Nathan Lambert's congressional testimony: China leads the open-weight model race

Nathan Lambert of the Interconnects AI newsletter has published the expanded version of remarks he prepared to brief U.S. Congressional members and staff on the state of open-weight models amid U.S.–China competition, dated September 21.

Open-weight versus open-source models

Lambert separates open-weight models — weights released publicly under a license, as with Llama, Qwen or DeepSeek — from genuinely open-source models, which also publish the training code and data needed to reproduce them. The best-known open-source models come from American non-profits.

Chinese leadership in numbers

Chinese labs took the lead on Hugging Face downloads in July 2025, mainly through Qwen. Lambert's tracking puts China's cumulative lead at about 1.6 billion downloads out of 3.2 billion — twice the American figure. On the Artificial Analysis Intelligence Index, the top Chinese models as of September 14 were GLM-5.3 (45 points) and Kimi K3 (44), against 26 for Inkling and 23 for Nemotron 3 Ultra.

He estimates Chinese open-weight models trail the closed American frontier by two to five months, while American open-weight models run six to nine months behind the leading OpenAI and Anthropic systems. Distillation from stronger models is widespread in China, but blocking it entirely would widen the gap by only one to two months.

Adoption beyond benchmarks

On OpenRouter, open-model traffic rose from about one trillion tokens a week in September 2025 to roughly 80 trillion today, with Chinese models growing from about 70% of that usage to over 80%. The open-source coding agent OpenCode reports around 95% of its inference on Chinese models.

In academia, a scan of the five most popular machine-learning categories on arXiv found mentions of any open model rising from 2% of papers in January 2023 to 50% in September 2026.

Risk and the policy question

As open models approach levels that can enable new harms — Lambert points to cybersecurity and the July 2026 OpenAI–Hugging Face incident — restricting access to the strongest Chinese models would mainly set back the American businesses that depend on them. Hugging Face itself used a Chinese open-weight model to analyse the intrusion because closed models refused the requests. Managing the risk, he concludes, is a matter of ecosystem preparation.