China drafts export controls covering 23 AI technologies and large models

China's commerce and science ministries have released a draft update to the nation's list of banned and restricted export technologies. The catalog targets 23 key AI technologies across large models, agents, compute clusters and data processing.

The harshest rule hits offensive tools. AI tools that can write or alter malware on their own to attack critical infrastructure, the military or government agencies face a total export ban. Certain AI tools for finding software vulnerabilities would not be banned outright, but they would face new export limits.

Large models are controlled in two cases. The first covers models trained on China's "important data", as defined by the Data Security Law and related rules, that offer services based on it. Lawfully public data is exempt. Up to 0.1 percent important data is also allowed if the data was cleaned beforehand, never shows up in the model's outputs, and does not substantially boost the model's R&D capacity for core technology in that field.

The second case covers the core technologies behind large models: weights, checkpoints, source code, architecture, and training or inference methods. These controls kick in once a model hits 10^26 FLOPs of training compute, has at least 2 trillion parameters, or ranks in the global top 10. Export officials decide top-10 status from public rankings by major domestic and foreign evaluators or industry platforms, looking at overall performance, specific tasks, safety or global usage numbers. AI technology that is already public through open-source releases or research papers is exempt from the new rules.

To avoid choking off its domestic open-source boom, China pairs the controls with a fast track. Publishing restricted AI tech and weights as open source qualifies for "automatic licensing". So do Chinese firms that run their own data centers abroad, tape out chips overseas, conduct research with foreign partners, or help set international standards. Under the ministry's rules, eligible exports are cleared the moment a company files the paperwork.

Using restricted models through API calls, apps, websites, SaaS tools or lightweight versions on local devices stays free of these export controls. Installing restricted AI tech on local servers abroad requires a standard export license, though companies can still use the streamlined process if the software runs on their own overseas cloud. Until the final rules take effect, companies must register affected exports through the commerce ministry's portal.

Beijing also pledges to back AI development in the Global South and ease exports for capacity building, tech sharing and joint rulemaking within the World Artificial Intelligence Cooperation Organization, a bloc China launched to counter Western AI bodies.

Dean W. Ball, head of strategic futures at OpenAI, writes that the move does not mean China is shutting down open-weight models, but that it establishes the optionality to shut down future ones at will. He adds that Beijing's framework is a "mirror image" of the model-weight export controls the US government has already put in place. The Decoder's own assessment is that the rules will likely push US companies relying on open-weight AI toward domestic models rather than Chinese alternatives, and that Europe is in a tougher spot, since its strategy of running open-weight models in domestic clouds falls apart if the best open models come from the US or China and can be cut off.

Key facts

The draft catalog targets 23 key AI technologies across large models, agents, compute clusters and data processing; autonomous malware tools aimed at critical infrastructure, the military or government agencies face a total export ban.
Controls on core large-model technology (weights, checkpoints, source code, architecture, methods) apply at 10^26 FLOPs of training compute, at least 2 trillion parameters, or a global top-10 ranking.
Models trained on China's "important data" that offer services based on it are also controlled; lawfully public data is exempt, and up to 0.1 percent is allowed under conditions.
Open-source releases of restricted tech and weights qualify for automatic licensing, cleared the moment paperwork is filed; already-public tech is exempt, and API, app and SaaS use stays free of controls.
Installing restricted AI tech on local servers abroad needs a standard export license; until final rules take effect, companies must register affected exports through the commerce ministry's portal.

Why it matters

China is writing export controls for AI models into its national list of restricted technologies, using concrete triggers: 10^26 FLOPs of training compute, at least 2 trillion parameters, or a top-10 global rank. Dean W. Ball, head of strategic futures at OpenAI, calls the framework a "mirror image" of the model-weight export controls the US has already put in place, and says it gives the Chinese government the option to shut down future open-weight models at will. The fast-track for open-source releases shows Beijing is trying to protect its domestic open-source boom at the same time.

Who it affects

Chinese AI labs and companies that export models, weights, code or training methods; Chinese firms running data centers abroad, taping out chips overseas or doing research with foreign partners. Foreign users of Chinese open-weight models are affected too. The article's own assessment is that US companies relying on open-weight AI will likely lean toward domestic models over Chinese ones, and that Europe is in a tougher position because its sovereignty strategy depends on running open-weight models in domestic clouds.

How to use it

For affected Chinese companies, the draft sets out the routes. Publishing restricted tech and weights as open source, running your own overseas data centers, taping out chips abroad, doing research with foreign partners or helping set international standards qualifies for automatic licensing, cleared on filing. Access through API calls, apps, websites, SaaS tools or lightweight on-device versions stays free of the controls. Installing on local servers abroad needs a standard export license, but the streamlined process is available if the software runs on the company's own overseas cloud. Until the final rules take effect, affected exports must be registered through the commerce ministry's portal.

How solid is it

This is a draft update, not final rules; the source refers to the final rules as still to take effect. The details come from The Decoder's report on the draft. The Ball quote is his commentary, and the source gives only his title, not an official OpenAI position. The conclusions about US companies and Europe are the article's own assessment, not Ball's.

Risks and caveats

The draft may change before it becomes final. The source gives no date for the release, for the end of any comment period, or for when the final rules take effect, and it names no specific models that meet the thresholds. The three model triggers are alternatives, so a model needs to meet only one. The top-10 trigger depends on officials' reading of public rankings across performance, tasks, safety or usage. Ball stresses that the rules do not mean China is shutting down open-weight models today, only that the option now exists for future ones.

“This move does not mean 'China is shutting down open-weight models,' but it does indicate the Chinese government establishing the optionality to shut down future open-weight models at will”

— Dean W. Ball, head of strategic futures at OpenAI