Interconnects AI · 2026-09-21 · notable
Nathan Lambert — Chinese open models get twice the downloads of US ones
Nathan Lambert briefed members of Congress and their staff on open-weight models, then published the remarks. Chinese open models reach 3.2 billion downloads to America's 1.6 billion, and hold over 80% of open-model usage on OpenRouter.
Chinese labs now lead open-weight AI on downloads, capability scores, research citations and commercial usage.
What is it?
Chinese open-weight models lead American ones on every measure Nathan Lambert tracks, and this post sets out the counts side by side. Lambert wrote it after being invited to brief a group of Congressional members and staff on open models through the lens of U.S.–China competition, then expanded the remarks for a general audience.
How does it work?
The comparison rests on four separate counts rather than one. Downloads: 3.2 billion for Chinese models against 1.6 billion for American ones. Capability: GLM-5.3, Kimi K3 and GLM-5.3-Flash score 45, 44 and 42 on the index Lambert cites, against 26 for Thinking Machines' Inkling and Inkling Small and 23 for Nvidia's Nemotron 3 Ultra. Research: any Chinese open-weight model is named in over 40% of arXiv papers, against 30% for American ones, with Qwen at 30% and Llama fallen to about 21%.
Why does it matter?
The usage numbers are the part that is hard to argue with: Lambert puts Chinese models above 80% of open-model traffic on OpenRouter, up from around 70%, and at roughly 95% or higher on OpenCode. He names Harvey, Cursor, DoorDash, Airbnb and Perplexity as American companies building on Chinese weights. On the gap to the closed frontier, he estimates Chinese open models are 2–5 months behind and American open models 6–9 months behind.
Who is it for?
people tracking open-weight models and AI policy