China moves to tighten export controls on open-weight models and 23 AI technologies

The draft sets restrictions on 23 AI technologies. Core large-model techniques, weights, checkpoints, source code, architecture, and training/inference methods trigger controls when they reach 10^26 FLOPs of training compute, 2 trillion parameters, or a global top-10 ranking. Models trained on and serving China's important data are also restricted, though lawfully public data or data under 0.1% share that has been cleaned is exempt. AI tools that can autonomously write or modify malware to attack critical infrastructure, military, or government agencies face a full export ban, while vulnerability-discovery AI tools face new export restrictions. To protect the domestic open-source ecosystem, the draft establishes a fast track: releasing restricted AI technologies and weights as open source can receive automatic licensing, with companies completing approval simply by submitting materials. Using restricted models via API, apps, websites, SaaS tools, or lightweight local devices is not controlled, but installing restricted AI technology on overseas local servers requires a standard export license. AI technologies already released as open source or published in papers are fully exempt. For overseas teams relying on Chinese open weights, obtaining the latest weights and training methods will require export approval once controls take effect, potentially forcing a shift to U.S. open-source models. The EU's strategy of maintaining tech sovereignty by running open models on domestic clouds could face challenges if the best open models come from the U.S. and China and can be cut off at any time. Users of this product are not affected through API calls for now, but self-hosted deployments need to assess compliance costs.