AI Models
Who builds notable AI models, how large they are and how openly they are released, from Epoch AI data in Stanford's AI Index Report 2026, plus AI activity on GitHub and Hugging Face.
AI Progress · 27 charts · 7 topics
- Source
- Stanford AI Index Report 2026
- Charts
- 27 with CSV data
- Last checked
- 20 Oct 2026
- Data
- All CSVs (zip)
64Notable AI models from US organisations in 2025, against 35 from ChinaView chart →39Arena points between the top US and top Chinese model, March 2026 (346 in May 2023)View chart →92%Industry's share of notable AI models in 2025, up from 15% in 2014View chart →135BMedian parameter count of 2025's notable AI models, against 16M in 2015View chart →21Notable AI models released by OpenAI in 2025, the most of any organisationView chart →5.6MAI-related projects on GitHub in 2025, up 54% from 3.6 million in 2024View chart →
US organisations released 64 notable AI models in 2025 and Chinese organisations 35, according to Epoch AI data in the AI Index. The performance gap between the two countries' best models has narrowed: China's top model trailed the top US model by 39 Arena points in March 2026, down from 346 in May 2023. Industry builds almost all notable models, 98 of the 107 released in 2025, while academia alone produced 2.
Read the full overview
Models are larger and take longer to train. The median notable model of 2025 had 135 billion parameters, against 16 million in 2015, and those that report a training time took a median of 78 days. The largest training run in Epoch AI's database, GPT-6 Astra, is estimated at 9.4×10²⁶ FLOP. Openness is limited: 55 notable models of 2025 were available only through an API or hosted product and 45 had open weights, and training code stayed unreleased for 86 of the 107.
The charts cover the US and China, open and closed models, the leading labs, industry and academia, model scaling, and open-source activity on GitHub and Hugging Face. OpenAI released the most notable models in 2025, 21, ahead of Google's 15. GitHub hosted about 5.6 million AI projects in 2025, up 54% in a year, while the US share of downloads of the most-downloaded Hugging Face models fell from 73% in 2020 to 17% in 2025.
The data come from Stanford's AI Index Report 2026, which draws on Epoch AI's database of notable models, the Arena leaderboards, Artificial Analysis, GitHub and Hugging Face. "Notable" is Epoch AI's selection of influential models, not a count of every release, and its live database keeps adding models. Medians for parameters, datasets and training time cover only models with a published or estimated figure, and compute for most frontier models is estimated rather than reported by developers.
AI Adoption · 2026AI Statistics: the pocket guide83Free for subscribersAll 83 key numbers on two pages, with sources.Every hub's headline numbers in one PDF, ready for a board paper or a slide.Get the guide
01
US vs China AI Models
US organisations released 64 notable AI models in 2025 and Chinese organisations 35, according to Epoch AI. The performance gap has narrowed: China's top model trailed the top US model by 39 Arena points in March 2026, down from 346 in May 2023. Since 2003, the US has produced 596 notable models and China 160.
US vs China AI Models64Notable AI models from US organisations in 2025, against 35 from China and 8 from South KoreaThe US produced 64 notable AI models in 2025; China produced 35Chart · CSVView →
US vs China AI Models35Notable AI models from China in 2025, up from 12 in 2022China's notable AI models rose for a third straight year to 35; Europe's fell to 2Chart · CSVView →
US vs China AI Models596Notable AI models from US organisations, 2003–2025 (China: 160; UK: 84)Since 2003, the US has produced 596 notable AI models, nearly four times China's 160Chart · CSVView →
US vs China AI Models38×GPT-6 Astra's estimated compute (9.4×10²⁶ FLOP) relative to China's largest, Doubao-pro (2.5×10²⁵)The largest US training run used about 38 times the compute of China's largestChart · CSVView →
US vs China AI Models1,464Arena score of the top Chinese model in March 2026, 39 points behind the top US model (1,503)China's top AI model trailed the top US model by 39 Arena points in March 2026, down from 346 in 2023Chart · CSVView →
US vs China AI Models1,618AI models released in the US, cumulative 2010–2025, up from 237 by 2018The US has released 1,618 AI models since 2010, almost twice China's 849Chart · CSVView →02
Open vs Closed AI Models
In 2025, 55 notable AI models were released only through an API or hosted product and 45 with open weights. The best open-weight model trailed the best closed model by 49 Arena points in March 2026, down from 174 in May 2023. Training code stayed unreleased for 86 of 2025's 107 notable models.
Open vs Closed AI Models55Of 2025's 107 notable models were available only through an API or hosted product; 45 had open weightsMore notable AI models launched behind an API than with open weights in 2025: 55 vs 45Chart · CSVView →
Open vs Closed AI Models86Of 2025's 107 notable AI models had unreleased training code; 4 were fully open sourceTraining code stayed closed for 86 of 2025's 107 notable AI modelsChart · CSVView →
Open vs Closed AI Models1,454Arena score of the top open-weight model in March 2026, 49 points behind the top closed-weight model (1,503)The best open-weight model trailed the best closed model by 49 Arena points in March 2026, down from 174Chart · CSVView →
Open vs Closed AI Models16Openness Index score out of 100 for K2 Think and Olmo 3 32B Think, the highest of 14 models; Claude 4.5 Haiku scores 2K2 Think and Olmo 3 32B Think top the Openness Index at 16 of 100; no other model scores on pre-training dataChart · CSVView →03
Top AI Labs Compared
OpenAI released the most notable AI models in 2025, 21, ahead of Google's 15 and Alibaba's 11. Since 2014 Google leads with 194 notable models, more than twice Meta's 87. On the Arena, the best models from Anthropic, xAI, Google and OpenAI sat within 22 points of each other in March 2026.
Top AI Labs Compared21Notable AI models released by OpenAI in 2025 (Google 15, Alibaba 11, Anthropic 7)OpenAI released the most notable AI models in 2025: 21, ahead of Google's 15Chart · CSVView →
Top AI Labs Compared194Notable AI models from Google, 2014–2025 (Meta 87, OpenAI 61)Google leads with 194 notable AI models since 2014, more than twice Meta's 87Chart · CSVView →
Top AI Labs Compared1,503Arena score of Anthropic's top model in March 2026, the highest of eight developers; xAI 1,495, Google 1,494, OpenAI 1,481Anthropic, xAI, Google and OpenAI's best models were within 22 Arena points in March 2026Chart · CSVView →04
Industry vs Academia
Industry built 98 of the 107 notable AI models released in 2025, a 92% share, up from 15% in 2014. Academia alone produced 2 notable models in 2025, down from 21 in 2014, when academia accounted for 64% of the year's notable models and was the largest sector.
Industry vs Academia98Of 2025's 107 notable AI models came from industry alone; academia alone produced 2Industry built 98 of the 107 notable AI models released in 2025Chart · CSVView →
Industry vs Academia92%Industry's share of 2025's notable AI models, up from 15% in 2014Industry's share of notable AI models rose from 15% in 2014 to 92% in 2025Chart · CSVView →05
AI Model Scaling
The median notable AI model of 2025 had 135 billion parameters, up from 16 million in 2015, and was trained on a median of about 15 trillion datapoints. The 14 notable models of 2025 that report a training time took a median of 78 days, against 15 days in 2023. GPT-6 Astra's training run is estimated at 9.4×10²⁶ FLOP.
AI Model Scaling135BMedian parameter count of 2025's notable AI models (57 with a published figure), against 16M in 2015The median notable AI model of 2025 had 135B parameters, up from 16M in 2015Chart · CSVView →
AI Model Scaling15TMedian training dataset of 2025's notable AI models (29 with a published figure)Notable AI models of 2025 trained on a median of 15 trillion datapointsChart · CSVView →
AI Model Scaling78 daysMedian training time of the 14 notable 2025 models that report one (2023: 15 days)Training a notable AI model took a median of 78 days in 2025Chart · CSVView →
AI Model Scaling9.4×10²⁶FLOP: Epoch's estimate for GPT-6 Astra (Sep 2026), the largest training run in its databaseThe largest AI training runs now exceed 10²⁶ FLOP, led by GPT-6 Astra at 9.4×10²⁶Chart · CSVView →06
GitHub AI Projects
GitHub hosted about 5.6 million AI-related projects in 2025, up 54% from 3.6 million in 2024, and about 207,000 had at least 10 stars. The US share of those starred projects fell from 75.3% in 2011 to 31.7% in 2025, though US-based projects have collected 30 million stars, twice any other region.
GitHub AI Projects5.6MAI-related projects on GitHub in 2025, up from 3.6 million in 2024GitHub hosted 5.6 million AI projects in 2025, up 54% in a yearChart · CSVView →
GitHub AI Projects207KAI projects on GitHub with at least 10 stars in 2025, up 18% from 175,000 in 2024About 207,000 GitHub AI projects had at least 10 stars in 2025Chart · CSVView →
GitHub AI Projects31.7%US share of GitHub AI projects with at least 10 stars in 2025, down from 75.3% in 2011The US share of notable GitHub AI projects fell from 75% in 2011 to 32% in 2025Chart · CSVView →
GitHub AI Projects18.2MStars given to AI projects on GitHub in 2025, up from 13.6 million in 2024AI projects on GitHub gained 18.2 million stars in 2025Chart · CSVView →
GitHub AI Projects30MCumulative GitHub stars on US-based AI projects by 2025, against 15.3M for the rest of the world and 13.0M for EuropeUS-based AI projects have collected 30 million GitHub stars, twice any other regionChart · CSVView →07
Hugging Face Models
Hugging Face users created about 1.18 million new models in 2025, including 332,000 in the fourth quarter. US developers' share of downloads of the most-downloaded models fell from 73.1% in 2020 to 16.6% in 2025, while China's reached 16.9%. Google's share fell from 42% in 2020 to 0.3% in 2025.
Hugging Face Models332KNew models created on Hugging Face in Q4 2025, plus 153,000 new datasetsHugging Face users created 332,000 new models in the last quarter of 2025Chart · CSVView →
Hugging Face Models16.6%Share of top Hugging Face model downloads from US-based developers in 2025, down from 73.1% in 2020; China 16.9%Models from US developers fell from 73% of top Hugging Face downloads in 2020 to 17% in 2025Chart · CSVView →
Hugging Face Models42% → 0.3%Google's share of top Hugging Face model downloads, Q2 2020 vs Q3 2025Google's share of top Hugging Face downloads fell from 42% in 2020 to under 1%Chart · CSVView →Related hubs
All 13 hubs →About the data: 27 interactive charts from the AI Index Report 2026 (Stanford HAI), using data from research partners credited on each chart. Each chart page explains what the chart shows, lists its key numbers, cites its source and links to the data as a CSV.