Andy Garcia · freelance dev
FRENES
Back to the blog
AI & Regulation

China Drafts Export Controls for AI Models of 2 Trillion Parameters or More. Weights Already Open Are Exempt, but the Next Release Would Need a Licence

October 19, 2026 7 min read

On Friday, October 16, China's Ministry of Commerce and Ministry of Science and Technology published Announcement No. 48 of 2026. It proposes adding 40 technologies to the country's catalogue of technologies prohibited or restricted from export, across artificial intelligence, biomedicine and solar. Twenty-three are AI technologies: large models, agents, computing clusters and information processing. Across the whole list, 14 would be banned from export and 26 restricted. The public can comment until November 15.

A temporary rule took effect the same day. Before exporting any listed technology, an exporter must register on a Ministry of Commerce platform, and deals that have started but are not finished are included.

The Associated Press reported that models with 2 trillion parameters or more could be restricted, and that Moonshot's Kimi K3 and Alibaba's Qwen 3.8-Max are among the Chinese models above that line. The draft itself names no model. The ministry's stated aims, in AP's translation: "reinforcing the security bottom line, preventing misuse and abuse, and ensuring that AI remains under human control."

I went to the annex and to the ministry's Q&A, because for anyone who buys models the line that matters is not the threshold. It is the exemptions.

Why I read it

I am CTO of a LegalTech built on Claude. In June, the US Commerce Department asked Anthropic to bar foreign nationals from using Fable 5 and Mythos 5, citing national security. AP recalled on October 16 that those restrictions were later lifted. I wrote on LinkedIn at the time that model risk had become the new vendor lock-in, and that I now treat a second provider as part of the architecture. Open weights are one place a second provider can come from.

The US move was about who may call a model. China's draft is about something else: where a model's weights may go.

What the draft controls

Both ministry documents are in Chinese, so the quotes from them below are translations.

The restricted large-model entry covers a model's "core technologies": weight parameters, key checkpoints, underlying source code, architecture design, and training-optimisation and inference-acceleration techniques. It applies to models that meet any one of three tests: at least 10^26 floating-point operations of compute, at least 2 trillion parameters, or capability "at a globally leading level".

That third test has a definition. It means a global top-10 ranking, at the time the export takes place, in public evaluations by mainstream evaluation bodies in China or abroad, or by large industry platforms. The criteria listed include overall performance, specialised capabilities, safety, training efficiency and "global call volume". The draft does not say which ranking counts, or whether one is enough. On that wording, a model that becomes popular with developers abroad could move into scope after its release.

Two more provisions concern training data. Models trained on China's "core data" and then offered as a service would be banned from export. Models trained on "important data" would be restricted, with carve-outs for trace amounts (0.1 percent or less, under further conditions) and for data that can be lawfully obtained in public.

The exemptions are the story

Three sentences change how the threshold reads.

First, the annex says technologies that have already entered the public domain through open source or academic papers are outside its scope.

Second, the ministry's spokesperson said that offering a restricted model through API calls, app or web interaction, SaaS, or "local on-device lightweight deployment" is not a large-model technology export.

Third, open-sourcing AI technology from now on gets "automatic licensing and other facilitation measures", which the ministry describes as "basically report and export immediately". The same treatment applies to Chinese companies' own computing centres abroad, chip tape-outs abroad, joint research using foreign resources and work on international standards.

Now look at the two models AP named. Moonshot's Hugging Face card for Kimi K3 lists 2.8 trillion total parameters and says the company releases "the full Kimi K3 model weights" under its own Kimi K3 License. Alibaba's Qwen3.8-2.4T-A95B repository holds a model with 2.4 trillion parameters, 95 billion of them active, and its card calls Qwen3.8-Max "the official version" built on those weights, with extra features such as vision input. On my reading of the annex, the weights already published are outside the draft's scope. API calls to either model are outside it too.

What changes is the next one. If the draft is adopted as written, a future release that crosses any of the three tests becomes a declared export before it becomes a download. Automatic, by the ministry's own description. But an automatic licence is still a step that exists, and this is a draft: the thresholds and the facilitation measures are both open for comment until November 15. Because key checkpoints are on the list, I would not assume that a new checkpoint of an already-public model inherits its exemption.

The private-deployment line

The ministry was asked directly about Chinese models going abroad. Its answer: for restricted AI technology, "overseas localised deployment" falls within export control. If the deployment sits on the company's own cloud abroad, where risk is "relatively controllable", automatic licensing can apply.

Put that next to the API exemption and the gradient is clear. If a European company calls a covered Chinese model over the API, nobody needs Beijing's permission. If it wants a covered model installed on its own hardware, so client documents never leave its perimeter, the vendor needs an export licence. The option a residency-conscious buyer prefers is the one that gets the paperwork.

The temporary registration already applies. The cases it lists include transferring ownership of a technology abroad and licensing a domestic technology for use abroad. The form asks for the exporter and the recipient, the destination country or region, the technology and its use, and the expected value. A private deployment licensed to a foreign customer reads to me like the second case. If you are negotiating one, your company's name, country, use case and deal size may well end up on that form. The ministry adds that exporters should judge for themselves whether a technology falls within the registration scope.

Agents have numbers too

The agent entry is unusually concrete. Restricted agent technology includes:

  • planning that breaks a natural-language instruction into executable chains of 20 steps or more, with at least 80 percent decomposition accuracy and a 90 percent average subtask success rate;
  • multi-agent architectures running 60 or more concurrent agents, with conditions on speed and accuracy;
  • tool use across 200 or more third-party tools, with at least 90 percent tool-selection accuracy and 90 percent end-to-end success on tasks that need three tools or more;
  • AI game-decision technology that, among other conditions, plans 20 steps deep or runs 10 game agents at once, decides in 200 milliseconds or less, and beats 90 percent of humans on a standard benchmark.

The ministry also says products made using a listed technology are not covered. On that logic, a Chinese agent product sold abroad as a service would stay outside; transferring the agent stack itself is what the entry targets. I orchestrate my own work with Claude Code, and these thresholds read like the numbers in an agent launch post. The draft names no benchmark for any of them.

Elsewhere on the AI list, AI-driven generation and mutation of malicious code would be banned from export, and AI-assisted vulnerability discovery restricted.

What I would do before November 15

List every Chinese model in your stack and how you reach it. Through an API, an app or SaaS, the ministry says this is not a technology export.

For open weights you already run, keep the exact version you validated in your own storage. The exemption covers what is public; a fallback should not depend on a re-download.

If a plan counts on the next open release of a Chinese frontier model, add a licensing step to its timeline. Automatic today is a policy, not a guarantee.

If you are buying a private deployment from a Chinese vendor, ask in writing whether they have registered under Announcement No. 48, and whether your setup counts as their own overseas cloud or needs a full licence. Tie delivery to the answer.

And if this affects you, say so. Comments on scope, thresholds, control policies and facilitation measures go to jinxianmulu@mofcom.gov.cn, or by fax or letter, until November 15.

The API stays open. The draft is about where the weights go next, and who has to be asked first.

Sources

A project like this one?

I design and deploy products like this. Let's talk.

Let's talk