On October 16, China's Ministry of Commerce and Ministry of Science and Technology released a draft revision of the Catalogue of Technologies Prohibited or Restricted from Export, proposing to add 40 technologies, 23 of them in artificial intelligence. Public comment is open until November 15, and related exports are already subject to registration management during the transition period.
1. Models themselves are covered for the first time. The weights, source code, checkpoints, architecture, and training and inference methods of large models are brought under export controls for the first time, along with agents, chip design and compute clusters. Controls previously targeted mainly underlying hardware; this time the trained systems themselves fall within scope.
2. Clear thresholds. Export of the core technology is restricted for models trained with 10^26 FLOPs of compute, with more than 2 trillion parameters, or ranked among the global top 10; models trained on important data also face restrictions. AI tools that generate or modify malware targeting critical infrastructure are listed as prohibited from export, with 14 technologies proposed for prohibition and 26 for restriction.
3. Everyday use is unaffected. Accessing a restricted model via API, SaaS, website or app does not count as export; technology already made public through open source or academic papers is excluded, and open-source releases may also receive automatic licenses. The draft is still at the public-comment stage, and the final scope may change.
Source: read the original. This text is a summary, for reference only.
