China Moves to Put Export Controls on Model Weights and 23 AI Technologies

China has drafted export controls covering 23 categories of artificial intelligence technology, a list that for the first time reaches into the artifacts that make modern models usable: trained weights, source code and the training methods behind them. The thresholds attached to the draft are large — systems trained above roughly 10^26 FLOPs or built with 2 trillion parameters or more would fall under the rules. That is frontier scale, not the mid-sized models most companies ship.

The draft does not shut the door on open-weight releases. Under a fast-track procedure, open-source models can still be cleared for release once they are registered, which keeps Chinese labs in the business of publishing weights rather than locking everything behind an approval queue. The distinction matters because China's open-weight ecosystem — DeepSeek, Qwen, Kimi, GLM and others — has become a default starting point for developers worldwide who want capable models without paying per-token API rates.

What the controls would change is the paperwork and the timing. Registration introduces a step between a model finishing training and its weights going public, and it gives Beijing a lever over which systems travel abroad and under what terms. For labs weighing whether to publish at all, that added friction could shift release decisions even when approval is routine.

The second-order effect lands on buyers. Western developers and enterprises that have leaned on Chinese open-weight models for cost reasons would face a supply that is legally and practically less predictable. The likely outcome is a push back toward US open-weight families — Llama, Mistral, Gemma and similar lines — as teams reduce exposure to a single regulatory regime they do not control.

China is not alone in treating model weights as an export-controlled item. US rules already restrict the flow of advanced chips and, in narrower cases, model access to certain countries, so both governments now treat frontier AI as strategic goods rather than ordinary software. The difference is direction: Washington has focused on compute and closed-model access, while Beijing is drafting language around the weights and training recipes that underpin its most influential exports.

The draft is not final, and the 10^26 FLOPs and 2 trillion parameter lines could move before it is. The question worth tracking is how the registration lane is administered in practice — whether approval is quick and predictable, or slow enough that Chinese labs quietly stop publishing at the frontier. Watch the first few model launches after the rule takes effect; they will show whether open weights from China remain a reliable option or become a case-by-case exception.