:dfdd47a7ee380f6b739f","metric":"parameters","value":11616256.0,"unit":"parameters","observationPeriod":"2009-11-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Locally connected: 128*128*4*4*4*5 + 64*64*8*4*4*4 + 32*32*16*4*4*8 + 16*16*32*4*4*16=11534336\nClassification head: 16*16*32*10=81920\nTotal: 11534336+81920=11616256\n\"five regular layers with the dimensions 256\u00d7256, 128\u00d7128, . . ., 16\u00d716.\"\n\"size of the receptive field to be 4 \u00d7 4 neurons\""},{"id":"epoch:dfdd47a7ee380f6b739f:training-compute","entityId":"epoch:dfdd47a7ee380f6b739f","metric":"training-compute","value":4181852160000000.0,"unit":"FLOP","observationPeriod":"2009-11-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2*11616256*3*60000*1000=4181852160000000"},{"id":"epoch:f4f594efc8a69806e619:parameters","entityId":"epoch:f4f594efc8a69806e619","metric":"parameters","value":16818176.0,"unit":"parameters","observationPeriod":"2009-11-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Locally connected: 128*128*4*4*4*10 + 64*64*8*4*4*4 + 32*32*16*4*4*8 + 16*16*32*4*4*16=16777216\nClassification head: 16*16*32*5=40960\nTotal: 16777216+40960=16818176\n\"five regular layers with the dimensions 256\u00d7256, 128\u00d7128, . . ., 16\u00d716.\"\n\"size of the receptive field to be 4 \u00d7 4 neurons\""},{"id":"epoch:f4f594efc8a69806e619:training-compute","entityId":"epoch:f4f594efc8a69806e619","metric":"training-compute","value":2452090060800000.0,"unit":"FLOP","observationPeriod":"2009-11-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2*16818176*3*24300 *1000=2452090060800000"},{"id":"epoch:0a01c76642f06f2f32b1:parameters","entityId":"epoch:0a01c76642f06f2f32b1","metric":"parameters","value":258800.0,"unit":"parameters","observationPeriod":"2009-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"CNN: 32*5*5+32*64*5*5=52000\nFC: 1024*200+200*10=206800\nTotal: 52000+206800=258800"},{"id":"epoch:0a01c76642f06f2f32b1:training-compute","entityId":"epoch:0a01c76642f06f2f32b1","metric":"training-compute","value":20754000000000.0,"unit":"FLOP","observationPeriod":"2009-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming no padding and stride 1\n\u201con the 50,000 non-validation training samples until the best error rate on the validation set was reached (this took 30 epochs)\u201c\nFirst CNN layer: 2*32*24*24*5*5=921600\nSecond CNN layer: 2*64*16*5*5*8*8=3276800\nFirst FC layer: 2*4*4*64*200=409600\nSecond FC layer: 2*200*10=4000\nTotal forward FLOP: 921600+3276800+409600+4000=4612000\nTotral training compute: 4612000*3*30*50000=20754000000000\n\nThey train another model in this paper that I think is at least an order of magnitude larger than this one; however, they do not provide the number of epochs it was trained for, so I'm only able to calculate the FLOPs for the smaller model (which still claims SOTA performance on its own dataset). \n\n"},{"id":"epoch:1e4620607069eff00ae8:parameters","entityId":"epoch:1e4620607069eff00ae8","metric":"parameters","value":14423.0,"unit":"parameters","observationPeriod":"2009-08-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"C1: 6x7x7\nC3: 61x7x7\nC5: 305x6x6\nF6: 80x2\nTotal = 14423"},{"id":"epoch:1e4620607069eff00ae8:training-compute","entityId":"epoch:1e4620607069eff00ae8","metric":"training-compute","value":306000000000000.0,"unit":"FLOP","observationPeriod":"2009-08-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(2*340 million) * 3 * 30000 * 5\n\"340 million connection network\"\noperations per fwd pass ~= 2*340 million\n\"With a fixed detection threshold, the system reaches a roughly 3% equal error rate on this dataset after only 5 training epochs through the training set.\"\n"},{"id":"epoch:bcc2677be03d62160cb2:parameters","entityId":"epoch:bcc2677be03d62160cb2","metric":"parameters","value":133140.0,"unit":"parameters","observationPeriod":"2009-08-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use a policy gradient\nalgorithm to estimate the parameters of a log-linear model for action selection [...] In total, there are 4,438 features [in the Windows domain]. [...]  This difficulty can be attributed in part to the large branching factor of possible actions at each step \u2014 on average, there are 27.14 choices per action in the Windows domain\""},{"id":"epoch:be63886e49ebd2a63275:parameters","entityId":"epoch:be63886e49ebd2a63275","metric":"parameters","value":80940.0,"unit":"parameters","observationPeriod":"2009-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We use a policy gradient\nalgorithm to estimate the parameters of a log-linear model for action selection [...] In total, there are 8,094 features [in the Crossblock domain]. [...]  This difficulty can be attributed in part to the large branching factor of possible actions at each step \u2014 on average, there are [...] 9.78 [actions] in the Crossblock\ndomain\""},{"id":"epoch:530fc86dbfb4eb0cb890:parameters","entityId":"epoch:530fc86dbfb4eb0cb890","metric":"parameters","value":1000000.0,"unit":"parameters","observationPeriod":"2009-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"roughly 1M parameters\""},{"id":"epoch:530fc86dbfb4eb0cb890:training-compute","entityId":"epoch:530fc86dbfb4eb0cb890","metric":"training-compute","value":1.0000000001e+18,"unit":"FLOP","observationPeriod":"2009-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Number of epochs is unknown.\n\nBased on 18,598 CPU-hours with Intel Xeon 5500 processors (speculative), this would be up to 8e17 FP64 FLOP or 1.6e18 FP32 FLOP. With 60% utilization it's around 1e18 FLOP."},{"id":"epoch:498d6dfc7c6cbe1b61e3:parameters","entityId":"epoch:498d6dfc7c6cbe1b61e3","metric":"parameters","value":100000000.0,"unit":"parameters","observationPeriod":"2009-06-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For example, we are able to reduce the time required to learn a four-layer DBN with 100 million free parameters from several weeks to around a single day.\""},{"id":"epoch:498d6dfc7c6cbe1b61e3:training-compute","entityId":"epoch:498d6dfc7c6cbe1b61e3","metric":"training-compute","value":1000000000000000.0,"unit":"FLOP","observationPeriod":"2009-06-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"https://www.getguesstimate.com/models/19602\n\n6435 GPU seconds for 1M examples\nSingle GTX 280 with 622.1 GFLOPS\nAll results are reported for 1M examples, unclear if they ran larger training experiments."},{"id":"epoch:90d64ce71255a7bf2b0d:parameters","entityId":"epoch:90d64ce71255a7bf2b0d","metric":"parameters","value":80000000.0,"unit":"parameters","observationPeriod":"2009-04-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Best performing model (see Figure 3.1) had 10,000 hidden units in one hidden layer and 8000 visible units"},{"id":"epoch:3b4938279bbaa3b12ed8:parameters","entityId":"epoch:3b4938279bbaa3b12ed8","metric":"parameters","value":12410.0,"unit":"parameters","observationPeriod":"2009-01-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"C1: 20x7x6\nC3: 369x6x5\nF4: 100x5\n\n\"This particular network has two convolutional\nlayers and one subsampling layer. The first convolutional layer has 20 7 \u00d7 6 filters, and the second layer has 369 6 \u00d7 5 filters. The layers are not fully connected; in particular, the Y, U, and V filters are kept\nseparate throughout the two layers. For\nthe purposes of training, a final 100-component, fully\nconnected layer is trained with five outputs\""},{"id":"epoch:3b4938279bbaa3b12ed8:training-compute","entityId":"epoch:3b4938279bbaa3b12ed8","metric":"training-compute","value":278721000000000.0,"unit":"FLOP","observationPeriod":"2009-01-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Every filter is only connected to \u2153 of the input channels: \u201cThe layers are not fully connected; in particular, the Y, U, and V filters are kept separate throughout the two layers.\u201d\nThe text of the paper and Figure 7 don\u2019t match, the text mentions an additional D=369 convolutional layer that is not represented in the Figure\nInput: 12*25\n1st CNN: 2*12*25*(3/3)*7*6*20=240000\nSize after layer: 6*20*20\nSize after pooling layer (1x4): 6*5*20\n2nd CNN: 2*6*5*(20/3)*6*5*369=4428000\nFC: 2*6*5*369*100=2214000\nTotal forward FLOP: 240000+4428000+2214000=6882000\nTotal training compute: 6882000*3*450000*30=278721000000000=2.8e14\n\n\n\"The network was initialized with random values\nand trained for 30 epochs\""},{"id":"epoch:c8af69955a38c7a18cb6:parameters","entityId":"epoch:c8af69955a38c7a18cb6","metric":"parameters","value":18030592.0,"unit":"parameters","observationPeriod":"2009-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"429*2048+2048*2048*4+2048*183=18030592\n\nTable 2: largest model has 5 layers of 2048 into 183 softmax, 429 input"},{"id":"epoch:15867d099fb6c4ac3eae:parameters","entityId":"epoch:15867d099fb6c4ac3eae","metric":"parameters","value":2600000.0,"unit":"parameters","observationPeriod":"2008-12-10","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:7a5c6656d3323966d0fe:parameters","entityId":"epoch:7a5c6656d3323966d0fe","metric":"parameters","value":30.0,"unit":"parameters","observationPeriod":"2008-12-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"5*5+5=30\n(3 layer network with 5 hidden neurons)"},{"id":"epoch:7a5c6656d3323966d0fe:training-compute","entityId":"epoch:7a5c6656d3323966d0fe","metric":"training-compute","value":1614600000.0,"unit":"FLOP","observationPeriod":"2008-12-09","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2*30 parameters *5000 epochs *207 training examples*26 nodes per example=1614600000"},{"id":"epoch:8c0cd59c95285bce7274:parameters","entityId":"epoch:8c0cd59c95285bce7274","metric":"parameters","value":1846400.0,"unit":"parameters","observationPeriod":"2008-12-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Except for where stated otherwise, the models used for the experiments used 100 dimensional feature vectors and a context size of 5.\"\n\"The vocabulary size for this dataset is 17964.\"\nEmbedding: 17964 * 100 = 1796400\nContext matrices: 5 * 100 * 100 = 50000\nUnembedding: 0 (tied embedding \u201cwhile the matrix of weights from the hidden layer to the output layer is simply the feature vector matrix R\u201d)\nTotal: 1796400 + 50000 = 1846400\n"},{"id":"epoch:e1bd3b53ed427475e4d4:parameters","entityId":"epoch:e1bd3b53ed427475e4d4","metric":"parameters","value":12198000.0,"unit":"parameters","observationPeriod":"2008-12-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"None given but it does say \" Since each non-leaf node in a tree has its own feature vector, the number of free parameters associated with the tree is linear in this quantity\", and the largest model (T7: ADAPATIVE(0.4) x 4) has 121980 of them. The feature vectors are 100-dimensional. I've done the dubious thing of multiplying the two to give an estimate."},{"id":"epoch:56cece334b04687a5846:parameters","entityId":"epoch:56cece334b04687a5846","metric":"parameters","value":3000000.0,"unit":"parameters","observationPeriod":"2008-07-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:758f186163c9b18fb325:parameters","entityId":"epoch:758f186163c9b18fb325","metric":"parameters","value":1500000.0,"unit":"parameters","observationPeriod":"2008-07-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"With a word vector size of 50 and a vocabulary size of 30,000, the embedding matrix has 1,500,000 parameters. There are also some small convolutional and dense layers with far fewer parameters."},{"id":"epoch:13477e6c51b8420a0fd1:parameters","entityId":"epoch:13477e6c51b8420a0fd1","metric":"parameters","value":100881.0,"unit":"parameters","observationPeriod":"2007-12-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"For the raw input representation,\nthere were 4 input units and a total of 100,881 weights"},{"id":"epoch:ee300ff552f927593afb:parameters","entityId":"epoch:ee300ff552f927593afb","metric":"parameters","value":300000000000.0,"unit":"parameters","observationPeriod":"2007-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:ee300ff552f927593afb:training-compute","entityId":"epoch:ee300ff552f927593afb","metric":"training-compute","value":1.4494464e+18,"unit":"FLOP","observationPeriod":"2007-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Assuming a Nehalem based processor with 8 FLOP/cycle (https://www.agner.org/optimize/microarchitecture.pdf#page=105.06) , 2 cores and 2.33 GHz clock speed: 8*2*2330000000=37280000000 FLOP/s\nTrained for 1 day on 1500 machines (Table 2)\nCompute: 1500*37280000000*1*24*60*60*0.3=1449446400000000000=1.4e18\n"},{"id":"epoch:2d0a7016a18ff0ea8623:parameters","entityId":"epoch:2d0a7016a18ff0ea8623","metric":"parameters","value":21000000000.0,"unit":"parameters","observationPeriod":"2007-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2"},{"id":"epoch:2d0a7016a18ff0ea8623:training-compute","entityId":"epoch:2d0a7016a18ff0ea8623","metric":"training-compute","value":7.7303808e+17,"unit":"FLOP","observationPeriod":"2007-06-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Trained for 2 days on 400 machines (Table 2)\nAssuming a Nehalem based processor with 8 FLOP/cycle (https://www.agner.org/optimize/microarchitecture.pdf#page=105.06) , 2 cores and 2.33 GHz clock speed: 8*2*2330000000=37280000000 FLOP/s\nCompute: 400*37280000000*2*24*60*60*0.3=773038080000000000=7.7e17"},{"id":"epoch:501c2f63a2858e53274f:training-compute","entityId":"epoch:501c2f63a2858e53274f","metric":"training-compute","value":9604800000000.0,"unit":"FLOP","observationPeriod":"2006-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(Just for natural images)\n\n\"All experiments were conducted on a Linux machine with AMD Opteron 2GHz CPU and 2GB RAM. ... For example, we were able to learn a set of 1,024 bases (each 14\u00d714 pixels)in about 2 hours and a set of 2,000 bases (each 20\u00d720 pixels) in about 10 hours.\"\n\nI filtered for 2GHz Opteron models that came out in 2005, of which there are five: https://www.techpowerup.com/cpu-specs/?mfgr=AMD&released=2005&generation=AMD%20Opteron&sort=name\n\nFound a source which indicates 3 cycles per 32-bit multiply, and 5 per 64 bit: https://www.cse.wustl.edu/~roger/569M/m2066.pdf \nAssuming 32-bit precision, 2e9 cycles/s / 3 cycle/FLOP = 6.67e8 FLOP/s\n\n10 * 3600 * 6.67e8 * 0.4 = 9.6048e12 FLOPs"},{"id":"epoch:1972c522bdb35a77f414:parameters","entityId":"epoch:1972c522bdb35a77f414","metric":"parameters","value":1600000.0,"unit":"parameters","observationPeriod":"2006-07-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:bc3a24ed42ddebf3703d:parameters","entityId":"epoch:bc3a24ed42ddebf3703d","metric":"parameters","value":3800000.0,"unit":"parameters","observationPeriod":"2006-07-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:6b28d493b46c9ede50bc:parameters","entityId":"epoch:6b28d493b46c9ede50bc","metric":"parameters","value":100.0,"unit":"parameters","observationPeriod":"2006-07-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The main learning algorithm is a logistic classifier. The input is a matrix M, where the rows are pairs of words, and the columns (variables) are counts of occurrences of synthetic dependency paths between those two words.\n\nSince there are on the order of 10~100 different types of syntactic relationships, this is the number of length-1 paths, and thus the number of parameters if only length-1 paths are used.\n\nHowever, if the length of the paths considered is longer (say, 5), then the parameters would be on the order of (10~100)^5. It's not clear to me which is the case"},{"id":"epoch:4ac58152f97a3a2a9562:parameters","entityId":"epoch:4ac58152f97a3a2a9562","metric":"parameters","value":114662.0,"unit":"parameters","observationPeriod":"2006-06-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The hidden layers were fully connected to themselves\nand the output layer, and fully connected from the input layer. The input layer was size 26, the softmax output layer size 62 (61 phoneme categories plus the blank label), and the total number of weights was\n114, 662.\"\n\nhttps://www.cs.toronto.edu/~graves/icml_2006.pdf"},{"id":"epoch:43a9ac46566eb3d76841:parameters","entityId":"epoch:43a9ac46566eb3d76841","metric":"parameters","value":37097.0,"unit":"parameters","observationPeriod":"2006-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Architecture described in figure 3"},{"id":"epoch:1d4f8277a1ce025daa71:parameters","entityId":"epoch:1d4f8277a1ce025daa71","metric":"parameters","value":3590057.0,"unit":"parameters","observationPeriod":"2006-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"For the CNN: \"In this case, the whole network has a total\nof 90,857 trainable parameters.\"\nFor the SVM: \nthere are 6 binary SVM models trained \nsupport vectors = 0.02x291600\ndimension = 100\nparameters = (0.02x291600)x100x6=3499200\nTotal = 3590057"},{"id":"epoch:1d4f8277a1ce025daa71:training-compute","entityId":"epoch:1d4f8277a1ce025daa71","metric":"training-compute","value":97977600000000.0,"unit":"FLOP","observationPeriod":"2006-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(291,600)*(8e6)*3*14\n\"A full propagation of one data sample through the network requires about 4 million multiply-add operations\" = 2*4e6 FLOPs\n\"The training and testing features are extracted with\nthe convolutional net trained after 14 passes\"\n\"The experiments here are run on a single CPU (AMD Opteron at 2GHz) with 4GB of memory\""},{"id":"epoch:1cc5828a892ec600dbb5:parameters","entityId":"epoch:1cc5828a892ec600dbb5","metric":"parameters","value":90857.0,"unit":"parameters","observationPeriod":"2006-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":""},{"id":"epoch:1cc5828a892ec600dbb5:training-compute","entityId":"epoch:1cc5828a892ec600dbb5","metric":"training-compute","value":745200000000000.0,"unit":"FLOP","observationPeriod":"2006-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training time: 5880+330=6210 minutes (Table 1)\n\u201cThe timing is normalized to hypothetical 1GHz single CPU.\u201d\nAssuming utilization of 0.5 and AMD Athlon 64 Processor with 4 FLOP/cycle\nCompute: 6210*60*4*1000000000*0.5=745200000000000 = 7.4e14\n"},{"id":"epoch:58aabc5f90e48ac05729:parameters","entityId":"epoch:58aabc5f90e48ac05729","metric":"parameters","value":72000.0,"unit":"parameters","observationPeriod":"2005-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The network has 3.15 Million connections and about 72,000 trainable parameters.\""},{"id":"epoch:eeb4df4276f1edf266ca:parameters","entityId":"epoch:eeb4df4276f1edf266ca","metric":"parameters","value":1472256.0,"unit":"parameters","observationPeriod":"2005-12-05","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In detail, we use different parameters (\u03b8r, \u03c31r, \u03c32r) for each row in the image, because the images we consider are taken from a horizontally mounted camera, and thus different rows of the image have different statistical properties.\"\n\"We collected a total of 425 image+depthmap pairs, with an image resolution of 1704x2272 and a depthmap resolution of 86x107\"\n\nThe dimensionality of each parameter set isn't totally clear, but from equation (1) it seems like \u03b8 is the same length as the absolute depth input features (646) and \u03c31 and \u03c32 are scalar. If so, then we should have:\n2272 * (646 + 1 + 1) = 1,472,256 parameters"},{"id":"epoch:8e0fba30f7ef65750711:parameters","entityId":"epoch:8e0fba30f7ef65750711","metric":"parameters","value":5711.0,"unit":"parameters","observationPeriod":"2005-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Model is \"a two layer net with 10 hidden units\"\nInput is of size 569 \"In all, we use 569 features\"\nParameters:\n569*10 + 10 for hidden layer\n10*1 + 1 for output layer"},{"id":"epoch:8e0fba30f7ef65750711:training-compute","entityId":"epoch:8e0fba30f7ef65750711","metric":"training-compute","value":3482081588304.0,"unit":"FLOP","observationPeriod":"2005-08-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"FLOPs per forward pass: 2*parameters = 11422\nFLOPs per pair (data point): two forward passes and one backward pass (\"A forward prop is performed for the first sample; each node\u2019s activation and gradient value are stored; a forward prop is then performed for the second sample, and the activations and gradients are again stored. The gradient of the cost is then *formula*\") = 2*11422 + 2*11422 = 45688\nTotal FLOPs = (FLOPs per pair = 45688)*(pairs = 3,464,289)*(epochs = 22) = 3.48E12"},{"id":"epoch:598f12fea81bfda84752:parameters","entityId":"epoch:598f12fea81bfda84752","metric":"parameters","value":152061.0,"unit":"parameters","observationPeriod":"2005-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The hidden layer sizes were chosen to ensure that all networks had roughly the same number of weights W (\u2248100,000). However, for the MLPs the network grew with the time-window size, and W varied between 22,061 and 152,061.\""},{"id":"epoch:598f12fea81bfda84752:training-compute","entityId":"epoch:598f12fea81bfda84752","metric":"training-compute","value":24124575958774.88,"unit":"FLOP","observationPeriod":"2005-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Extracted from AI and Compute (https://openai.com/blog/ai-and-compute/) charts by using https://automeris.io/WebPlotDigitizer/.\n\nForward FLOP: 2*200000=400000\nTrained on 4158 utterances (TIMIT)\n\"We found that large networks, of around 200,000 weights,\ngave good performance\"\nTIMIT has around 5 hours total. Estimated utterance length: 5*60*60/(4620+1680)=2.86s\nThe frame size was 5 ms\nFrames per utterance: 2.86/0.005=572\nTraining FLOP for one epoch: 3*400000*572*4158=2854051200000\n\nEpochs unclear, based on OpenAIs estimate it would be 8. "},{"id":"epoch:addd2baa4122cdd12fd4:parameters","entityId":"epoch:addd2baa4122cdd12fd4","metric":"parameters","value":120000000.0,"unit":"parameters","observationPeriod":"2005-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Very unsure, but the paper mentions \n\"We ran the training process of Section 3 on the same data, obtaining a grammar of 24M rules\" \nand \n\"For our experiments we used the following features, analogous to Pharaoh\u2019s default feature set:\n\u2022 P(\u03b3 | \u03b1) and P(\u03b1 | \u03b3), the latter of which is not\nfound in the noisy-channel model, but has been\npreviously found to be a helpful feature (Och\nand Ney, 2002);\n\u2022 the lexical weights Pw(\u03b3 | \u03b1) and Pw(\u03b1 | \u03b3) (Koehn et al., 2003), which estimate how well the words in \u03b1 translate the words in \u03b3;\n2\n\u2022 a phrase penalty exp(1), which allows the\nmodel to learn a preference for longer or\nshorter derivations, analogous to Koehn\u2019sphrase penalty (Koehn, 2003).\"\n\nSuggesting 24M rules * 5 features per rule (?)"},{"id":"epoch:528b0d90f6a6e3cb2cf0:parameters","entityId":"epoch:528b0d90f6a6e3cb2cf0","metric":"parameters","value":178.0,"unit":"parameters","observationPeriod":"2005-04-22","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From https://www.bnlearn.com/bnrepository/"},{"id":"epoch:4f2d7f27d6bafa738ca1:training-compute","entityId":"epoch:4f2d7f27d6bafa738ca1","metric":"training-compute","value":115848000000000.0,"unit":"FLOP","observationPeriod":"2005-01-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The computations were performed on Athlon processors with a 1.2 GHz clock\"\nFP32 per cycle: 4 (\"The bottom line is that the Athlon is capable of delivering as many as four 32-bit, single-precision floating-point results per clock cycle\", https://www.pctechguide.com/amd-technology/amd-athlon) \nTraining time per epoch: 1609s (table 1)\nEpochs: 30 \"Training is performed over about 20 to 30 epochs according to validation set perplexity (early stopping).\"\nAssumed utilization: 0.5 \nCompute estimate: 0.5*1200000000*4*30*1609=115848000000000=1.16e14"},{"id":"epoch:ba027c3bd984f8a4196b:parameters","entityId":"epoch:ba027c3bd984f8a4196b","metric":"parameters","value":4096000.0,"unit":"parameters","observationPeriod":"2004-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"64*64*1000=4096000\n\"100000 samples of natural scenes of 64 \u00d7 64 pixels were given as X\"\n\"LMICA was carried out in 1000 layers\"\n"},{"id":"epoch:ba027c3bd984f8a4196b:training-compute","entityId":"epoch:ba027c3bd984f8a4196b","metric":"training-compute","value":2782080000000000.0,"unit":"FLOP","observationPeriod":"2004-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"69*60*60*8*2800000000*0.5=2782080000000000=2.78e15\n\"it consumed about 69 hours with Intel 2.8GHz CPU\"\n- Assuming they used an Intel Pentium 4 processor with 8 FLOP/cycle (https://en.wikipedia.org/wiki/FLOPS)"},{"id":"epoch:156f3c211696e8d58c19:parameters","entityId":"epoch:156f3c211696e8d58c19","metric":"parameters","value":16636.0,"unit":"parameters","observationPeriod":"2004-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"We employ a network architecture similar to LeNet5 [6]. The difference is in the number of maps. In our architecture we have 8 feature maps in the bottom convolutional and subsampling layers and 20 maps in the next two layers. The last layer has 9 outputs to encode two pose parameters\"\nUsing the general architecture of LeNet5 from \"Gradient-based learning applied to document recognition\" (LeCun, 1998), we get:\nC1: CNN, D = 8, K=5, C=1; Parameters = 200\nC3: CNN, D=20, K=5, C=8; Parameters = 4,000\nC5: NN, D=20, K=5, C=20; Parameters = 10,000\nF6: NN, N=20, M=84; Parameters = 1,680\nOutput: NN, N=84, M=9; Parameters = 756\nPooling layers: negligible\nParameters = 16,636"},{"id":"epoch:156f3c211696e8d58c19:training-compute","entityId":"epoch:156f3c211696e8d58c19","metric":"training-compute","value":52603200000000.0,"unit":"FLOP","observationPeriod":"2004-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"26*60*60*[(2.81E+03)*((1E6)/2.5)]*0.5\n\n\"Training was performed using LUSH [1], and the total training\ntime was about 26 hours on a 2Ghz Pentium 4\"\n\n\"The Progress of Computing\" (Nordhaus, 2001) says that a Pentium 4 runs at 2.81E+03 MSOPS, and that \"Most benchmarks find that 1 million flops correspond to between 2 and 3 MSOPS\"\n\nA utilization rate of 0.5 was estimated since this is a smaller model that ran on a CPU (both of which suggest a slightly higher utilization rate than 0.3 to 0.4)"},{"id":"epoch:c845fe1beadf86f6f17c:parameters","entityId":"epoch:c845fe1beadf86f6f17c","metric":"parameters","value":100000.0,"unit":"parameters","observationPeriod":"2004-07-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For the first modification of the Rosenblatt perceptron 10 neurons were included into the R-layer. [...] The number of the A-layer neurons was 256,000\" The relation between the S-layer and A-layer is hardcoded"},{"id":"epoch:fe68749ff54a4dfc7777:parameters","entityId":"epoch:fe68749ff54a4dfc7777","metric":"parameters","value":90575.0,"unit":"parameters","observationPeriod":"2004-06-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The network has a total of 90,575 trainable parameters.\""},{"id":"epoch:fe68749ff54a4dfc7777:training-compute","entityId":"epoch:fe68749ff54a4dfc7777","metric":"training-compute","value":974230000000.0,"unit":"FLOP","observationPeriod":"2004-06-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"A full propagation through the network requires 3,896,920 multiply-adds.\" - it's not entirely clear whether this refers to a forward pass or forward + backward pass (I assumed the latter)\n\"We used a stochastic version of the Levenberg-Marquardt algorithm with diagonal approximation of the Hessian [7], for approximately 250,000 online updates.\"\n3896920*250000=974230000000=9.7e11"},{"id":"epoch:c43dd3e0e215f8343864:parameters","entityId":"epoch:c43dd3e0e215f8343864","metric":"parameters","value":100.0,"unit":"parameters","observationPeriod":"2003-10-13","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"citation from section 2.1: In the constellation model, the dimensionality of \\theta is large (~ 100) "},{"id":"epoch:23554230414c68ed8e37:parameters","entityId":"epoch:23554230414c68ed8e37","metric":"parameters","value":451.0,"unit":"parameters","observationPeriod":"2003-06-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 1\n"},{"id":"epoch:11c91a922dec71922496:parameters","entityId":"epoch:11c91a922dec71922496","metric":"parameters","value":9178890.0,"unit":"parameters","observationPeriod":"2003-05-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"There are various components to the system:\n\n- Translation probability model phi\n- The distortion probability distribution d\n- A langage model p_LM\n- A length factor w\n\nSeveral translation probability models are considered. The most performant one is the AP word alignment model. The sentence length preferred by the authors is 3 words maximum. In the biggest corpus considered (320k phrase pairs) it produces a phrase translation probability table of 1996k entries.\n\nThe distortion probability model d is taken from  (Marcu and Wong, 2002).\n\nThe distortion probability model must have ~10 parameters at most\n\nThe language model p_LM is a back off trigram model from (Seymore and Rosenfeld,1997). AFAIK the cutoff used is not specified. Based on the example on section 4.3 of (Seymore and Rosefeld, 1997), a trigram probability model has about 3866964 + 2674322 + 641604 parameters.\n\n\"For each possible phrase translation anywhere in the sentence (we call it a translation option), we multiply its phrase translation probability with the language model probability for the generated English phrase. As language model probability we use the unigram probability for the first word, the bigram probability for the second, and the trigram probability for all following words\"\n\nThe length factor w is an additional single parameter.\n\n\"In order to calibrate the output length, we introduce a\nfactor w for each generated English word in addition to\nthe trigram language model \"\n\nIn summary, the parameter count seems to be dominated by the trigram language model and the word alignment phrase translation model. "},{"id":"epoch:f5b4e51003e9e0456150:parameters","entityId":"epoch:f5b4e51003e9e0456150","metric":"parameters","value":4124233.0,"unit":"parameters","observationPeriod":"2003-03-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The number of free parameters is |V|(1 + nm + h) + h(1 + (n \u2212 1)m) [...] For example, consider the following architecture used in the experiments on the AP (Associated Press) news data: the vocabulary size is |V| = 17,964, the number of hidden units is h = 60, the order of the model is n = 6, the number of word features is m = 100\"\n\nBrown corpus: n=5, h=100, m=30, |V|=16383\n16383*(1+5*30+100)+100*(1+(5-1)*30)=4124233"},{"id":"epoch:f5b4e51003e9e0456150:training-compute","entityId":"epoch:f5b4e51003e9e0456150","metric":"training-compute","value":132076260000000.0,"unit":"FLOP","observationPeriod":"2003-03-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For example, consider the following architecture used in the experiments on the AP (Associated Press) news data: the vocabulary size is |V| = 17,964, the number of hidden units is h = 60, the order of the model is n = 6, the number of word features is m = 100. The total number of numerical operations to process a single training example is approximately |V|(1+nm+h)+h(1+nm)+nm\"\n\nBrown corpus: n=5, h=100, m=30, |V|=16383, dataset size = 800000, epochs=20\nForward FLOP: 16383*(1+5*30+100)+100*(1+5*30)+5*30=4127383\nAdjusting for backward pass with 1:1 ratio, as the by far largest layer (embedding) doesn't require gradients w.r.t. inputs.\nTotal FLOP: 2*4127383*800000*20=1.3207626e+14"},{"id":"epoch:ff469bda36852d191860:parameters","entityId":"epoch:ff469bda36852d191860","metric":"parameters","value":11904264.0,"unit":"parameters","observationPeriod":"2003-03-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The number of free parameters is |V|(1 + nm + h) + h(1 + (n \u2212 1)m) [...] For example, consider the following architecture used in the experiments on the AP (Associated Press) news data: the vocabulary size is |V| = 17,964, the number of hidden units is h = 60, the order of the model is n = 6, the number of word features is m = 100\"\n\nAP News: n=6, h=60, m=100, |V|=17964\n17964*(1+6*100+60)+60*(1+(6-1)*100)=11904264"},{"id":"epoch:ff469bda36852d191860:training-compute","entityId":"epoch:ff469bda36852d191860","metric":"training-compute","value":1666869200000000.0,"unit":"FLOP","observationPeriod":"2003-03-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"For example, consider the following architecture used in the experiments on the AP (Associated Press) news data: the vocabulary size is |V| = 17,964, the number of hidden units is h = 60, the order of the model is n = 6, the number of word features is m = 100. The total number of numerical operations to process a single training example is approximately |V|(1+nm+h)+h(1+nm)+nm\"\n\nAP News: n=6, h=60, m=100, |V|=17964, dataset=13994528, epochs=5\nForward FLOP: 17964*(1+6*100+60)+60*(1+6*100)+6*100=11910864\nAdjusting for backward pass with 1:1 ratio, as the by far largest layer (embedding) doesn't require gradients w.r.t. inputs.\nTotal FLOP: 2*11910864*13994528*5=1.6668692e+15"},{"id":"epoch:683218d594ab9480593c:parameters","entityId":"epoch:683218d594ab9480593c","metric":"parameters","value":12000.0,"unit":"parameters","observationPeriod":"2001-12-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Iteratively learns decision tree features, where each feature has 2 parameters (threshold and parity). They learn 6000 features total: \"The complete face detection cascade has 38 stages with over\n6000 features\"\n\nAdditional weights (see Table 1) are a temporary variable of the optimization and not part of the final model. "},{"id":"epoch:683218d594ab9480593c:training-compute","entityId":"epoch:683218d594ab9480593c","metric":"training-compute","value":63000000000000.0,"unit":"FLOP","observationPeriod":"2001-12-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\nThe training compute can be tediously worked out from the pseudocode. I think for dataset size D, number of filters T, the training compute is roughly 180k * D * 3 * T = 6.3e13 FLOPs\n\nThe training evaluates each of 180k candidate features at each step, repeated for 6000 steps (as 1 feature is selected per round). \nThe operations for one feature evaluation are unclear, but should be low (they only compare specific integer positions in the image). Estimated at 10op. \n180000*6000*10*14460=1.56168e+14"},{"id":"epoch:405fcd55f761542c7271:parameters","entityId":"epoch:405fcd55f761542c7271","metric":"parameters","value":3925310.0,"unit":"parameters","observationPeriod":"2000-11-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"10 models, one for each digit. Largest models: 500 epochs, 500 hidden units (Table 2)\n\"The largest network was the best, even though each digit model contains 392,500 parameters trained on only 4,400 images\"\n\"the classification network had 30 inputs and therefore 300 weights and 10 output biases.\"\n\nTotal: 392500*10 + 310 = 3,925,310"},{"id":"epoch:405fcd55f761542c7271:training-compute","entityId":"epoch:405fcd55f761542c7271","metric":"training-compute","value":51810000000000.0,"unit":"FLOP","observationPeriod":"2000-11-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Each model was trained on 4400 examples: \"The largest network was the best, even though each digit model contains 392,500 parameters trained on only 4,400 images.\"\n\nTable 2, largest network trained 500 epochs.\n10 * 6 * 392500 * 4400 * 500 = 51,810,000,000,000"},{"id":"epoch:e86cc4e831699f173561:parameters","entityId":"epoch:e86cc4e831699f173561","metric":"parameters","value":6906980.0,"unit":"parameters","observationPeriod":"2000-11-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"(30959*100) + (8*100*120) + (120*30959) = 6,906,980\n\"This is obtained with a network with the direct architecture, 100 randomly initialized words features, 120 hidden units, and n = 8 words of context.\"\n\"The Hansard corpus (Canadian parliament proceedings, French version) is a stream of about 34 million words, of which 32 millions (set A) was used for training, 1.1 million (set B) was used for validation, and 1.2 million (set C) was used for out-of-sample tests. The original data has 106, 936 different words, and those with frequency <= 10 were merged into a single token, yielding IVI = 30,959 different words.\"\n"},{"id":"epoch:e86cc4e831699f173561:training-compute","entityId":"epoch:e86cc4e831699f173561","metric":"training-compute","value":6339000000000000.0,"unit":"FLOP","observationPeriod":"2000-11-28","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The authors use a trick to avoid having to calculate the final layer for all possible words in the vocabulary. They precompute a \"short list\" of the most common word following any 2 precursor words with a smoothed trigram model, and then only calculate the softmax over words on the short list. This means only a negligible fraction of the unembedding parameters get used, so the effective number of parameters appears to be (30959*100) + (8*100*120) = 3,191,900\n\n\"Apparent convergence of the stochastic gradient descent procedure was obtained after around 10 epochs for Hansard\"\n\n6ND:\n6*3191900*33100000*10=6.339e15\n"},{"id":"epoch:f3017d43eb2c064d5ba4:parameters","entityId":"epoch:f3017d43eb2c064d5ba4","metric":"parameters","value":17.0,"unit":"parameters","observationPeriod":"2000-07-26","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In absence of the 3 peephole connections there are 14 adjustable weights\""},{"id":"epoch:85f9f54c080fe0119d34:parameters","entityId":"epoch:85f9f54c080fe0119d34","metric":"parameters","value":324.0,"unit":"parameters","observationPeriod":"1999-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The size of the credibility network is 256-64-4. The 64 middle layer units are meant to encode low level features, while each of the 4 top level units are meant to encode a digit class"},{"id":"epoch:85f9f54c080fe0119d34:training-compute","entityId":"epoch:85f9f54c080fe0119d34","metric":"training-compute","value":5443200.0,"unit":"FLOP","observationPeriod":"1999-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"=6*324 parameters *2800 datapoints"},{"id":"epoch:d8949a527f6b9cce8e9b:parameters","entityId":"epoch:d8949a527f6b9cce8e9b","metric":"parameters","value":112000.0,"unit":"parameters","observationPeriod":"1999-06-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8*61*160+160*160+160*52=112000\n\"a network with 61 input units, 52 outputs, 160 hidden traits and 8 (4+I+3) delayed inputs\"\nTable 1\n"},{"id":"epoch:d8949a527f6b9cce8e9b:training-compute","entityId":"epoch:d8949a527f6b9cce8e9b","metric":"training-compute","value":3864000000000.0,"unit":"FLOP","observationPeriod":"1999-06-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" 2*112000*3*11.5*5000*100=3864000000000=3.9e12\n\"was trained for 100 epochs using the 5,000 pairs\"\n\"The length of the non-categorized Spanish sentences ranged from 3 to 20 and the length of the non-categorized English sentences, from 3 to 17\""},{"id":"epoch:49db556bdd183a0151bc:parameters","entityId":"epoch:49db556bdd183a0151bc","metric":"parameters","value":66780.0,"unit":"parameters","observationPeriod":"1999-06-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*50*140+140*140+140*37=66780\nTable 1\n\n"},{"id":"epoch:49db556bdd183a0151bc:training-compute","entityId":"epoch:49db556bdd183a0151bc","metric":"training-compute","value":8013600000000.0,"unit":"FLOP","observationPeriod":"1999-06-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2*66780*3*500*5000*8=8013600000000=8e12\n\"The number of words of the categorized sentences ranged from 3 to 13 for the Spanish ones and from 3 to 12 for the English ones.\"\n\"was trained up to 500 epochs using the 5,000 categorized pairs\""},{"id":"epoch:730a9fa64f66978aae65:parameters","entityId":"epoch:730a9fa64f66978aae65","metric":"parameters","value":276.0,"unit":"parameters","observationPeriod":"1999-01-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"See Table 1"},{"id":"epoch:fb7fe1e9b11e17e76ee6:parameters","entityId":"epoch:fb7fe1e9b11e17e76ee6","metric":"parameters","value":60000.0,"unit":"parameters","observationPeriod":"1998-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"[LeNet5] contains 390408 connections, but only 60000 trainable free parameters because of the weight sharing\""},{"id":"epoch:fb7fe1e9b11e17e76ee6:training-compute","entityId":"epoch:fb7fe1e9b11e17e76ee6","metric":"training-compute","value":2810937600000.0,"unit":"FLOP","observationPeriod":"1998-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"[LeNet5] contains 390408 connections\" = multiply-adds\nMNIST - 60,000 data points\n20 epochs\n390408*60000*6*20=2.810938e+12"},{"id":"epoch:be1a7d9e820e4f2f3737:parameters","entityId":"epoch:be1a7d9e820e4f2f3737","metric":"parameters","value":7512.0,"unit":"parameters","observationPeriod":"1998-05-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The RNN generated a total of eigt output prosodic parameters. [...] The numbers of nodes in the first and second hidden layers were determined empirically and set to be 35 and 30, respectively\"\n\nFigure 1 contains an overview of the architecture.\n\nLayer 1: (102 + 35 + 1)*35 parameters\nLayer 2: (43 + 35 + 1)*30 parameters\nOutput layer: (30+8+1)*8 parameters"},{"id":"epoch:be1a7d9e820e4f2f3737:training-compute","entityId":"epoch:be1a7d9e820e4f2f3737","metric":"training-compute","value":226690156032.01834,"unit":"FLOP","observationPeriod":"1998-05-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Extracted from AI and Compute (https://openai.com/blog/ai-and-compute/) charts by using https://automeris.io/WebPlotDigitizer/."},{"id":"epoch:211374ceef0c86afb27a:parameters","entityId":"epoch:211374ceef0c86afb27a","metric":"parameters","value":10504.0,"unit":"parameters","observationPeriod":"1997-11-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table 2\n\nhttp://www.bioinf.jku.at/publications/older/2604.pdf"},{"id":"epoch:211374ceef0c86afb27a:training-compute","entityId":"epoch:211374ceef0c86afb27a","metric":"training-compute","value":31512000000000.0,"unit":"FLOP","observationPeriod":"1997-11-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Due to limited computation time, training is stopped after 5 million sequence presentations\"\n\nEach sequence has p=100 elements in the long-delay setting.\n\nCOMPUTE = PRESENTATIONS * PRESENTATION LENGTH * UPDATE COMPUTE PER TOKEN\n\n5000000*100*6*10,504.0="},{"id":"epoch:a0cedc9029c38afab1ba:parameters","entityId":"epoch:a0cedc9029c38afab1ba","metric":"parameters","value":13000.0,"unit":"parameters","observationPeriod":"1997-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Page 7: \"The structures of all networks are adjusted so that\neach of them has about the same number of free parameters\n(approximately 13 000 here\""},{"id":"epoch:187c7e2cab24560d9a75:parameters","entityId":"epoch:187c7e2cab24560d9a75","metric":"parameters","value":8000.0,"unit":"parameters","observationPeriod":"1997-05-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The new chess chip had a completely redesigned evaluation function, going from around 6400 features to over 8000\""},{"id":"epoch:21520b6088b510882d85:parameters","entityId":"epoch:21520b6088b510882d85","metric":"parameters","value":32015.0,"unit":"parameters","observationPeriod":"1997-01-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"SOM: 125\nL1: 20*3*3*3=540\nL2: 25*3*3*20*0.3=1350\nL3: 25*5*6*40=30000\nTotal: 125+540+1350+30000=32,015\n(see Table 1)"},{"id":"epoch:21520b6088b510882d85:training-compute","entityId":"epoch:21520b6088b510882d85","metric":"training-compute","value":31431600000.0,"unit":"FLOP","observationPeriod":"1997-01-31","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Forward pass flop based on architecture (see Table 1): \nFirst CNN layer: 23*28 -> 3*3*20\n2*21*26*3*3*20=196560\nSecond CNN layer: 20*11*13 -> 3*3*25*0.3\n2*9*11*3*3*20*25*0.3=267300\nFC layer: 25*5*6 -> 40\n2*25*5*6*40=60000\nTotal forward FLOP: 196560+267300+60000=523860\nTotal training compute: 523860*3*20000=31431600000\nTraining steps: \"The network was trained with backpropagation [13] for a total of 20 000 updates.\""},{"id":"epoch:77a0def659decd1a5d2f:parameters","entityId":"epoch:77a0def659decd1a5d2f","metric":"parameters","value":6452.0,"unit":"parameters","observationPeriod":"1996-06-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"System 11 is a combination of Network 1 and Network 2\n\nNetwork 1 has 2095 connections and network 2 has 4357 connections (see table 1)"},{"id":"epoch:77a0def659decd1a5d2f:training-compute","entityId":"epoch:77a0def659decd1a5d2f","metric":"training-compute","value":25859616000.0,"unit":"FLOP","observationPeriod":"1996-06-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Since there is no parameter sharing, the forward compute is roughly twice that of the number of parameters. We use a 2:1 forward-backward ratio as this is a shallow network, with most connections in the first layer.\n\nNumber of passes (Section 2.1):\n* \"Nearly 1,050 face examples were gathered from face databases [...]\"\n* \"Fifteen face examples are generated for the training set from each original image\"\n\nTraining loop:\n1. \"initial set of nonface images by generating 1,000 random images\"\n2. Train (presumably on whole set)\n3. Run + collect false positives\n4. \"Select up to 250 of these subimages [...] and add them into the training set [...] Go to step 2\"\n\n\"A typical training run selects approximately 8,000 nonface images \"\n\nSelecting 8,000 nonface images implies 8000/250 = 32 loops.\n\nAssuming compute is 3 * N * D, we have\n* Loop 1: D = 15*1050 + 1000\n* Loop 2: D = 15*1050 + 1000 + 250\n* So on.\n\nHence D overall is 32*(15*1050 + 1000) + 250*32/2*(32+1) = 668,000.\n\nHence compute = 3 * 12904 * 668e3 = 25859616000"},{"id":"epoch:755c5fdfd3ca6233745d:parameters","entityId":"epoch:755c5fdfd3ca6233745d","metric":"parameters","value":13607.0,"unit":"parameters","observationPeriod":"1996-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"230*55+56*15+16*6+7*3=13607 (Figure 2)"},{"id":"epoch:755c5fdfd3ca6233745d:training-compute","entityId":"epoch:755c5fdfd3ca6233745d","metric":"training-compute","value":881733600000.0,"unit":"FLOP","observationPeriod":"1996-06-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2*13607*3*10800000=881733600000=8.8e11\nTraining steps: 400*27000=10800000\n\"After 400 epochs the error of the network\""},{"id":"epoch:44a0f953eb015ff3a6d7:parameters","entityId":"epoch:44a0f953eb015ff3a6d7","metric":"parameters","value":432800.0,"unit":"parameters","observationPeriod":"1995-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Total connections 32*32*20*20+20*20*48+20*20*10=432800\nInput: 32*32, Lissom: 20*20, Output: 10 (Figure 1a), up to 48 lateral connections per Lissom neuron (Figure 1b)\n"},{"id":"epoch:44a0f953eb015ff3a6d7:training-compute","entityId":"epoch:44a0f953eb015ff3a6d7","metric":"training-compute","value":195544800000.0,"unit":"FLOP","observationPeriod":"1995-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Lissom connections: 32*32*20*20+20*20*48=428800\nLissom compute: 2*428800*3*38*2000=195532800000=1.96e11\nPerceptron connections: 20*20*10=4000\nPerceptron compute: 2*4000*3*500*1700=20400000000=2e10\nTotal compute: 195532800000+12000000=195544800000=1.96e11\n\"LISSOM was trained with 2000 patterns\"\n\"The initial self-organizing map was formed in 8 epochs over the training set, gradually reducing the neighborhood radius from 20 to 8. The lateral connections were then added to the system, and over another 30 epochs,\"\n\"Of these, 1700 were used to train the perceptron layer, \"\n\"After the SOM and LISSOM maps were organized, a complete set of activation patterns on the two maps were collected. These patterns then formed the training input for the perceptron layer. Two separate versions were each trained for 500 epochs,\"\n\n"},{"id":"epoch:3c1fbde436524475b0ee:parameters","entityId":"epoch:3c1fbde436524475b0ee","metric":"parameters","value":100000000.0,"unit":"parameters","observationPeriod":"1995-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Section 6.2.2: \"...polynomials\nof degree 4 (that have more than 10^8 free parameters)...\"\nThey used 4-degree polynomials for MNIST"},{"id":"epoch:891926eaf183be55e05c:parameters","entityId":"epoch:891926eaf183be55e05c","metric":"parameters","value":4446.0,"unit":"parameters","observationPeriod":"1994-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Inputs are 12x12 pixels (Figure 3). They first train the architecture in Figure 2 a), then freeze it and train the additional layer in Figure 3 b).\n\nFigure 2 a): 2*(12*12*15) + 2*15 = 4,350\nFigure 2 b): 2*(12*12*15) + 2*15 + 2*(15*3) + 2*3 = 4,446"},{"id":"epoch:891926eaf183be55e05c:training-compute","entityId":"epoch:891926eaf183be55e05c","metric":"training-compute","value":80828280.0,"unit":"FLOP","observationPeriod":"1994-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Training 2a):  \"The learning took about 3000 iterations of steepest descent\" Assuming each iteration refers to a single image.\n6 * 4446 * 3000 = 80,028,000\n\nTraining 2b): \"While keeping the first layer of weights frozen, this network was trained using exactly the same cost function as the first layer for about 30 iterations using a gradient-based learning method.\"\n\n6 * 4446 * 30 = 800,280\n\nTotal: 80,828,280"},{"id":"epoch:ca7ef0609e34b56046bf:parameters","entityId":"epoch:ca7ef0609e34b56046bf","metric":"parameters","value":206910.0,"unit":"parameters","observationPeriod":"1994-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"5*80*430+430+430*80+80=206910\n\"P has nk input units and k output units. n is called the \"time-window size\"\n\"Note that the time-window was quite small (n = 5).\"\n\"alphabet consisted of k = 80 possible characters\"\n\"P had 430 hidden units\""},{"id":"epoch:ca7ef0609e34b56046bf:training-compute","entityId":"epoch:ca7ef0609e34b56046bf","metric":"training-compute","value":18621900000000.0,"unit":"FLOP","observationPeriod":"1994-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2*206910*3*15000000=18621900000000=1.86e13\n\"The training phase consisted of 25 sweeps through the training set\""},{"id":"epoch:060e813bb8c86c63c401:parameters","entityId":"epoch:060e813bb8c86c63c401","metric":"parameters","value":384000.0,"unit":"parameters","observationPeriod":"1994-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\u201cIn the example we describe, 7000 training images are sufficient to fit 384,000 parameters\u201c"},{"id":"epoch:060e813bb8c86c63c401:training-compute","entityId":"epoch:060e813bb8c86c63c401","metric":"training-compute","value":453600000000.0,"unit":"FLOP","observationPeriod":"1994-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"0.3*12*60*60*35000000=453600000000=4.54e11\nAssuming a utilization of 0.3 and interpreting \"overnight\" as 12 hours.\n\u201cthe training procedure is fast enough to do the fitting overnight on an R4400-based machine. \u201c\nR4400 has 35MFLOPS (\u201cCompare this to the 200MHz R4400 which is rated at about 35MFLOPS\u201d, http://www.sgidepot.co.uk/perf.html)"},{"id":"epoch:2a61ae1d032037139c31:parameters","entityId":"epoch:2a61ae1d032037139c31","metric":"parameters","value":72251.0,"unit":"parameters","observationPeriod":"1994-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Prior to learning an evaluation function, the model M (175 input, 165 hidden, and 175 output units)\" = 58,090 parameters\n\"NeuroChess then learns an evaluation network V (175 input units, 0 to 80 hidden units, and one output units).\" = 14,161 parameters\nTotal: 58,090 + 14,161 = 72,251"},{"id":"epoch:2a61ae1d032037139c31:training-compute","entityId":"epoch:2a61ae1d032037139c31","metric":"training-compute","value":858730812676.0,"unit":"FLOP","observationPeriod":"1994-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Lower bound: 0.3*2*24*60*60*1400000=72576000000=7.26e10\nUpper bound: 0.3*14*24*60*60*1400000*20=10160640000000=1.02e13\nGeometric mean: 858730812676=8.59e11 (speculative)\n\"Thus far, experiments lasted for 2 days to 2 weeks on I to 20 SUN Sparc Stations. \"\nSparcStation has 1.4 MFLOPS (https://ieeexplore.ieee.org/document/63671)\n"},{"id":"epoch:8d20fcf958be7cb5ced9:parameters","entityId":"epoch:8d20fcf958be7cb5ced9","metric":"parameters","value":2447124.0,"unit":"parameters","observationPeriod":"1994-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The total number of free parameters is then:\n(Nw - 1).NT + (NT - 1).NT.NT.\"\nWhere:\nNw= Vocabulary size\nNT = Number of tags\n\n\"In the treebank 159 different tags are used. These tags were projected on a smaller system of 76 tags designed by Evelyne Tzoukermann and Peter Brown (see Appendix). The results quoted in this paper all refer to this smaller system\"\nSo NT = 76\n\nhttps://www.aclweb.org/anthology/J94-2001/\n\nThere is no direct reference to Nw, but the data is from \"Lexicon and grammar in probabilistic tagging of written English.\" which says\n\n\"(the new CLAWS lexic\u00f3n has almost 26,500 entries)\"\nSo tentatively Nw=26500\n\nhttps://dl.acm.org/doi/10.3115/982023.982049"},{"id":"epoch:bcd42f3caed9b7938b57:parameters","entityId":"epoch:bcd42f3caed9b7938b57","metric":"parameters","value":1888.0,"unit":"parameters","observationPeriod":"1994-01-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Parameters: 100*16 + 16*16 + 16*2 = 1888\nArchitecture: \"The topology of the classifier was X-Y-Y-2, where X is the number of input components, Y is the number of neurons in each hidden layer and the number of neurons in the output layer is two, which is the number of classes. The two hidden layers were considered to have the same number of nodes for simplification purposes. \"\nInput size: \"Each pattern consists of a 10 x 10 pixel sub-image.\"\nHidden size: \"Experiments have been made on networks with 6, 9, 12 and 16 hidden nodes. \""},{"id":"epoch:bcd42f3caed9b7938b57:training-compute","entityId":"epoch:bcd42f3caed9b7938b57","metric":"training-compute","value":4531200000.0,"unit":"FLOP","observationPeriod":"1994-01-07","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Compute estimate: 2*1888*3*400000=4531200000=4.53e9\nTraining steps: \"In Fig. 6 we report the classification results obtained on the testing set in the 12 and 16 component compressed data after 400000 training iterations\""},{"id":"epoch:bff8be13eb870434e39e:parameters","entityId":"epoch:bff8be13eb870434e39e","metric":"parameters","value":5620.0,"unit":"parameters","observationPeriod":"1993-11-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"15*15*20+20+50*10*2+100=5620\nHidden layer is split, 15*15 image input, 2*50 output neurons (see Figure 2)\nHidden size up to 20 neurons (\"This architecture was used with varying numbers of hidden units in the single, divided, hidden layer; experiments with 10, 16 and 20 hidden units were performed. \")"},{"id":"epoch:bff8be13eb870434e39e:training-compute","entityId":"epoch:bff8be13eb870434e39e","metric":"training-compute","value":17534400000.0,"unit":"FLOP","observationPeriod":"1993-11-29","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2*5620*3*520000=17534400000\nTraining examples: 2000*260=520000\n\"As mentioned before, 2000 image/position pairs were gathered for training\"\n\"All of the networks described in this paper are trained with the same parameters for 260 epochs\""},{"id":"epoch:9cc11e2348b13cb3ed97:parameters","entityId":"epoch:9cc11e2348b13cb3ed97","metric":"parameters","value":744.0,"unit":"parameters","observationPeriod":"1993-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The input is 8 by 200 units, the first convolutional layer is 6 by 192 units with each unit's receptive field covering 8 by 9 units of the input. The first averaging layer is 6 by 64 units, the second convolution layer is 4 by 57 with 6 by 8 receptive fields and the second averaging layer is 4 by 19\"\n\"Two separate sub-networks based on Time Delay Neural Networks (Lang and Hinton, 1988, Guyon et al. 1990) act on each input pattern to extract features,\"\n\"All weights could be learnt, but the two sub-networks were constrained to have identical weights.\"\nL1: H=1, W=200, C=8, K=9, D=6\nL2: H=1, W=64, C=6, K=8, D=4\nParameters:  7*9*8+5*8*6=744"},{"id":"epoch:9cc11e2348b13cb3ed97:training-compute","entityId":"epoch:9cc11e2348b13cb3ed97","metric":"training-compute","value":12869570138112.0,"unit":"FLOP","observationPeriod":"1993-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"8073216*3*7701*69=12869570138112=1.29e13\nForward pass flop: 2*(2*200*200*8*6+2*64*64*6*4)=8073216\n\"We used up to 7,701 signature pairs\"\nEpochs: 69 (Table 1)\n"},{"id":"epoch:b713e06e9ac62def7487:parameters","entityId":"epoch:b713e06e9ac62def7487","metric":"parameters","value":1658364.0,"unit":"parameters","observationPeriod":"1993-06-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The model is initiallized with 2.44E+09 translation probabilities, which are progressively culled until 1,658,364 remain. There are other parameters in the models (eg the fertility probabilities that relate each word in the input to the number of words it will align to) but the parameter count is dominated by the translation probabilities."},{"id":"epoch:b7d58b9b499bfd12472d:training-compute","entityId":"epoch:b7d58b9b499bfd12472d","metric":"training-compute","value":5292000000.0,"unit":"FLOP","observationPeriod":"1993-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"0.3*35*60*60*140000=5292000000=5.29e9\n\u201cTraining on the data from 1979-1985 took approximately 35 hours\u201d\n\u201cThe neural networks were trained on a 386Dx-25mhz processor\u201d\n0.14 MFLOPS taken from this benchmark: https://www.vogons.org/viewtopic.php?t=46350"},{"id":"epoch:e08af31f3aaf18db257a:parameters","entityId":"epoch:e08af31f3aaf18db257a","metric":"parameters","value":2578.0,"unit":"parameters","observationPeriod":"1992-11-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\u201cThe network has 4645 neurons, 2578 different weights, and 98442 connections.\u201c"},{"id":"epoch:b03355f52e905bd2b7d4:parameters","entityId":"epoch:b03355f52e905bd2b7d4","metric":"parameters","value":594.0,"unit":"parameters","observationPeriod":"1992-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\u201cThe network shown has 60 input PEs, one for each cell line, and 6 output PEs\u201c\n\u201cNeural networks with three to nine hidden layer PEs used\u201d\n9*60 + 6*9 = 594"},{"id":"epoch:b03355f52e905bd2b7d4:training-compute","entityId":"epoch:b03355f52e905bd2b7d4","metric":"training-compute","value":53460000.0,"unit":"FLOP","observationPeriod":"1992-10-16","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2*594*3*15000=53460000=5.35e7\n\u201cThe extent of training was 15,000 presentations\u201c"},{"id":"epoch:8440eba9676a861610ee:parameters","entityId":"epoch:8440eba9676a861610ee","metric":"parameters","value":1166.0,"unit":"parameters","observationPeriod":"1992-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Table II: \"he neural network has three hidden layers, with m hidden nodes in each layer\", m = 20, input dim. = 9, output dim. = 6\n\n9*20+20*20+20*20+6*20+66=1166"},{"id":"epoch:8440eba9676a861610ee:training-compute","entityId":"epoch:8440eba9676a861610ee","metric":"training-compute","value":1403117760.0,"unit":"FLOP","observationPeriod":"1992-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1166 params * 2 FLOP/param * (3 for forward + backward pass) * 460 epochs * 436 examples"},{"id":"epoch:004650086accdc88bd44:parameters","entityId":"epoch:004650086accdc88bd44","metric":"parameters","value":25000.0,"unit":"parameters","observationPeriod":"1992-05-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The best performance was obtained with a network containing 80 hidden units and over 25,000 weights.\""},{"id":"epoch:004650086accdc88bd44:training-compute","entityId":"epoch:004650086accdc88bd44","metric":"training-compute","value":18232157622832.703,"unit":"FLOP","observationPeriod":"1992-05-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Extracted from AI and Compute (https://openai.com/blog/ai-and-compute/) charts by using https://automeris.io/WebPlotDigitizer/.\n\nOpenAI estimate: 1.8e13\n\nHardware estimate (likely overestimates due to simulation effort)\n\"on an IBM RS/6000 workstation, the smallest network was trained in several hours, while the largest net required two weeks of simulation time.\"\nIBM RS/6000 achieves 1.5 GFLOPS on Linpack (https://link.springer.com/rwe/10.1007/978-0-387-09766-4_232) \n14*24*60*60*0.5*1500000000=9.072e+14\n\nOperation counting estimate: \nForward FLOP: 50000\nEach legal move had to be evaluated separately, assuming an average of 10 move options (+2 for backward passes of the chosen move):\n50000*12*300000*21=3.78e+12\n\nKeeping the OpenAI estimate as the median estimate. "},{"id":"epoch:c80c181f82de0f9f281d:parameters","entityId":"epoch:c80c181f82de0f9f281d","metric":"parameters","value":8386.0,"unit":"parameters","observationPeriod":"1991-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" 7*26*40+40+40*26+26=8386\n\"The network had 7 x 26 input units, 40 hidden units and 26 output units\""},{"id":"epoch:c80c181f82de0f9f281d:training-compute","entityId":"epoch:c80c181f82de0f9f281d","metric":"training-compute","value":75474000000.0,"unit":"FLOP","observationPeriod":"1991-12-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2*8386*3*1500000=75474000000=7.55e10\n\"It was trained on 400 to 5000 random words from the data base of around 20.000 words,\"\n\"The top full line corresponds to the generalization error after 300 epochs\"\n"},{"id":"epoch:797482f5033ebe004c88:parameters","entityId":"epoch:797482f5033ebe004c88","metric":"parameters","value":429409.0,"unit":"parameters","observationPeriod":"1991-06-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"From https://www.bnlearn.com/bnrepository/"},{"id":"epoch:5f40c00fa9860182cb48:parameters","entityId":"epoch:5f40c00fa9860182cb48","metric":"parameters","value":1536.0,"unit":"parameters","observationPeriod":"1990-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Largest model:\n\"A 48-16-48 RAAM learned to construct representations \"\nParameters 48*16*2 = 1536"},{"id":"epoch:5ad4866c717d55635942:parameters","entityId":"epoch:5ad4866c717d55635942","metric":"parameters","value":72940.0,"unit":"parameters","observationPeriod":"1990-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"900*40*2+40+900=72940\n\u201cImages sampled at 30x30 were compressed using a 900x40x900 fully connected back-propagation network\u201d"},{"id":"epoch:5ad4866c717d55635942:training-compute","entityId":"epoch:5ad4866c717d55635942","metric":"training-compute","value":78775200000.0,"unit":"FLOP","observationPeriod":"1990-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2*72940*3*90*2000=78775200000\n\u201cThe compression network trained for 2000 runs on each of 90 faces\u201d"},{"id":"epoch:7905e02b99ce5ed781b9:parameters","entityId":"epoch:7905e02b99ce5ed781b9","metric":"parameters","value":1640.0,"unit":"parameters","observationPeriod":"1990-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Largest classification model: 40*40 + 40=1640 (Figure 2)"},{"id":"epoch:67d32be65acbaff531f9:parameters","entityId":"epoch:67d32be65acbaff531f9","metric":"parameters","value":36.0,"unit":"parameters","observationPeriod":"1990-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The input layer consisted of the five nodes, one for each of the ratios. The hidden layer consisted of 5 node.;. The output layer consisted of only one neuron\"\n30 weights + 6 biases = 36"},{"id":"epoch:67d32be65acbaff531f9:training-compute","entityId":"epoch:67d32be65acbaff531f9","metric":"training-compute","value":3059337600.0,"unit":"FLOP","observationPeriod":"1990-06-17","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" 2*36*3*74*191400=3,059,337,600=3.06e9\n\"Convergence was reached after 191,400 iterations\""},{"id":"epoch:1b561e8760a5e49192b3:parameters","entityId":"epoch:1b561e8760a5e49192b3","metric":"parameters","value":27480.0,"unit":"parameters","observationPeriod":"1990-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"203*120+120*26=27480\n\u201c203 input units, 120 hidden units, and 26 output units\u201d"},{"id":"epoch:1b561e8760a5e49192b3:training-compute","entityId":"epoch:1b561e8760a5e49192b3","metric":"training-compute","value":35811936000.0,"unit":"FLOP","observationPeriod":"1990-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Updated FLOP estimate: 2*27480*3*1000*7.24*30=35811936000=3.6e10"},{"id":"epoch:9a39c89883e76394d93d:parameters","entityId":"epoch:9a39c89883e76394d93d","metric":"parameters","value":9760.0,"unit":"parameters","observationPeriod":"1989-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In summary, the network has 1256 units, 64,660 connections, and 9760 independent parameters\""},{"id":"epoch:9a39c89883e76394d93d:training-compute","entityId":"epoch:9a39c89883e76394d93d","metric":"training-compute","value":1496338054440.0,"unit":"FLOP","observationPeriod":"1989-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Its a deep CNN so we assume a backward-forward ratio of 2:1\n 2*64660*3*23*167693=1496338054440\n\"The network was trained for 23\npasses through the training set (167,693 pattern presentations).\""},{"id":"epoch:65d8dca66553d5290371:parameters","entityId":"epoch:65d8dca66553d5290371","metric":"parameters","value":10.0,"unit":"parameters","observationPeriod":"1989-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Each net has 5 units"},{"id":"epoch:65d8dca66553d5290371:training-compute","entityId":"epoch:65d8dca66553d5290371","metric":"training-compute","value":120000000.0,"unit":"FLOP","observationPeriod":"1989-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"10 params * 6 FLOP/param/pass * 4 datapoints * 1000 epochs * 50 individuals * 10 generations"},{"id":"epoch:c0635471f7c1ddeef3c6:parameters","entityId":"epoch:c0635471f7c1ddeef3c6","metric":"parameters","value":36627.0,"unit":"parameters","observationPeriod":"1989-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1217*29 + 29*46 =36627\n\u201cEach of these 1217 input units is fully connected to the hidden layer of 29 units, which is in turn fully connected to the output layer. The output layer consists of 46 units, divided into two groups.\u201d"},{"id":"epoch:c0635471f7c1ddeef3c6:training-compute","entityId":"epoch:c0635471f7c1ddeef3c6","metric":"training-compute","value":10548576000.0,"unit":"FLOP","observationPeriod":"1989-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2 * 36627 * 3 * 40 * 1200 = 10548576000 = 1.05e10\n36627 parameters\n\"After 40 epochs of training on the 1200 simulated road snapshots\""},{"id":"epoch:8707c0cc6281f0b52078:parameters","entityId":"epoch:8707c0cc6281f0b52078","metric":"parameters","value":3040.0,"unit":"parameters","observationPeriod":"1989-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\u201cThe MLP consisted of 64 inputs (the DFf coefficients. each nonnalized between zero and one), a single hidden layer of 40 units, and 12 output units;\u201d"},{"id":"epoch:8707c0cc6281f0b52078:training-compute","entityId":"epoch:8707c0cc6281f0b52078","metric":"training-compute","value":7485696000.0,"unit":"FLOP","observationPeriod":"1989-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2*3040*3*410400=7485696000=7.49e9\n\u201cThe network was trained on 100 iterations through the 4104 training vectors.\u201d"},{"id":"epoch:dba570bb969018017a7f:parameters","entityId":"epoch:dba570bb969018017a7f","metric":"parameters","value":2578.0,"unit":"parameters","observationPeriod":"1989-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"In summary, the network has 4635 units, 98442 connections, and 2578 independent parameters.\u201c"},{"id":"epoch:dba570bb969018017a7f:training-compute","entityId":"epoch:dba570bb969018017a7f","metric":"training-compute","value":181440000000.0,"unit":"FLOP","observationPeriod":"1989-11-27","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1.4e6 * 3 * 24 * 60* 60 * 0.5 = 181440000000 = 1.81e11\n\"A complete training session (30 passes through the training set plus test) takes about 3 days on a SUN SPARCstation 1\"\nSparcstation 1 has an estimated compute of 1.4 MFLOPS (source: https://ieeexplore.ieee.org/document/63671 )"},{"id":"epoch:609b82199f6044a36feb:training-compute","entityId":"epoch:609b82199f6044a36feb","metric":"training-compute","value":27000000000.0,"unit":"FLOP","observationPeriod":"1989-06-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"0.5*6*60*60*2.5e6 = 27000000000 = 2.7e10\nTrained for 6h on a SUN-4 (section 4)\nAssumed utilization of 0.5\nSUN-4 is estimated at 2.5e6 FLOP/s (Nordhaus, 2007)"},{"id":"epoch:9d36b07a6442848db515:parameters","entityId":"epoch:9d36b07a6442848db515","metric":"parameters","value":805.0,"unit":"parameters","observationPeriod":"1989-06-18","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"6*25+25+8*45+45*6=805 (see Figure 6)"},{"id":"epoch:7312d9af3bb184294bc5:parameters","entityId":"epoch:7312d9af3bb184294bc5","metric":"parameters","value":10000.0,"unit":"parameters","observationPeriod":"1988-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\u201cFor an MLN of about 10,000 links, the time was 115 CPU msecs for the recognition of a spoken letter and 317 msecs for the learning of a spoken letter on the SUN 4/280. A 20% reduction was obtained on the VAX 8650\u201d"},{"id":"epoch:7312d9af3bb184294bc5:training-compute","entityId":"epoch:7312d9af3bb184294bc5","metric":"training-compute","value":296425000.0,"unit":"FLOP","observationPeriod":"1988-08-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\u201cFor an MLN of about 10,000 links, the time was 115 CPU msecs for the recognition of a spoken letter and 317 msecs for the learning of a spoken letter on the SUN 4/280. A 20% reduction was obtained on the VAX 8650\u201d, \u201cLearning and recognition were performed on a VAX 8650.\u201d\nDataset: 70*10*2=1400 (Train) 10*10*2=200 (Test)\n\u201cThe ten words of the El set were pronounced twice by 80 speakers (40 males and 40 females)\u201d\n\u201cThe data from 70 speakers were used as a training set while the data from the remaining 10 speakers (6 males and 4 females) were used for the test\u201d\nVAX 8650 FLOPS  = 1.67E+06 (Nordhaus)\nTraining time: 317ms * 1400  * 0.8 = 355040ms = 355s\nEstimate: 0.5 * 1.67e6 * 355 = 296425000 = 2.96e8"},{"id":"epoch:782ff82c0796bd1fb3c3:parameters","entityId":"epoch:782ff82c0796bd1fb3c3","metric":"parameters","value":110.0,"unit":"parameters","observationPeriod":"1988-07-24","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 3"},{"id":"epoch:64abc2501834e9055a8e:parameters","entityId":"epoch:64abc2501834e9055a8e","metric":"parameters","value":816.0,"unit":"parameters","observationPeriod":"1987-06-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Network: 12-60-6-16\nWeights: 6*60+60*6+6*16=816\nLayer 2 was only sparsely connected to input layer"},{"id":"epoch:64abc2501834e9055a8e:training-compute","entityId":"epoch:64abc2501834e9055a8e","metric":"training-compute","value":18032947200.0,"unit":"FLOP","observationPeriod":"1987-06-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"FLOPs: 2*816*3*160*23020=18032947200=1.8e10"},{"id":"epoch:fb0943015c7d6dfe7e8e:parameters","entityId":"epoch:fb0943015c7d6dfe7e8e","metric":"parameters","value":18629.0,"unit":"parameters","observationPeriod":"1987-06-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The connections in the network are specified by a total of 18629\nweight parameters (including a variable threshold for each unit)\""},{"id":"epoch:fb0943015c7d6dfe7e8e:training-compute","entityId":"epoch:fb0943015c7d6dfe7e8e","metric":"training-compute","value":28328002560.0,"unit":"FLOP","observationPeriod":"1987-06-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"18629 params * 2 FLOP/param * (3 for forward + backward pass) * 55 epochs * 1024 words/epoch * 4.5 letters/word"},{"id":"epoch:924906810a4a0d2bf057:parameters","entityId":"epoch:924906810a4a0d2bf057","metric":"parameters","value":18629.0,"unit":"parameters","observationPeriod":"1987-06-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The connections in the network are specified by a total of 18629 weight parameters (including a variable threshold for each unit)\""},{"id":"epoch:924906810a4a0d2bf057:training-compute","entityId":"epoch:924906810a4a0d2bf057","metric":"training-compute","value":27664065000.0,"unit":"FLOP","observationPeriod":"1987-06-06","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"18629 params * 2 FLOP/param * (3 for forward + backward pass) * 55 epochs * 1000 words/epoch * 4.5 letters/word (estimated number of letters)"},{"id":"epoch:c0541c786a6c6ad8712c:parameters","entityId":"epoch:c0541c786a6c6ad8712c","metric":"parameters","value":720.0,"unit":"parameters","observationPeriod":"1986-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Architecture in Figure 3: \n\n24+12 input -> 6 + 6 hidden -> 12 hidden -> 6 hidden -> 24 output\n\nParameters: 6*24+6*12+12*12+6*12+24*12=720"},{"id":"epoch:c0541c786a6c6ad8712c:training-compute","entityId":"epoch:c0541c786a6c6ad8712c","metric":"training-compute","value":673920000.0,"unit":"FLOP","observationPeriod":"1986-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"We assume that the number of mult-adds per pass is equal to the number of parameters -> 2*720=1440 FLOP per forward pass.\n\n\"We trained the network for 1500 sweeps\"\nThere are 104 relationship triplets (\"[...] of the 104 possible triplets\")\n\nFLOP: 1500*104*3*1440=673920000\n\n"},{"id":"epoch:ef7ebcb1f1b3e315520b:parameters","entityId":"epoch:ef7ebcb1f1b3e315520b","metric":"parameters","value":432.0,"unit":"parameters","observationPeriod":"1986-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Parameters: 24*6 + 12*6 + 12*12 + 12*6 =432\n\"Figure 5: The activity levels in a five-layer network after it has learned. The bottom layer has 24 input units on the left for representing person 1 and 12 units on the right for representing the relationship. The white squares inside these two groups show the activity levels of the units. There is one active unit in the first group (representing Colin) and one in the second group (representing has-aunt). Each of the two groups of input units is totally connected to its own group of 6 units in the second layer. These two groups of 6 must learn to encode the input terms as distributed patterns of activity. The second layer is totally connected to the central layer of 12 units, and this layer is connected to the penultimate layer of 6 units.\"\n"},{"id":"epoch:ef7ebcb1f1b3e315520b:training-compute","entityId":"epoch:ef7ebcb1f1b3e315520b","metric":"training-compute","value":388800000.0,"unit":"FLOP","observationPeriod":"1986-08-15","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":" 2*432*3*1500*100=388800000=3.9e8\n\"After 1500 sweeps through all 100 training examples the weights were very stable \""},{"id":"epoch:ccce8cce94377bce7b35:parameters","entityId":"epoch:ccce8cce94377bce7b35","metric":"parameters","value":211600.0,"unit":"parameters","observationPeriod":"1986-01-03","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Source: https://files.eric.ed.gov/fulltext/ED267419.pdf\np.9: network architecture is given, with two layers of hidden units. The hidden units are  called \u201cWickelfeature representation\u201d. The \u201cmodifiable connections\u201d are only between the hidden units. p.19: \u201cAll in all then, we used only 460 of the 1,210 possible Wickelfeatures. Using this representation, a verb is represented by a pattern of activation over a set of 460 Wickelfeature units.\""},{"id":"epoch:4947f20ed9ccd473e21f:parameters","entityId":"epoch:4947f20ed9ccd473e21f","metric":"parameters","value":9315.0,"unit":"parameters","observationPeriod":"1984-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Parameters 5*5*9*3*3 + 3*3*9*3*3*9 + 9*3*3*9 = 9315\n\"The numbers of excitatory cells in these four layers were: 7x7 in U0, 5x5 in  U1, 3x3 in U2, and 9 in U3\"\n\"Each feature-extracting cell in layer U1 receives excitatory modifiable afferent connections from 3x3 cells in layer U0\"\n\"On the other hand, each feature, each extracting cell in layers U2 and U3 receives excitatory modifiable connections from all 9 cells in each of the 3 x 3 hypercolumns in the layer preceding it. Therefore, it receives 3 x 3 x 9 afferent excitatory modifiable connections altogether\"\n"},{"id":"epoch:0562abd5b4b7a24b9d83:parameters","entityId":"epoch:0562abd5b4b7a24b9d83","metric":"parameters","value":324.0,"unit":"parameters","observationPeriod":"1983-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The system consists of two parts: ACE and ASE, each with 162 weights (=324 parameters). Found in Figures 2 and 3."},{"id":"epoch:0562abd5b4b7a24b9d83:training-compute","entityId":"epoch:0562abd5b4b7a24b9d83","metric":"training-compute","value":324000000.0,"unit":"FLOP","observationPeriod":"1983-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"324 * 2 * 500000 = 324000000 = 3.24e8. The calculation assumes \"compute per forward pass\" = \"number of parameters\" = \"compute per backward pass\". Their model only has a single layer and is trained with simple update rules instead of gradient descent. Training details are described in Section IX.\n\nNote that this is the compute for a single run; they appear to have repeated training 10 times for the ASE+ACE system."},{"id":"epoch:fd9f0835e832454adb41:parameters","entityId":"epoch:fd9f0835e832454adb41","metric":"parameters","value":9900.0,"unit":"parameters","observationPeriod":"1982-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"My understanding is that the biggest Hopfield networks they studied had N=100 units. \n\nEach unit has 99 synapses Tij from each other unit, for a total of 100*99 parameters"},{"id":"epoch:0bbdca0d2cda04e7db8c:parameters","entityId":"epoch:0bbdca0d2cda04e7db8c","metric":"parameters","value":4096.0,"unit":"parameters","observationPeriod":"1981-07-25","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The input vectors are 3D.\nI could not find the grid size, but from the images it looks 8x8.\nSo the network was 8x8x3 parameters."},{"id":"epoch:67afef1fdae327096f0a:parameters","entityId":"epoch:67afef1fdae327096f0a","metric":"parameters","value":1140576.0,"unit":"parameters","observationPeriod":"1980-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The synaptic connections from S-layers to C-layers\nare fixed and unmodifiable. [...]\nThe numbers of excitatory cells in these seven layers are: 16x16 in U0, 16x16x24 in Us1, 10x10x 24 in Uc1, 8x8x24 in Us2, 6x 6x 24 in Uc2, 2x2x24 in Us3, and 24 in Uc3 \n[...]\n the number of input synapses to each S-cell is 5 x 5 in layer Us1 and 5x5x24 in layers Us2 and Us3\n[...]\nThe number of excitatory input synapses to each C-cell is 5x5 in layers Uc1 and Uc2, and is 2x2 in\nlayer Uc3\n\"\n\nThe number of synapses into each S-layer is:\n\nS1: (16*16*24)*(5*5) \nS2: (8*8*24)*(5*5*24)\nS3: (2*2*24)*(5*5*24)\n\nWe assume one parameter a per synapse into each cell in a S-layer, and one parameter b per each cell in a S-layer."},{"id":"epoch:67afef1fdae327096f0a:training-compute","entityId":"epoch:67afef1fdae327096f0a","metric":"training-compute","value":273738240.0,"unit":"FLOP","observationPeriod":"1980-04-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"It does not necessarily mean that all of these input synapses are\nalways fully reinforced. In usual situations, only some of these input synapses are reinforced, and the rest of them remains in small values [...] Each of the five stimulus patterns has been presented 20 times to the network. By that time, self organization of the network has almost been completed.\"\n\nWe multiply by 2 to account for multadds in the forward pass. \nThere is no real backward pass, weights are only updated sparsely. Estimating 20% additional weight update compute compared to the forward pass:\n2*1,140,576.0*1.2*100=273738240\n"},{"id":"epoch:d004b0544716aded6412:parameters","entityId":"epoch:d004b0544716aded6412","metric":"parameters","value":21600.0,"unit":"parameters","observationPeriod":"1975-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4 layers, 288 neurons per layer, weights connect each neuron to only 25 neurons in the previous layer.\nOnly 3 layers with learnable weights, the first layer is just an input representation (see Fig 5)\n3*288*25 parameters"},{"id":"epoch:d004b0544716aded6412:training-compute","entityId":"epoch:d004b0544716aded6412","metric":"training-compute","value":5184000.0,"unit":"FLOP","observationPeriod":"1975-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"No real backward pass as weights are sparsely updated. Assuming 20% additional compute for the weight update.\nTotal compute estimate: 100*2*3*288*25*1.2 = 5184000\n"},{"id":"epoch:4e549b3bc7011b2d061d:parameters","entityId":"epoch:4e549b3bc7011b2d061d","metric":"parameters","value":357.0,"unit":"parameters","observationPeriod":"1973-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"16 input features + bias = 17 input features\n7*6/2 = 21 Hyperplanes\n17*21 = 357 parameters\n\"For the multicategory problem involving NR categories, a total of NR(NR - 1)/2 hyperplanes are used to partition the pattern space.\"\n\"The input variables to the classifier consisted of the mean variance of the four textural features (f1,f2,f3, andfg obtained from the distance 1 gray-tone spatial-dependence matrices) and eight spectral features (comprised of the mean variance of the image gray-tone values) in each of the four spectral bands\"\n"},{"id":"epoch:16510c4adf0790c35ed6:parameters","entityId":"epoch:16510c4adf0790c35ed6","metric":"parameters","value":21.0,"unit":"parameters","observationPeriod":"1973-09-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Fig. 1 shows that there is a bias term, while Fig. 5 shows that the input is a sequence of 20 bits, corresponding to 20 weights. So the total number of parameters is 21."},{"id":"epoch:ca72c0b4034031410d0a:parameters","entityId":"epoch:ca72c0b4034031410d0a","metric":"parameters","value":2450.0,"unit":"parameters","observationPeriod":"1969-05-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"5 adaline were trained on binary decisions (p. 1)\nEach adaline had up to 490 input weights (\u201cmeshes\u201d)\nTotal parameters = 5*490=2450\n"},{"id":"epoch:56f4804787076b008f05:parameters","entityId":"epoch:56f4804787076b008f05","metric":"parameters","value":40.0,"unit":"parameters","observationPeriod":"1967-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The total number of parameters used at any one time has been varied from a very few to as many as 40\""},{"id":"epoch:2b8215433df0ed880d33:parameters","entityId":"epoch:2b8215433df0ed880d33","metric":"parameters","value":2061.0,"unit":"parameters","observationPeriod":"1966-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"2 connected systems, 1st level LTE and 2nd level LTE.\n1st Level: 1810 parameters (\"Thus, every 20 msec after the beginning of the utterance, the 15 filter amplitudes were each represented by a 12-bit code, resulting in a 180-bit time sample of the spectrum for that interval. Each time sample was fed to the first-level LTE's, which reduced it to a 10-bit code\")\n2nd Level: 251 parameters (\"This resulted in a 250-bit input pattern to the second level for the first half-second of each utterance. Each 250-bit pattern was then classified by the LTE into one of two classes\")"},{"id":"epoch:2b8215433df0ed880d33:training-compute","entityId":"epoch:2b8215433df0ed880d33","metric":"training-compute","value":105917060.0,"unit":"FLOP","observationPeriod":"1966-11-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"1st and 2nd level system are trained separately, multiple versions of both are trained, I chose the largest clearly described training runs.\n\n1st level LTE compute: 2*1810*28700=103894000=1.04e8\n1st level steps: 28700 (\"Only 287 samples were selected to train the 10 LTE's. The same algorithm was used as that used with the 100-class gain. Two LTE's\nconverged before 100 training passes.\")\n\n2nd level LTE compute: 2*251*4030=2023060=2e6\n2nd level steps: 4030 (31 epochs, 130 training examples, see Table 3)\n\nTotal compute: 103894000+2023060=105917060=1.06e8 (assuming no backward pass since they didn't use backpropagation)\n"},{"id":"epoch:f40738700c0e6671950f:training-compute","entityId":"epoch:f40738700c0e6671950f","metric":"training-compute","value":1080000.0,"unit":"FLOP","observationPeriod":"1965-10-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 10 shows their largest system is trained for 3h and was trained on an analog IBM 1620 that was simulated on a digital IBM 1710.\nNordhaus, 2007 lists the IBM 1620 at 200 multiplications per second and doesn\u2019t contain the 1710\nFlops estimate: 0.5 * 3 * 60 * 60 * 200 = 1080000 = 1.08e6\nAssumed utilization of 0.5\n"},{"id":"epoch:c4c7f32ac73a77579bf6:training-compute","entityId":"epoch:c4c7f32ac73a77579bf6","metric":"training-compute","value":22500000.0,"unit":"FLOP","observationPeriod":"1963-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"0.5*2.5*60*60*5000 = 22500000 = 2.25e7\nAssumed utilization of 0.5\nTrained for 2-3h on an IBM 7090 (from Introduction)\nEstimated IBM 7090 at 5000 FLOP/s based on multiplications per second (Nordhaus, 2007)\nNote: the Nordhaus estimate is very different from Wikipedia's estimate of 100000 FLOP/s, which cites a PowerPoint as source."},{"id":"epoch:f741ecfda925e91e5042:training-compute","entityId":"epoch:f741ecfda925e91e5042","metric":"training-compute","value":1559250.0,"unit":"FLOP","observationPeriod":"1962-06-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"0.5*45*35*1980 = 1559250 = 1.56e6\nTrained using IBM punched cards, computation took 45 * 35s for all 10 digits (Section Estimating the Linear Decision Function).\nMultiplications per second estimate based on publication year: 1.98e3 (regression on Nordhaus data).\nAssumed utilization of 0.5"},{"id":"epoch:b21857b0cb92bd0b1f36:parameters","entityId":"epoch:b21857b0cb92bd0b1f36","metric":"parameters","value":17.0,"unit":"parameters","observationPeriod":"1960-06-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The machine's total experience is stored in the values of the weights a0,...,a16\""},{"id":"epoch:b21857b0cb92bd0b1f36:training-compute","entityId":"epoch:b21857b0cb92bd0b1f36","metric":"training-compute","value":6600.0,"unit":"FLOP","observationPeriod":"1960-06-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"The method of searching that has proven most useful is the method of steepest descent\"\n\nApparently each pattern was only shown once to the system.\n\nSo the training compute is (forward pass compute) * (3 for backprop) * dataset size\n\nThis is a single layer (and single neuron) which does not require gradients w.r.t. inputs - 1:1 forward-backward ratio\n\n"},{"id":"epoch:37a4a6ddc0e3e70ddf92:parameters","entityId":"epoch:37a4a6ddc0e3e70ddf92","metric":"parameters","value":1000.0,"unit":"parameters","observationPeriod":"1960-03-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\" The first program was designed to handle\nup to 1000 A units, and a 72 by 72 sensory mosaic. It\nwas found that this large sensory system presented\nstimuli with a fineness of grain considerably better than\nthe limits of discrimination of a thousand-unit percep-\ntron, and at the same time, required an excessive\namount of time for stimulus transformations, since each\nilluminated point in the stimulus must be transformed\nindividually into its image point.\""},{"id":"epoch:37a4a6ddc0e3e70ddf92:training-compute","entityId":"epoch:37a4a6ddc0e3e70ddf92","metric":"training-compute","value":720000000.0,"unit":"FLOP","observationPeriod":"1960-03-30","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"4000 * 12000 * 15\nfrom the text \"This program uses the IBM 704 computer to simulate per-ceptual learning, recognition, and spontaneous classification of visual stimuli in the perceptron,\"\nfrom https://en.wikipedia.org/wiki/IBM_704 The 704 can execute up to 12,000 floating-point additions per second.\n\" For the first system, the computing time averaged about 15 seconds per stimulus cycle, \"\nIn Fig 10 we see up to 4000 stimuli"},{"id":"epoch:d099f764d70520a28f0e:parameters","entityId":"epoch:d099f764d70520a28f0e","metric":"parameters","value":2625.0,"unit":"parameters","observationPeriod":"1959-12-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"A two bit state is recorded for each of the 75 cell pairs and each of the 25+10 characters recognized."},{"id":"epoch:9ce1cf4fb91a61d99b49:parameters","entityId":"epoch:9ce1cf4fb91a61d99b49","metric":"parameters","value":16.0,"unit":"parameters","observationPeriod":"1959-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"with 16 terms for generalization learning\"\n\n\"Mention has been made several times of the procedure\nfor replacing terms in the scoring polynomial. The program, as it is currently running, contains 38 different\nterms (in addition to the piece-advantage term), 16 of\nthese being included in the scoring polynomial at anyone\ntime and the remaining 22 being kept in reserve.\""},{"id":"epoch:9ce1cf4fb91a61d99b49:training-compute","entityId":"epoch:9ce1cf4fb91a61d99b49","metric":"training-compute","value":428400000.0,"unit":"FLOP","observationPeriod":"1959-07-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"it can learn to do this in a remarkably short period of time 8 or 10 hours of machine-playing time)\"\n\n\"The availability of a larger and faster machine (the IBM 704), coupled with many detailed changes in the programming procedure, leads to a fairly interesting game being played, even without any learning.\"\n\n\"The Type 704 is the first large-scale, commercially available computer to employ fully automatic floating point arithmetic commands. [...]. Floating point addition or subtraction operations require 84 microseconds.\"\n\nsource: https://www.ibm.com/ibm/history/exhibits/mainframe/mainframe_PP704.html\n\n\"An idea of the learning ability of this procedure can be gained by analyzing an initial test series of 28 games\"\n\n\"Each game averaged 68 moves (34 to a side), of which approximately 20 caused changes to be made in the scoring polynomial.\"\n\n10*3600*1000000/84=428571428"},{"id":"epoch:e7312cc1eea097a2f5a3:training-compute","entityId":"epoch:e7312cc1eea097a2f5a3","metric":"training-compute","value":600000000.0,"unit":"FLOP","observationPeriod":"1959-02-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The paper mentions using an IBM 704, which can execute up to 12,000 floating-point additions per second (https://wikiless.org/wiki/IBM_704). My best guess as to how long it ran for ranges between 1h to 2 days, which when plugged into guesstimate (https://www.getguesstimate.com/models/19625), i.e., taking the log mean, gives a mean estimate of 600M"},{"id":"epoch:787b11189d7462d799b8:parameters","entityId":"epoch:787b11189d7462d799b8","metric":"parameters","value":1000.0,"unit":"parameters","observationPeriod":"1957-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"\"Figure 4.8 Illustration of the Mark 1 perceptron hardware. The photograph on the left shows how the inputs were obtained using a simple camera system in which an input scene, in this case a printed character, was illuminated by powerful lights, and an image focussed onto a 20 \u00d7 20 array of cadmium sulphide photocells, giving a primitive 400 pixel image. The perceptron also had a patch board, shown in the middle photograph, which allowed different configurations of input features to be tried. Often these were wired up at random to demonstrate the ability of the perceptron to learn without the need for precise wiring, in contrast to a modern digital computer. The photograph on the right shows one of the racks of adaptive weights. Each weight was implemented using a rotary variable resistor, also called a potentiometer, driven by an electric motor thereby allowing the value of the weight to be adjusted automatically by the learning algorithm.\"\n\nsource: Bishop, Christopher M. (2006). Pattern Recognition and Machine Learning\n\nThe Perceptron had a 400-pixel visual input and 1000 neurons in the hidden layer. https://twitter.com/DiegoKuonen/status/1130352233223262208"},{"id":"epoch:787b11189d7462d799b8:training-compute","entityId":"epoch:787b11189d7462d799b8","metric":"training-compute","value":694894.9377361819,"unit":"FLOP","observationPeriod":"1957-01-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Extracted from AI and Compute (https://openai.com/blog/ai-and-compute/) charts by using https://automeris.io/WebPlotDigitizer/.\n\nAdditional experiment described in https://babel.hathitrust.org/cgi/pt?id=coo.31924004657973&seq=70 \n- 400 input (20x20) - 512 hidden with 40 fixed connections each (not learned) - 1 output (learned)\nParameters: 512*1=512\nForward flop: 41*512=20992\nForward + \u201cbackward flop\u201d: 43*512=22016 (only last layer was adjusted)\n100*22016=2201600\n"},{"id":"epoch:dd2787ab41b1431c6286:parameters","entityId":"epoch:dd2787ab41b1431c6286","metric":"parameters","value":225.0,"unit":"parameters","observationPeriod":"1955-03-01","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"Figure 4 contains the learnt weight matrix"},{"id":"epoch:2fc4929ecc20cb5f4ee0:parameters","entityId":"epoch:2fc4929ecc20cb5f4ee0","metric":"parameters","value":40.0,"unit":"parameters","observationPeriod":"1952-01-08","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The link below seems to suggest the SNARC had 40 cells, each with a dial that acts as a configurable weight.\n\nhttps://www.webofstories.com/play/marvin.minsky/137"},{"id":"epoch:1ded709e181fdf0416f8:parameters","entityId":"epoch:1ded709e181fdf0416f8","metric":"parameters","value":40.0,"unit":"parameters","observationPeriod":"1950-07-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The learned part is the maze configuration. There are 25 squares of the maze. The 16 squares to the left top corner have each one adjacent square down and one adjacent square up, for a total of 16*2 walls. We only need to count the 8 spare walls connecting the squares in the right side and the bottom side. In total there are 16*2+8 walls."},{"id":"epoch:1ded709e181fdf0416f8:training-compute","entityId":"epoch:1ded709e181fdf0416f8","metric":"training-compute","value":40.0,"unit":"FLOP","observationPeriod":"1950-07-02","updatedAt":"2026-10-23T12:29:34.203347+00:00","source":"https://epoch.ai/data/all_ai_models.csv","methodologyVersion":"epoch-model-history-v1","publicationStatus":"approved","note":"The \"training\" consists on the mouse running around and checking each wall (assuming each relay switch is one operation)."}]}