A draft update to China’s national catalog of restricted exports has been circulated by the country’s commerce and science ministries, and this version reaches straight into the open-weight model ecosystem.
Spanning large models, agents, compute clusters, and data processing, the list covers 23 AI technologies, with the sharpest edges falling on two classes of model technology.
Where the lines fall
Data is the first trigger. Controls apply to models trained on what Chinese law defines as important data when services are offered on top of it. Lawfully public material is exempt, and a narrow allowance exists – up to 0.1 percent – for cleaned data that never surfaces in outputs and does not meaningfully boost core research capacity.
A second trigger takes in the core machinery behind large models. Architecture, source code, weights, checkpoints, and training or inference methods all fall under it. Three thresholds bring the controls into play – 10^26 floating-point operations of training compute, at least 2 trillion parameters, or a global top-ten ranking. Officials would settle that last status using published rankings from major evaluators and industry platforms.
A fast lane built in
An automatic licensing path was paired with the draft, so Beijing avoids strangling its own open-source scene. Publishing restricted technology as open source clears the moment paperwork is filed, as does international standards participation and overseas chip work by Chinese firms.
APIs, apps, websites, and lightweight on-device versions keep restricted models outside the rules. A standard license is required, though, to install them on servers abroad.
For everyone outside China the draft matters most. Hosting open-weight models in domestic clouds has been Europe’s sovereignty play, and that plan weakens when the strongest open models come from two countries able to cut off access. Already-public research and open-source releases skip the new rules.
