China leads the open-weight AI model race. As of July 2025, Chinese models, primarily Alibaba's Qwen, overtook American models in Hugging Face downloads. The gap has since grown to 1.6 billion downloads, with China's total at 3.2 billion, exactly twice the American total. Models like GLM-5.2 and Kimi K3 have crossed the same agentic capability threshold that Anthropic's Claude Code crossed in December 2025. Chinese companies have held the lead in both technical capability and adoption metrics for approximately 18 months.
The terminology matters for understanding what is actually at stake. Open-weight models, including Meta's Llama, Alibaba's Qwen, Google's Gemma, and DeepSeek's releases, make weights publicly available under licenses but withhold training data and code. True open-source models, the rarer category, include everything needed to reproduce the model from scratch. The author, a former Allen Institute for AI researcher who helped build the OLMo models, notes that genuine open-source work remains an American nonprofit stronghold, with AllenAI, OpenAthena's Marin, and EleutherAI's Pythia representing the field. Chinese dominance is in the open-weight category, which is the one that drives commercial deployment.
The full piece is worth reading for the policy context: these remarks were delivered directly to Congressional members and staff as a briefing on U.S.-China AI competition. The author also runs the ATOM Project, a public dashboard tracking download data by region, and the piece includes specific methodology on how that data is collected and interpreted. If you want to understand how policymakers are being briefed on open models right now, this is the primary source.
[READ ORIGINAL →]