China’s commerce and science ministries have proposed new export controls for AI technologies, including advanced models, AI agents and tools that can create or alter malware. The draft also outlines ways eligible open-source releases and other exports could receive automatic approval, making the proposal both a restriction framework and a licensing system.
What the draft would cover
The proposal updates China’s list of banned and restricted export technologies. It identifies 23 key AI areas across large models, agents, compute clusters and data processing.
Some cyber capabilities would face particularly strict treatment. AI tools able to write or alter malware on their own to attack critical infrastructure, the military or government agencies would be banned from export. Certain tools for finding software vulnerabilities would face export limits rather than a total ban.
Large models could fall under controls in two ways. One concerns models trained on China’s “important data” that provide services based on that data. The source describes important data by reference to the country’s Data Security Law and related rules.
Lawfully public data is exempt. The draft also allows amounts up to 0.1 percent when the data has been cleaned beforehand, does not appear in AI outputs and does not substantially improve research and development capacity for core technology in the field.
Thresholds for model technology
A separate set of controls concerns the technology behind large models. The draft lists weights, checkpoints, source code, architecture, and training or inference methods.
Those controls would apply if a model reaches any one of three thresholds: 10^26 FLOPs of training compute, at least 2 trillion parameters, or a place among the global top 10. Export officials would assess top-10 status using public rankings from major domestic and foreign evaluators or industry platforms.
The ranking assessment could consider overall performance, specific tasks, safety or global usage numbers. That gives the proposal more than one way to identify a model as covered: a system could meet a technical threshold or qualify through its position in public evaluations.
Dean W. Ball, head of strategic futures at OpenAI, writes that the move does not mean China is shutting down open-weight models, but suggests the government is creating the option to restrict future models. He describes the framework as a “mirror image” of US model-weight export controls.
How open-source and overseas use fit in
The draft includes a fast-track process intended to avoid slowing China’s domestic open-source activity. Publishing restricted AI technology and weights as open source would qualify for “automatic licensing.” The same process would apply to certain activities by Chinese firms, including running their own data centers abroad, taping out chips overseas, conducting research with foreign partners or helping set international standards.
For eligible exports, approval would take effect when a company files the required paperwork. The proposal also distinguishes between ways of accessing a model and installing it. API calls, apps, websites, SaaS tools and lightweight versions on local devices would remain outside these export controls.
Installing restricted AI technology on local servers abroad would require a standard export license. A company could use the streamlined process if the software runs on its own overseas cloud.
Technology already made public through open-source releases or research papers is exempt from the new rules, according to the source. The distinction means publication status could matter alongside model capability and deployment method.
Why the proposal matters beyond China
Export controls can affect who can obtain or deploy a model, even when the model itself is open weight. The article notes that US companies relying on open-weight AI may turn more toward domestic models instead of Chinese alternatives.
It also points to a challenge for Europe, whose strategy has included running open-weight models in domestic clouds to support technological independence, while companies host closed top-tier models on local servers. If access to leading open models from the US or China can be interrupted by export controls, that approach may be harder to sustain.
China says it will support AI development in the Global South and ease exports for capacity building, technology sharing and joint rulemaking within the World Artificial Intelligence Cooperation Organization. Until final rules take effect, companies must register affected exports through the commerce ministry’s portal.