China wants a license before its biggest AI models leave the country. Using them through an app stays free.
Beijing published a draft on October 16 that would restrict exports of model weights and chip designs. The wording leaves the API, the app and open-source code alone.
China’s Ministry of Commerce and Ministry of Science and Technology published a draft on October 16 that would put 40 technologies on the country’s list of restricted or prohibited exports. Twenty-three of them are AI. That’s the number the headlines picked up, and it’s the least interesting part.
The more useful question is what actually has to be stopped at the border. Going by the ministry’s notice (Announcement No. 48) and the reporting around it, the answer is narrower than “China restricts AI exports” suggests.
Files and designs, not services
The draft targets the thing itself: model weights, checkpoints, source code, architecture, training methods, chip designs. Bloomberg’s report says it covers technology for models with at least 2 trillion parameters, and it names Moonshot’s Kimi K3 and Alibaba’s Qwen3.8-Max as examples. It also covers models ranking in the world’s top 10 on performance, usage or safety measures.
What stays outside is anything a foreign customer touches through a front door. The Commerce Ministry said companies that let overseas users reach a restricted model through apps, websites or online services wouldn’t be subject to the controls. Cybernews read that the same way: people abroad could keep using services like DeepSeek’s. Open-source technology and research already in the public domain is excluded too, and the ministry promises streamlined procedures such as automatic licensing for open-source exports.
So a Chinese lab could keep selling API access to a US company. What it couldn’t do, without a license, is hand that company the weights to run on its own servers. Bloomberg adds that deploying a covered model overseas counts as a technology export, though deployments on infrastructure the Chinese provider itself owns abroad could qualify for automatic licensing.
The numbers, and where reports differ
Here is what each source says about the same draft.
| Item | What’s reported | Source |
|---|---|---|
| Technologies in the draft | 40 (23 AI, 5 biomedicine, 12 solar) | Ministry notice |
| Prohibited / restricted | 14 / 26 | Ministry spokesperson |
| AI items prohibited / restricted | 6 / 17 | CTOL’s count of the annex |
| Model-weight threshold | 2 trillion parameters, or 10^26 training operations, per CTOL | CTOL |
| Chip design threshold | 14-nanometer or more advanced, or 2.5D/3D packaging | Bloomberg |
| Comment period | October 16 to November 15 | Ministry notice |
Bloomberg’s wording on the model threshold reads as if a model must be both trained on “important data” and have 2 trillion parameters. CTOL, which read the annex, says a model is restricted if any one of three conditions is met. I couldn’t open the annex itself, so I’d trust the annex over either report. The final text is also still a draft, and the numbers can still move before the comment period ends.
The spokesperson’s own examples are worth a look. Prohibited items include AI-driven malicious code generation and mutation. Restricted items include AI-assisted vulnerability discovery. Those sit closer to the cyber worries the US labs keep raising than to a trade fight over chatbots.
Registration starts now
One piece isn’t a draft. From October 16, exporters have to register certain deals with the ministry before they go ahead: transferring ownership of a covered technology abroad, licensing it for foreign use, or letting a foreign party take ownership or control of the company that owns it. Deals already underway but not finished are included.
That last category matters for anyone watching Chinese AI startups take foreign money. A change of control now needs a registration. The ministry says the point is to understand demand and manage risk, and that the policy isn’t aimed at any country, region or entity.
What I make of it
I don’t think this changes much for a US developer this month. Nearly everyone outside China uses Chinese models through hosted APIs or by downloading open weights, and the draft leaves both alone. The open-weights point is the one I’d watch, because “already in the public domain” protects what’s out, not what comes next. A lab that wants to release the weights of a next flagship may have to ask first, and nobody outside the ministry knows how that will go.
The timing is awkward for the labs too. Bloomberg notes Alibaba said last month it plans models of 5 to 10 trillion parameters, which would sit well above the 2 trillion line.
For Western open-model efforts, the draft cuts both ways. Meta’s Muse Spark Open and Reflection’s Beam were pitched as the American answer to Chinese open weights. If the best Chinese weights get harder to obtain, that pitch matters more. If they stay available at a smaller size, it matters less. Both outcomes are plausible, and the earlier GLM 5.3 story shows why security officials already watch these releases closely.
The comment window closes on November 15. The final list is the thing to read, not this draft.
Sources
- Ministry of Commerce and Ministry of Science and Technology, Announcement No. 48 of 2026 (in Chinese)
- Ministry of Commerce spokesperson’s Q&A on the draft (in Chinese)
- Bloomberg via The Straits Times: China plans AI export curbs to exert control over key tech
- CTOL Digital: China’s draft export list controls model files and chip blueprints, not APIs
- Cybernews: China AI export controls target 40 technologies