China has taken the lead in open-weight models, Nathan Lambert argues

Nathan Lambert's read on the open-model landscape, drawn from his congressional testimony, is blunt: Chinese labs have overtaken the US in open-weight models, reversing the lead Meta once held with Llama. He backs it with usage data rather than vibes. Chinese models account for roughly 3.2 billion downloads against 1.6 billion for US models, and their share of open-model traffic on OpenRouter has climbed from about 70% to over 80%. On OpenCode, a coding agent, close to 95% of open-model inference now runs on Chinese models.

The capability gap points the same way. Lambert cites top Chinese models such as GLM-5.3 and Kimi K3 scoring in the low-to-mid 40s on the Artificial Analysis Index, while leading US open models like Nemotron sit in the mid-20s, with Chinese labs shipping comparable models two to six months earlier. In research, Qwen alone shows up in about 30% of arXiv papers that name an open model, ahead of Llama's 21%. He notes the Chinese open models still trail the closed US frontier by two to five months, and argues distillation explains only one or two months of that gap, so copying is not the main story.

His recommendation to policymakers is to keep investing in US open models, both to hold domestic research leadership and to reduce US companies' reliance on Chinese models.

Why it matters

If you build on open weights, the default choice is increasingly a Chinese model, which carries real consequences for data governance, export rules, and long-term support. Lambert's numbers give you the concrete case to bring to that decision, rather than treating "US versus China" as an abstraction.