Chinese open-source AI models accounted for 17.1% of global downloads on Hugging Face over the year ending August 2025, narrowly edging the US share of 15.86%, according to a joint study by researchers at MIT and Hugging Face. It's the first time Chinese models have led on that metric. The gap is small. The trajectory is not.
Close numbers, wider trend
It looks like a photo finish at the top line. 17.1 versus 15.86 percent is the kind of margin you'd shrug at in most contexts. Then you look at the more recent slice. Between February 2025 and February 2026, Chinese models accounted for 41% of Hugging Face downloads against 36.5% for US-developed ones, according to more recent data analyzed by The New Stack. The yearlong study captured a crossover. The monthly numbers since show a widening lead.
What the study counts is downloads of new model releases, not total compute used or revenue earned. Those last two still tilt heavily toward American labs. OpenAI, Anthropic, and Google charge for access to closed systems that businesses pay real money for. Downloads, by contrast, measure developer mindshare. They matter, but in a different way.
Where the gap actually shows up
The more interesting story is what happens after the download. In Stanford's index work, 83% of Chinese respondents said AI products are more beneficial than harmful, against 39% of Americans and 36% in the Netherlands. That gap survived the post-ChatGPT recalibration and shows up in product metrics. As of late last year, more than 600 million people in China were using generative AI tools, according to the China Internet Network Information Center, a 142% year-over-year jump. ByteDance's Doubao chatbot reportedly hit around 100 million daily active users during Chinese New Year.
Lizzi Lee of the Asia Society Policy Institute described Chinese users in recent reporting as "basically acting as real-time testers at scale." That's the part that's hard to manufacture elsewhere. A billion-person market willing to install something the day it ships is a different feedback loop than a US one nervous about job displacement.
So what does open really mean here
Alibaba's Qwen family has surpassed 300 million downloads on Hugging Face, and the platform now lists more user-built Qwen variants than variants of Google and Meta models combined. "Open source is a more widespread trend in China than in the United States. American companies are not willing to do it; they want to make money and don't want to reveal their trade secrets," Wendy Chang of the Mercator Institute for China Studies told the Financial Times. Fair enough as description. Less convincing as thesis. Chinese labs aren't open because they're idealistic. They're open because US export controls cut them off from Nvidia's best chips, and a global developer base helps compound the work they can do with what's left.
"Open" also doesn't mean what it usually means. Most Chinese releases ship model weights, not training data or full code, which falls short of the OSI definition of open source AI. The models also carry Chinese content moderation defaults, which matters if you're building anything politically textured.
What's next
China's "AI Plus" national plan targets 90% penetration of AI terminals and agents by 2029. That's the kind of round number set in five-year plans and then quietly adjusted. The closer thing to watch is the next DeepSeek and Qwen release cycles over the coming months. If the download lead widens past noise, the adoption story stops being a thesis and starts being market structure.
