China overtakes US in open-weight AI models

Summary

Nathan Lambert's expanded congressional testimony argues that Chinese open-weight models (Qwen, GLM, Kimi) have significantly outpaced their US counterparts in download counts and benchmark performance since mid-2025. According to the testimony, Chinese models are only 2–5 months behind the US's closed-source frontier models. In contrast, US open models are lagging behind their Chinese counterparts by 6–9 months.

Why it matters

This assessment shifts the narrative of the US-China AI race from closed models to the dominance of open-weight models. This has direct implications for global AI adoption, cybersecurity risks, and US technology policy. If open-weight models become the primary vehicle for enterprise and sovereign AI deployment, the current gap implies that US businesses are becoming increasingly dependent on Chinese model ecosystems. The testimony also highlights a structural policy dilemma: restricting Chinese open models would negatively impact US users more than Chinese users.

Highlights

• China leads by approximately 1.6 billion in Hugging Face download counts. China's total downloads stand at 3.2 billion, compared to the US's roughly 1.6 billion. This gap has been widening since August 2025.
• Top Chinese open models score 42–45 points on the AAII benchmark test. The best US open models (Inkling, Nemotron 3 Ultra) score 23–26 points.
• According to Lambert, completely blocking distillation from US closed-source models would only widen the capability gap by an additional 1–2 months. This indicates that China's lead is structural rather than purely data-driven.

Critical take

The article was written by a former AI2 researcher who contributed to the development of Olmo. This suggests a slight institutional interest in keeping the US 'truly open-source' category visible. Additionally, the benchmark gap of 45 versus 26 may overstate real-world capability differences. The author themselves notes that Chinese labs are optimized for narrower, benchmark-friendly task distributions. Furthermore, the regulatory risk section lacks concrete details. It merely references the OpenAI-HuggingFace incident and does not provide specifics on existing policy tools to mitigate the misuse of open-weight models.