I was recently invited to brief Congressional members on the state of open-weight models in U.S.-China competition. Open language models have publicly available weights for inspection or downstream use, contrasting with closed models accessed via APIs like GPT-4 or Claude Opus 4.5. Open-weight models, such as Meta’s Llama, Alibaba’s Qwen, Google’s Gemma, and Deep Seek’s models, are governed by licenses and often include inference code.

Since April 2025, Chinese AI companies have led in open-weight models. True open-source models include weights, licenses, inference code, training code, and training data, with prominent examples from U.S. nonprofits like the Allen Institute’s Olmo, Open Athena’s Marin, and Eleuther AI’s Pythia. Open and closed models exist on a spectrum; Nvidia’s Nemotron releases training data but not all data, while closed models vary in API transparency.

Chinese models like GLM-5.2 and Kimi K3 have achieved commercial viability and agentic capabilities comparable to Anthropic’s Claude Code. America led early via Meta’s Llama, but Chinese open-weight models surpassed U.S. counterparts in key areas 18 months ago. Hugging Face Downloads show China leading since July 2025, driven by Alibaba’s Qwen, with a download lead of 1.6B out of 3.2B total—twice America’s.

On the Artificial Analysis Intelligence Index (AAII), top Chinese models as of September 14, 2026 are Z.ai’s GLM-5.3 (45), GLM-5.3-Flash (42), and Moonshot AI’s Kimi K3 (44). Leading American models—Thinking Machines’ Inkling and Inkling Small (26 each) and Nvidia’s Nemotron 3 Ultra (23)—lag behind, released less frequently and updated later than Chinese counterparts.