China’s commerce and science ministries have published a draft that expands the list of AI technologies subject to export controls. The draft covers 23 AI tools, including large models, agents, compute clusters and data‑processing software. Tools that can automatically write or modify malware aimed at critical infrastructure are banned from export. Other vulnerability‑finding tools will face new limits rather than a full ban.
How the model rules work
Two cases trigger export limits for large models. First, models trained on China’s “important data” – data defined under the Data Security Law – are restricted unless the data is public, cleaned to less than 0.1 % of the training set, never appears in model outputs, and does not materially boost R&D for core tech. Second, the underlying technologies – weights, checkpoints, source code, architecture and training or inference methods – are controlled once a model reaches 10^26 FLOPs of compute, has at least 2 trillion parameters, or ranks in the global top‑10 according to public evaluator rankings.
What is still allowed
Using restricted models via API calls, SaaS apps, websites or lightweight local versions remains free of the new controls. Exporting the same software to run on foreign servers, however, requires a standard export license. China also introduced a fast‑track “automatic licensing” for AI tech released as open source, and for Chinese firms that operate data centres abroad or collaborate internationally.
Why it matters
Companies that rely on open‑weight AI models may need to shift to domestic alternatives or secure export licences, adding cost and complexity. Researchers and developers can still use open‑source releases, but any commercial deployment abroad will face new paperwork.
