Nathan Lambert: Chinese open-weight models lead US by 1.6B downloads and months on benchmarks
Interconnects (Nathan Lambert) · rss · 2026-09-21
Nathan Lambert shares his congressional briefing on the state of open models: Chinese labs have clearly led open-weight AI since April 2025.
The landscape
- Open-weight models (Llama, Qwen, Gemma, DeepSeek) release weights under licenses; true open-source models with training code/data are led by US nonprofits (AI2's Olmo, EleutherAI's Pythia).
- China passed the US in Hugging Face downloads in July 2025 and now leads by 1.6B (3.2B total, double America's).
- On the Artificial Analysis Index, GLM-5.3 (45), Kimi K3 (44), and GLM-5.3-Flash (42) far outpace America's best open models (Inkling at 26, Nemotron 3 Ultra at 23). The top US open models trail 15 Chinese models.
- Chinese open-weight models sit 2-5 months behind the US closed frontier; US open models trail OpenAI/Anthropic by 6-9 months. China is closest on agentic coding.
Why and what's next
- Chinese labs ship faster with narrower task focus; in 2026 top labs began buying frontier RL environment data from US and Chinese startups.
- Even if distillation were fully blocked, Lambert estimates the gap would only widen by 1-2 months.
Policy: restricting Chinese open models would hurt US businesses more (HuggingFace needed a Chinese open model to analyze a cyberattack); he argues the US should keep investing in open models to manage globally shared risks.
More from AGI Musings
- Lego Bionicle experiment: why AI researchers chase hard problems, not just big paychecks — c_valenzuelab · 2026-09-21
- 18 months ago his 'hundreds of arXiv papers a year' prediction drew gasps; he still underestimated AI — prof_g · 2026-09-21
- AI slop is the direct result of bad models on hard tasks, says Abacus.AI CEO — bindureddy · 2026-09-21
- Casual prompts can't measure AI: data contamination breaks amateur evals — iamKierraD · 2026-09-21
- Uncle Bob on AI risk: install guardrails but push as hard and fast as we can — mark_k · 2026-09-21
- Molecule team to discuss agentic AI reshaping healthcare value chain at Roche's RoX Health event — melnykowycz · 2026-09-21