What happened?
In a report titled "State of Open Models: Summer 2026 Observations" published on August 14, 2026, Hugging Face examined the open model ecosystem from January to August 2026. The number of model repositories on the platform rose from 2.43 million to 2.96 million, dataset counts climbed from 711 thousand to 1 million, and Spaces grew from 1.00 million to 1.44 million. Despite this growth, 85.6% of models received fewer than 200 downloads ever, and 1.5% of repositories accounted for 99.2% of all downloads.
According to the report, Chinese labs skipped the traditional progression from small to large models in 2026, releasing massive models directly. The largest Chinese-origin open model ranged from 754 billion to 2.78 trillion parameters month to month, while the largest U.S.-origin model stayed under 130 billion parameters in five of the seven months. The exceptions were NVIDIA's 561-billion-parameter Nemotron 3 Ultra, released in May and June, and Thinking Machines Lab's Inkling model.
Moonshot, MiniMax, Xiaomi and Z.ai released almost no models under 70 billion parameters, while Tencent and Alibaba's Qwen family span the whole range from under 1 billion up. Xiaomi and Meituan, absent from open source last year, crossed the trillion-parameter threshold this year.
Why does it matter?
On the U.S. side, the organizations releasing the most new open models aren't model labs but hardware makers: AMD and NVIDIA topped the list with more than 200 new model repositories each, and LiquidAI ranked third with roughly 100. Google and Meta, pioneers of open releases in past years, fell behind NVIDIA in new model counts; Meta moving its flagship models to closed source reinforces the shift.
Most U.S.-origin releases above 100 billion parameters aren't new models but adaptations built on Chinese-origin models. The list of genuinely original American models at this scale is short: Thinking Machines' 952B Inkling, NVIDIA's 561B Nemotron 3 Ultra and 124B Nemotron 3 Super, and Arcee AI's 399B Trinity-Large. AMD did extensive conversion work at this scale but released no original models — work that helps trillion-parameter Chinese models run efficiently on American hardware.
When Hugging Face compared the top 25 most-downloaded repositories with the top 25 most-liked ones, only one model appeared on both lists. The all-MiniLM-L6-v2 model was downloaded 1.55 billion times over seven months but received only 5,156 likes; by contrast, the Kimi-K3 model averaged 60 downloads for every like it received. The report emphasizes that likes signal interest in a model, while downloads indicate it has been put into production.
Nearly all of MiniMax's 2026 downloads come from models over 70 billion parameters; that figure is 88% for Moonshot, 55% for DeepSeek, and 39% for Z.ai. By contrast, almost none of Google's, Microsoft's, or IBM Granite's 2026 downloads come from models above 70 billion parameters; that share is 14% for NVIDIA and 9% for Meta. Moonshot's strategy of focusing solely on its largest models brought in 37 million downloads over the year, while the Qwen family's broad range of scales reached 2.045 billion downloads — roughly 55 times more than Moonshot.
Comparison: Licensing approach
| License type | Chinese labs (20B+ parameters) | U.S. labs (same scale) |
|---|---|---|
| Apache 2.0 | 59% | 29% (Apache or MIT combined) |
| MIT | 22% | - |
| Proprietary/restricted terms | 0 | 41% |
| License not disclosed | - | 30% |
In the dataset of 178 major Chinese model releases examined, none carried a license restricting commercial use. DeepSeek and Z.ai distribute their models — ranging from 700 billion to 1.65 trillion parameters — under a plain MIT license, meaning Chinese labs license their largest models with nearly the same freedom as their smallest ones.
What's next?
Hugging Face repeats this analysis every six months; the report does not specify the next period or its publication date. It notes that value in open source comes not from license revenue but from indirect channels — APIs, cloud services, hardware positioning — and that the valuations of Z.ai and Kimi point the same way.
What we know
- Report period: January-August 2026, based on Hugging Face Hub data.
- Model repository count: grew from 2.43 million to 2.96 million.
- Dataset count: grew from 711 thousand to 1 million.
- Qwen family total downloads: 2.045 billion (2.061 billion including all repositories).
- Moonshot total downloads: 37 million.