China proposes AI export controls with a fast track for open models

What could require a license

The draft gives two grounds for controlling exports of large models. One covers models trained on China’s “important data,” as defined by its data security law and related rules, when they provide services based on that data. Data lawfully available to the public is exempt.

A model may also use a small amount of important data—up to 0.1%—if it is cleaned beforehand, does not appear in the model’s answers and does not significantly expand research and development capabilities in relevant key technologies.

The second ground covers key model technologies: weights, checkpoints, source code, architecture, and training and inference methods. Controls apply if a model meets any of three thresholds:

Export authorities would determine top-10 status using public rankings from major Chinese and foreign evaluators or industry platforms. They could consider overall performance, results on specific tasks, safety and the number of users worldwide.

Tools that autonomously create or modify malware for attacks on critical infrastructure, military sites or government agencies would face a full export ban. Some software vulnerability-finding tools would instead face restrictions.

Technologies already published openly, including in open-source projects or academic papers, are exempt from the new rules.

Open access, with a faster route out

The draft also creates “automatic licensing” for publishing restricted AI technologies and model weights openly. The same fast-track process would cover Chinese companies building data centers abroad, designing chips overseas, conducting research with foreign partners or helping develop international standards. Exports that meet the conditions would be approved once the company submits its documents.

Use of controlled models through APIs, apps, websites, SaaS services or lightweight versions on local devices would remain outside export controls. Installing the technologies on local servers abroad would require a regular export license. Companies could still use the fast-track process if the software runs in their own cloud overseas.

China also says it will support AI development in the Global South and ease exports for capacity-building, technology exchange and joint work on rules through the World Artificial Intelligence Cooperation Organization, which China established as a counterweight to Western AI organizations. Until the final rules take effect, companies must register affected exports through the commerce ministry’s portal.

The tension in the rules

I think the most consequential distinction is not simply between open and closed models, but between making model technology available and shipping it in a form that can be installed abroad. Open publication gets a fast track, while a local server deployment can still require a regular license. That leaves room for distribution, but makes the route to independent operation less predictable.

The draft could make US companies using open-weight models more likely to choose domestic alternatives. European companies face a sharper problem: their digital-sovereignty strategy has relied on running open models in local clouds, and some had planned to host top-tier closed models on local servers. That strategy is harder to sustain if the strongest open models come only from the US or China and exports can be restricted.

What I’d want to know is how consistently the authorities will apply the top-10 test, and how much uncertainty companies will face before a model is ranked. The draft allows officials to draw on several kinds of public rankings and measures, but does not identify a single benchmark. The fast track may speed up approved exports; it does not remove the uncertainty over which models qualify for control in the first place.