Hugging Face CEO Clément Delangue said China has taken the lead in open-model artificial intelligence and could soon challenge U.S. developers at the technological frontier, as a wave of increasingly large and inexpensive Chinese systems reshapes the global race.
Speaking on CNBC’s Squawk on the Street on Monday, Delangue said Chinese developers were “clearly dominating on open models right now,” pointing to a culture of collaboration and rapid knowledge-sharing across the country’s AI ecosystem.
He added that he would not be surprised to see China begin leading frontier development by the end of 2026 or in 2027 if its current pace continued. By contrast, Delangue said many U.S. laboratories were “building in silos,” limiting the wider exchange of research and engineering advances.
Open-model downloads exposed the new balance
Delangue’s warning follows a March report from Hugging Face that showed Chinese models had already overtaken their U.S. counterparts in both monthly and overall downloads on the company’s widely used platform.
Chinese-developed models accounted for 41% of Hugging Face downloads during the preceding year, giving them the largest national share. The report also found that most models reaching the platform’s trending rankings after being created in 2025 were either developed in China or built from Chinese foundations.
That momentum extends beyond a few high-profile launches, as Chinese companies increasingly release model files that developers can download, adapt and run independently, helping the technology spread across research, commercial products and real-world applications.
Hugging Face acknowledged that automated software pipelines can inflate download totals, particularly for smaller models, but said the broader move toward compact, deployable systems remained clear.
Alibaba, DeepSeek and Kimi accelerate China’s AI race
The race accelerated on Monday when Alibaba introduced Qwen3.8-Max, its largest model yet, with 2.4 trillion parameters and a context window of up to one million tokens.
To reduce computing demands, the model activates 95 billion parameters during operation rather than running its entire network at once, a design intended to lower costs and shorten response times.
DeepSeek has taken a different route, placing lower deployment costs at the centre of its updated V4-Flash model. The July 31 release charges $0.14 per million uncached input tokens and $0.28 per million output tokens, giving developers another low-cost route into advanced reasoning and coding tools.
Moonshot AI has also entered the contest with Kimi K3, released on July 16 with 2.8 trillion parameters, native visual understanding and a one-million-token context window.
Together, the launches suggest the AI race is moving beyond benchmark rankings into a wider battle over cost, scale, openness and the speed at which new technology reaches developers worldwide.
