China’s commerce and science ministries have proposed tighter export controls covering 23 AI technology areas, including frontier models, agents, compute systems and data processing. The draft would ban exports of autonomous tools capable of creating or modifying malware to attack military, government or critical infrastructure targets.
Core model technology could face restrictions when a system reaches at least 2 trillion parameters, uses 10^26 floating-point operations for training or ranks among the world’s top 10. Covered material includes model weights, checkpoints, source code, architecture and some training or inference methods. Models trained on legally defined important Chinese data could also be restricted, with exceptions for public or small, cleaned portions of data.
The proposal does not amount to a blanket ban on open-weight AI. Technology that is already public would be exempt, and domestic developers could register an open release for automatic approval under a fast-track process. Export officials would use public rankings and industry evaluations to determine whether a model meets the top-10 threshold.
Because the text is a draft, its final wording and enforcement may change. If adopted broadly, it would give Chinese authorities a mechanism to stop selected future model releases while preserving routine publication. Developers outside China may need to consider export eligibility before building critical systems around unreleased Chinese models.