Chinese AI Labs Now Dominate Open-Source Models. Here Is What the Numbers Actually Say.
A Hugging Face data deep-dive covering January to August 2026 finds Chinese labs releasing the biggest open models by far, while the tools developers actually depend on are years-old American ones nobody talks about.
Key points
- Public model repositories on Hugging Face grew from 2.43 million to 2.96 million between January and August 2026.
- Chinese labs released open models as large as 2.78 trillion parameters in 2026; most U.S. labs stayed under 130 billion.
- The most-downloaded model of the period, all-MiniLM-L6-v2 (a small text-processing tool released in 2022), was pulled 1.55 billion times in seven months against just 5,156 likes.
- 59% of large Chinese open models (above 20 billion parameters) ship under the Apache 2.0 licence, which lets anyone use or modify them commercially, and none carry a non-commercial restriction.
- AMD and NVIDIA each published more than 200 new model repositories in 2026, more than any AI research lab.
Every few months someone declares that the AI race has a new leader. The summer 2026 data from Hugging Face, the platform where researchers and companies share AI models the way developers share code, is more complicated than that story. The headline is Chinese dominance at the frontier. The footnote, which matters more for most people, is that developers keep reaching for the same handful of old, boring, reliable tools.
How big is the gap between Chinese and American open models?
Very large, and growing. Chinese labs released models as large as 2.78 trillion parameters (a parameter is roughly one adjustable dial inside an AI model; more dials generally means more capability, but also more computing power to run). U.S. labs stayed under 130 billion parameters in five of the seven months covered, with NVIDIA's Nemotron 3 Ultra at 561 billion parameters being the main exception.
Moonshot, MiniMax, Xiaomi and Z.ai publish almost nothing below 70 billion parameters. For context, a model that size will not run on a typical laptop or even a high-end gaming PC. Alibaba's Qwen family and Tencent cover the full range, from under one billion parameters up, which means a student or small business can actually download and use them.
The licensing terms are striking too. Of 178 Chinese model releases above 20 billion parameters, 59% carry the Apache 2.0 licence and 22% carry MIT, both of which allow free commercial use. Not one carries a non-commercial restriction. On the American side of the same size band, only 29% is Apache or MIT; 41% sits under custom terms.
What are developers actually using day to day?
Not the shiny new things. All-MiniLM-L6-v2, a compact text-matching model released in 2022, recorded 1.55 billion downloads in the first seven months of 2026. Kimi-K3, one of the most-talked-about new releases, received about 60 downloads per like it got. Likes reflect excitement; downloads reflect what is quietly running inside real software pipelines.
Thirteen of the top 25 most-downloaded repositories date from 2022. Not one model published in 2026 makes that list.
I tested this dynamic myself when helping a teacher friend set up an automated tool to sort student feedback. We reached for a two-year-old embedding model, a type of AI that turns text into numbers so a computer can compare meaning, because it was stable, well-documented and free. The newest frontier model never came up.
Does this mean U.S. labs have stepped back from open source?
Partly, yes. Google and Meta now rank well below NVIDIA in new model releases, and Meta has moved its flagship models behind closed doors. The companies publishing the most new open models in 2026 are hardware makers: AMD and NVIDIA. Each cleared 200 new repositories. Open models, for them, are essentially marketing. A free model optimised for your chip is a live demo that the chip works.
Where U.S. organisations still dominate is in the workhorse models that power everyday applications: smaller text models, embedding models, speech and vision tools from Google, Microsoft and IBM Granite that rack up hundreds of millions of downloads a year.
| Layer | Who leads | Example |
|---|---|---|
| Frontier scale (100B+ parameters) | Chinese labs | Qwen 3.8 Max (2.4 trillion parameters) |
| Hardware-optimised releases | NVIDIA, AMD | Nemotron 3 Ultra (561B) |
| Most-downloaded single model | U.S. (2022 vintage) | all-MiniLM-L6-v2 (1.55B downloads) |
| New open-model repositories | NVIDIA, AMD (200+ each) | Various |
| Permissive licensing at scale | Chinese labs | 59% Apache 2.0, 0% non-commercial |
Common questions
Does any of this affect the AI tools I use at work?
Probably not directly. Most consumer apps and business tools are built on a layer of established models that rarely changes. The frontier activity described here matters more to developers building new products than to people using existing ones.
Why do Chinese labs give away such large models for free?
The weights, the files you download, are free. The return comes elsewhere: cloud and API fees when businesses pay to run the models at scale, and the strategic benefit of having the whole world's developers build on your technology.
