Chinese-developed models pulled 41% of all downloads on Hugging Face over the past year, outpacing US versions at 36.5%. That's a massive swing from just months earlier, when the gap was barely visible at 17% versus 15.8% in August 2025.
Hugging Face CEO Clément Delangue has been watching this unfold and isn't shy about naming the shift. Open-weight models from Chinese labs like Alibaba's Qwen and Zhipu AI's GLM family are matching or beating American counterparts on performance benchmarks. More importantly, they're shipping with permissive licenses that strip away legal friction, letting companies deploy them commercially without the overhead that sometimes tangles US alternatives.
When the safety guardrails became a liability
The real story isn't just numbers on a dashboard. In July 2026, Hugging Face itself got hit by a cyberattack. The company's incident response team reached for US-developed models first, assuming they'd handle the adversarial scenario. They didn't. Safety guardrails that normally protect users instead neutered the models when Hugging Face needed raw capability under pressure.
So they switched to Zhipu AI's GLM 5.2. It worked. That single pivot says more about competitive reality than any market share chart.
What's actually moving the needle
Delangue points to three concrete advantages of open models: they run locally on your own hardware without per-query API costs, they let teams customize them for specific problems, and they eliminate the pricing treadmill of services from OpenAI or Anthropic. For companies building at scale, that math gets compelling fast.
Chinese labs understood this before the US caught up. They positioned permissive licensing as a feature, not an afterthought. Now they're reaping downloads and mindshare from engineers who actually ship products.
This is informational analysis, not financial or investment advice. Market dynamics in AI development continue to shift rapidly.
