The current balance of power in open models

Nathan Lambert published an expanded version of testimony he prepared for a briefing to Congressional members and staff on the state of open-weight models in the context of U.S.-China competition. In the Interconnects post, dated September 21, 2026, he framed the remarks as a state of the union on open models.

Lambert distinguishes open-weight models, such as Meta's Llama, Alibaba's Qwen, Google's Gemma and DeepSeek's models, from true open-source models, which also include training code and data needed for reproduction. He writes that the most prominent open-source models have been built in the United States, led recently by the Allen Institute for AI's Olmo models, which he says he helped build over 2.5 years there, along with OpenAthena's Marin and EleutherAI's Pythia models. He describes these categories as existing on a spectrum.

On competition, Lambert writes that America was the early leader in open language models mainly through Meta's Llama, but that Chinese open-weight models surpassed American ones about 18 months ago. He cites Hugging Face Downloads, where China took the lead in July 2025, largely through Qwen, and says China's download lead has grown to about 1.6B, with a total of 3.2B downloads, twice America's total, since he first published the American Truly Open Models (ATOM) Project in August 2025.

On the Artificial Analysis Intelligence Index as of September 14, 2026, he lists the top three Chinese models as Z.ai's GLM-5.3 and GLM-5.3-Flash and Moonshot AI's Kimi K3, with scores of 45, 42 and 44. The leading American models, he writes, are Thinking Machines' Inkling and Inkling Small, both scoring 26, and Nvidia's Nemotron 3 Ultra at 23. He states the top American models were released in June and July 2026 and are updated less frequently than Chinese counterparts.

Lambert estimates Chinese open-weight models are roughly 2-5 months behind the closed American frontier, while American open-weight models are about 6-9 months behind OpenAI and Anthropic. He attributes Chinese labs' output partly to faster releases and narrower task focus, calling the reasons an open debate. He estimates that fully preventing distillation, for example through know-your-customer tools at Anthropic and OpenAI, would widen the gap from the strongest American models to Chinese open-weight models by only 1-2 months.

On adoption, he cites OpenRouter data showing open-model usage growing from about 1T tokens processed in a week of September 2025 to about 80T tokens per week, with Chinese models rising from about 70% to over 80% of usage. He names Harvey, Cursor, DoorDash, Airbnb and Perplexity as companies building on Chinese models. He reports that scanning five popular arXiv ML categories showed mentions of any open model rising from 2% in January 2023 to 50% in September 2026, with Qwen mentioned in 30% of papers, any Chinese open-weight model above 40%, and U.S. models around 30%.