China’s open‑source AI surge

The U.S.-China Economic and Security Review Commission published a staff research paper titled "Two Loops: How China’s Open AI Strategy Reinforces Its Industrial Dominance" on March 23, 2026, authored by Senior Policy Analyst Ngor Luong. (uscc.gov ) The paper documents that Alibaba’s Qwen family had become the largest model ecosystem on Hugging Face by the report’s publication, with the commission citing “over 100,000 derivatives” built on Qwen variants. (uscc.gov ) The report distinguishes a digital/open-model cycle from a physical/deployment cycle—arguing U.S. chip-focused export controls largely constrain the former but do not directly address model improvements driven by mass deployment and industrial data collection. (uscc.gov ) Reuters coverage cited in syndication notes U.S. export restrictions on advanced chips began in 2022 while Washington approved exports of Nvidia’s second‑most‑advanced accelerator in December 2025, illustrating partial relaxation in hardware controls. (yahoo.com ) The syndicated reporting and the USCC paper cite an industry estimate that roughly 80% of U.S. AI startups now build on Chinese open‑source models, a statistic the commission used to highlight rapid global adoption driven by cost and availability. (yahoo.com ) Hugging Face usage and community activity corroborate heavy adoption of Chinese base models—Qwen’s public organization page and model cards show large followings and numerous community uploads across Qwen families. (huggingface.co ) The narrative places recent Chinese open releases—Moonshot’s Kimi K2.5 among them—alongside Alibaba and MiniMax as examples of widely distributed, open‑weight models that industry coverage says have led spikes in global downloads and downstream forks. (techcrunch.com )