Chinese open-weight models reach 3.2B Hugging Face downloads, double US total as of September 2026

Chinese firms overtook US open-weight releases in July 2025 on Hugging Face download volume. Qwen and subsequent GLM releases drove the shift, reaching 3.2B total downloads against 1.6B for US models tracked in the ATOM Project. Licenses and inference stacks remain comparable, yet release cadence and benchmark iteration favor Chinese labs by two to six months on agentic tasks. Benchmark gaps widened after December 2025. GLM-5.3, GLM-5.3-Flash, and Kimi K3 posted AAII scores of 45, 42, and 44; US models released June-July 2026 sit at 26 and 23. Download leadership and capability metrics now align, confirming sustained Chinese dominance rather than isolated spikes. Open-weight access creates asymmetric leverage in US-China competition. US policy focus on closed APIs leaves downstream fine-tuning and inspection advantages to Chinese weights. True open-source releases remain US-led via Allen Institute Olmo and EleutherAI Pythia, yet these trail on commercial adoption. Continued Chinese iteration at current pace projects further separation on agentic benchmarks by mid-2027. US responses will hinge on whether additional data or training code disclosures occur before the next model cycle closes.