Hugging Face says open-model activity on its platform surged during 2025, with Chinese models taking a 41% plurality of recent downloads. The larger point is less triumphal: participation is widening, but attention remains spectacularly concentrated.
Hugging Face Watch analysis
What happened
In its State of Open Source on Hugging Face: Spring 2026 report, published on 17 March, the company says its platform grew to 13 million users, more than two million public models and over 500,000 public datasets during 2025.
The analysis covers downloads, repository creation, derivative models, geography and research activity. These are Hugging Face platform signals rather than a census of every open model in existence, but the Hub is large enough to make the direction difficult to dismiss with a tasteful shrug.
Key findings
- China took the download lead
Chinese models generated a 41% plurality of downloads over the preceding year, overtaking the US in Hugging Face’s analysis. - A tiny elite commands attention
The top 200 models, about 0.01% of those available, captured 49.6% of all downloads. - Most models remain obscure
Roughly half of the models on Hugging Face recorded fewer than 200 total downloads. - Independent builders gained ground
Their download share rose from 17% to 39%, while industry’s share fell from about 70% to roughly 37%. - Qwen became derivative-model infrastructure
Hugging Face counted over 113,000 Qwen derivatives, or more than 200,000 models when every Qwen-tagged model was included.
Why it matters
The open-model contest is no longer a simple US-versus-China leaderboard. Fine-tuners, quantisers and small collectives increasingly determine which models become affordable, portable and useful on ordinary hardware. Distribution is becoming a form of technical power.
Yet abundance is not the same as diversity in practice. Millions of repositories can still funnel most usage towards a sliver of model families, making compatibility decisions, licences and community tooling around those families unusually consequential.
Our read
The smart response is neither to crown a national winner nor to count repositories like medals. Teams choosing an open-model stack should examine the surrounding derivative ecosystem, hardware fit and deployment flexibility alongside benchmark scores. Downloads reveal momentum, not quality by divine decree.
What to watch
- Whether China’s download lead persists across future reporting periods.
- Whether Qwen derivatives translate into durable production adoption.
- How much usage continues shifting from original labs to independent adapters and quantisers.
- Whether concentration around the most popular models rises despite continuing repository growth.
Discussion spark: Does China’s lead in open-model downloads change which model families your team tests first, or do cost and hardware still decide?
Sources and evidence
- State of Open Source on Hugging Face: Spring 2026 (2 September 2026, 06:25 UTC)
not affiliated with or endorsed by Hugging Face