Chinese open-weight models lead U.S. counterparts in downloads and benchmarks, per Congressional testimony

A recent Congressional testimony expanded into a public analysis argues that Chinese open-weight AI models have surpassed U.S. models in both popularity and technical performance. The author, Nathan Lambert, presents data indicating that Chinese open-weight models now account for roughly twice as many downloads as U.S. open-weight models—about 1.6 billion more downloads, totaling 3.2 billion compared to 1.6 billion for U.S. models.

The analysis cites the Artificial Analysis Intelligence Index (AAII) to compare model performance, showing that as of September 14, 2026, the top three Chinese open-weight models—Z.ai’s GLM-5.3 (score: 45), GLM-5.3-Flash (score: 42), and Moonshot AI’s Kimi K3 (score: 44)—outperform the leading U.S. open-weight models. The top U.S. performers are Thinking Machines’ Inkling and Inkling Small (both score: 26) and Nvidia’s Nemotron 3 Ultra (score: 23).

The author attributes China’s lead to faster release cycles and a focus on tasks with clear user demand, such as agentic coding, which flatters performance on public benchmarks. Chinese labs have also begun purchasing advanced training data, including reinforcement learning environments for agentic tasks, from both U.S. and Chinese providers, further narrowing the gap to frontier capabilities.

Despite China’s lead in open-weight models, the analysis notes that U.S. closed models remain 2–5 months ahead of Chinese open-weight models, and 6–9 months ahead of U.S. open-weight models. The testimony also highlights that Chinese open-weight models are closest to U.S. closed models in agentic coding tasks but lag in open-ended scientific domains like physics and biology.

The document discusses regulatory implications, including concerns about the risks of open-weight models enabling misuse, such as cybersecurity incidents. It argues that restricting access to open-weight models could disproportionately harm U.S. businesses, citing an example where Hugging Face used a Chinese open-weight model to analyze a cyberattack after closed models refused to respond.