In a summary text adapted from a presentation to U.S. lawmakers and congressional staff, a researcher who helped create the Allen Institute for AI’s Olmo models provides an overview of open-weight AI models. According to him, Chinese labs are now the leaders, both in performance rankings and in usage.
Open-weight models make their parameters publicly accessible for examination or reuse. They are distinct from “closed” models like GPT-4 or Claude Opus 4.5, which are only accessible via an API, an interface for querying the model remotely. Open-weight models include Meta’s Llama, Alibaba’s Qwen, Google’s Gemma, and those from DeepSeek. So-called “open source” models go further by also publishing the code and training data. The text cites the Allen Institute for AI’s Olmo models, OpenAthena’s Marin, and EleutherAI’s Pythia in this regard.
A Chinese Lead in Downloads and Rankings
According to the author, China took the lead in Hugging Face downloads in July 2025, primarily thanks to Qwen models. Its lead amounts to about 1.6 billion downloads, for a total of 3.2 billion, which is twice the U.S. total. On the Artificial Analysis Intelligence Index (AAII) as of September 14, 2026, the top three Chinese models are GLM-5.3 (score of 45), Kimi K3 (44), and GLM-5.3-Flash (42), from the publisher Z.ai and Moonshot AI. The top U.S. models, Thinking Machines’ Inkling and Inkling Small, score 26, and Nvidia’s Nemotron 3 Ultra scores 23. The text estimates that Chinese models have a 2 to 6 month lead over their U.S. equivalents.
The OpenRouter platform, which allows switching between models, shows a similar trend. According to its data, the volume of tokens processed by open models has risen from about 1 T per week in September 2025 to about 80 T today, and the share of Chinese models has increased from about 70% to over 80% of usage.
In the corporate sector, several players rely on Chinese models: Harvey, specialized in law; Cursor, in code; DoorDash, which uses Kimi models; Airbnb, which uses Qwen; and Perplexity, which uses DeepSeek. In research, the author analyzed papers from five major arXiv categories dedicated to machine learning. The share of publications mentioning an open model rose from 2% in January 2023 to 50% in September 2026. Qwen appears in about 30% of papers, and Llama in about 21%.
A Gap That Is Narrowing, According to the Author
The author assesses distillation, a technique involving training a model on the outputs of a more powerful model, as giving Chinese labs a 1 to 2 month lead. He also estimates that comprehensive prevention, for example using customer identity verification tools at Anthropic and OpenAI, would only widen the gap by 1 to 2 months. He further notes that Chinese labs are now purchasing more training data, particularly reinforcement learning environments for agentic tasks.
The text also discusses risks, notably in cybersecurity, exemplified by the incident involving OpenAI and Hugging Face. It reports that Hugging Face used an open-weight Chinese model to analyze a cyberattack, as closed models had refused these requests. The author argues that the United States should continue to support investment in open models to better anticipate these risks and reduce U.S. companies’ dependence on models designed in China. He acknowledges, however, some unknowns: data is lacking on usage outside the U.S. and China, including in tracking open U.S. models, which he launched in August 2025 under the name American Truly Open Models (ATOM) Project.
