Open models, summer 2026: Chinese laboratories skip the small model stage

Hugging Face's half-year report shows 2.96 million model filings, 151,448 Qwen derivatives, and a striking size gap between Chinese and U.S. publications.

Hugging Face published its state of play of open-weight models on August 14. Between January and August 2026, the number of model filings on the platform increased from 2.43 million to 2.96 million, and the number of datasets from 711,000 to one million.

The geography of the sector is clearly apparent from the figures. Chinese laboratories publish frontier models ranging from 754 billion to 2,780 billion parameters each month — Moonshot's Kimi-K3 reaches 2,800 billion, Qwen peaks at 2,400 billion, DeepSeek-V4-Flash at 284 billion. On the American side, releases have remained below 130 billion parameters five months out of seven, with the notable exception of NVIDIA's Nemotron 3 Ultra (561 billion). In other words, Chinese teams no longer follow the classic progression from the small model to the large: they publish directly at the top.

The derivative ecosystem is following the same movement: 151,448 Qwen-derived models are listed, i.e. 2.6 times the footprint of Meta, with 39.6 million monthly downloads in GGUF format. Finally, the report highlights a discrepancy between visibility and actual use: models with less than a billion parameters represent 83% of cumulative downloads, only one repository is in both the top 25 downloads and favorites, and most publications remain under a permissive license (Apache 2.0, MIT).