Hugging Face finds Chinese labs lead open model releases at frontier scale while U.S. hardware vendors dominate new model uploads

Between January and August 2026, the number of public model repositories on the Hugging Face Hub grew from 2.43 million to 2.96 million, datasets from 711,000 to 1 million, and Spaces from 1.00 million to 1.44 million. Roughly 85.6% of models have fewer than 200 lifetime downloads, and 1.5% of repositories account for 99.2% of all downloads.

At the frontier scale, Chinese labs released the largest open models in most months of 2026, with parameter counts ranging from 754 billion to 2.78 trillion. American labs’ monthly ceiling stayed under 130 billion in five of seven months, with exceptions including NVIDIA’s Nemotron 3 Ultra at 561 billion and Thinking Machines Lab’s Inkling at 952 billion.

Two patterns explain China’s frontier lead: building large models ceased being a differentiator, and community quantization layers enable running large models on consumer hardware within days. This allows labs to stake a position via size without shipping small models first.

U.S. hardware vendors led new model uploads, with AMD and NVIDIA each releasing more than 200 new model repositories, far ahead of other organizations. LiquidAI ranked third with around 100 new repositories. Google, Microsoft, IBM Granite, and OpenAI’s older vision and speech models collectively generated hundreds of millions of downloads annually, indicating continued U.S. participation in open source via smaller and embedding models.

Only a few major original American models appeared above 100 billion parameters in 2026: Thinking Machines’ Inkling (952B), NVIDIA’s Nemotron 3 Ultra (561B), Nemotron 3 Super (124B), and Arcee AI’s Trinity-Large (399B). AMD contributed many conversions but no original model at this scale, reflecting a shift toward distribution and optimization rather than model creation.

Chinese open models are increasingly optimized for domestic chips, mirroring the reverse of the earlier U.S.-led hardware-software co-design trend.

Among the top 25 model repositories by downloads in 2026, none published in 2026 appeared in the list, while thirteen dated from 2022. The most-downloaded model, all-MiniLM-L6-v2, recorded 1.55 billion pulls in seven months, while frontier models like Kimi-K3 received about 60 downloads per like, illustrating that downloads accrue to small, stable models over years rather than to recent frontier releases.

Chinese frontier labs such as MiniMax, Moonshot, and Z.ai recorded the majority of their 2026 downloads from models above 70 billion parameters, whereas large U.S. publishers like Google, Microsoft, and IBM Granite recorded essentially none of their 2026 downloads above 70 billion. Qwen’s broad release strategy across model sizes reached 2,045 million downloads, about 55 times Moonshot’s 37 million, underscoring the value of covering multiple sizes for adoption.

Of 178 Chinese releases above 20 billion parameters in 2026, 59% carried Apache 2.0 licenses and 22% MIT, with none carrying non-commercial restrictions.