China wants to control exports of any AI model over 2 trillion parameters

China wants a say over which of its most powerful AI models leave the country. Its commerce and science ministries on Friday proposed adding 23 AI technologies to the national list of technologies banned or restricted from export, covering large language models, AI agents, computing clusters and AI chip design, and ordered companies to register any such exports from today.

The draft, Announcement No. 48 of 2026 from the Ministry of Commerce and the Ministry of Science and Technology, covers 40 technologies in all, adding 5 in biomedicine and 12 in solar manufacturing. Comments are open until November 15. A ministry spokesperson said the revision is routine and "does not target any specific country, region or entity".

Any model over 2 trillion parameters

The heart of the AI list is a restricted category for the "core technology" of large models: weights, key checkpoints, underlying source code, architecture designs, and training and inference optimisation techniques. It applies to any model that meets one of three tests:

• Compute: trained with 10²⁶ floating point operations or more.
• Size: 2 trillion parameters or more.
• Rank: capability at a "globally leading level", which the draft defines as a top 10 place in public global rankings by mainstream evaluators or big industry platforms when the export happens.

That would catch the giant models China's labs are now planning. DeepSeek wants to train an 8 trillion parameter model on Chinese chips, and Alibaba has said it is planning a 10 trillion parameter Qwen model. The draft does not name any company or model.

Models trained on China's legally defined "important data" are also restricted, unless that data made up no more than 0.1% of training and was cleaned so it cannot resurface in outputs. Models trained on "core data", a stricter category under China's Data Security Law, would be banned from export outright.

Agents and AI hacking tools

The list also sets numerical bars for AI agents. Restricted items include planning systems that can break a plain-language instruction into 20 or more steps with at least 80% accuracy and 90% success, setups that run 60 or more agents together, and tool-use systems that work with 200 or more third-party tools and pick the right one at least 90% of the time.

The strictest items are about cyberattacks. The draft bans the export of AI that writes and mutates malicious code for attacks on critical infrastructure, defence or government systems, where it beats multi-engine malware detectors 60% of the time or a single mainstream antivirus product 85% of the time. AI-assisted vulnerability hunting would be restricted if it can, for example, build complete working exploits for known flaws in mainstream operating systems or browsers at least half the time, or autonomously find five or more undisclosed high-severity flaws in a major open-source project or operating system in one go.

The list also covers AI chip designs needing 14 nanometre or finer manufacturing or 2.5D and 3D packaging, and the networking behind clusters of more than 10,000 AI accelerators.

Open source, APIs and apps are left out

The draft excludes anything already in the public domain through open-source releases or published papers. The spokesperson said the controls are "precise and limited" and that "the vast majority of AI technologies can still be exported freely" once they take effect (MadRobot's translations).

Crucially for Chinese AI apps abroad, the spokesperson said offering a restricted model through an API, an app or website, a SaaS product or a lightweight on-device version does not count as a technology export. Deploying a restricted model on servers abroad does, though deployment on a company's own overseas cloud would get an automatic licence. So would open-sourcing AI technology, running Chinese-owned data centres overseas and making chips at foreign plants, which the ministry said would be cleared almost "as soon as reported".

Until the final list takes effect, companies must register on the commerce ministry's platform before transferring ownership of any listed technology abroad, licensing it to foreign users, or passing control of its owner to a foreign entity.

"Under human control"

The spokesperson said the aim is to "hold the security bottom line, prevent misuse and abuse, and ensure that AI always remains under human control", echoing what Xi Jinping told Donald Trump in September. The ministry pointed out that "all major economies" already control these technologies to some degree. In June, the US Commerce Department restricted exports of Anthropic's Mythos 5 and Fable 5 worldwide.

Why it matters

China has built its AI reputation on giving models away, and this draft keeps that door open while putting the frontier behind a licence. If it passes as written, Chinese labs that keep their biggest models closed will need Beijing's approval to hand the weights to anyone abroad, much as US labs now need Washington's.

Sources: Ministry of Commerce and Ministry of Science and Technology, Announcement No. 48 of 2026 and its annex; Ministry of Commerce spokesperson Q&A.