China Proposed Export Controls on 23 AI Technologies, Including Weights for Models Over 2 Trillion Parameters. Hosted APIs Are Exempt. Here Is the Dependency Check I'd Run Before It Is Final.
On October 16, China's Ministry of Commerce and Ministry of Science and Technology published a draft that would add 40 technologies to the country's export restriction catalogue, 23 of them in AI. Public comment is open until November 15. If your product runs on Qwen, DeepSeek, GLM or Kimi because the weights are free and good, this is the first policy document that gives Beijing a lever over the next release, not the last one.
What the draft covers
Everything below comes from a Cloud Security Alliance research note on MOFCOM Announcement No. 48, cross-checked against press coverage of the same announcement. I could not retrieve the original Chinese text, so treat the details as secondhand until you read it.
Model weights, checkpoints, source code, architecture and training know-how would be restricted if a model meets any one of three tests: at least 10^26 FLOPs of training compute, at least 2 trillion parameters, or a global top-10 position on public rankings, call volume or similar measures. Agents are controlled separately, with thresholds such as 20-step subtask chains, 60 concurrent agents and 200 integrated tools. AI-generated malicious code is prohibited and AI-assisted vulnerability discovery is restricted.
Two things are explicitly outside the controls. Offering a model through an API, app, web page, SaaS product or lightweight on-device deployment is not treated as a large-model export. And technology already public, through open-source release or published papers, is exempt, so weights sitting on your disk are not affected.
What is not exempt
Running a restricted model on infrastructure outside China counts as an export and needs a license. Newly developed models that meet the thresholds would go through reporting before an open-source release, a process the note describes as automatic licensing. The note also points out that "open source" is not defined in the draft, so it is unclear how restricted-use licenses would be treated.
There is also an interim registration requirement already in effect, covering things like transferring technology ownership abroad or licensing technology for overseas use. That part is not a draft.
What this changes for a solo builder
Nothing you run today stops working. The change is on the supply side: a Chinese lab that trains a frontier-scale model may now need to report before it publishes weights, and the draft gives the state a mechanism to slow or stop that. One commentator, Dean W. Ball, called it "the optionality to shut down future open-weight models at will," as relayed by developersdigest.tech. That is a description of leverage, not an announcement of a shutdown.
I would also not assume every model is in scope. The 2 trillion parameter and 10^26 FLOP tests are high bars, but the top-10 ranking test is fuzzy and moves. Whether a particular Kimi or Qwen release would count is something I cannot tell from the draft.
Where this could be wrong
This is a draft. The catalogue could change before it is final, and the exemptions could widen or shrink. Past export-control drafts in other countries have been softened and have also been tightened. I have no information about how aggressively MOFCOM would enforce the report-and-export step, and "automatic licensing" suggests it may be mostly paperwork.
There is also a cost to overreacting. Moving a stable product off a model you like because of a draft policy is usually worse than doing nothing.
What I'd actually do
Spend 30 minutes on a dependency check this week. List every Chinese open-weight model you rely on, and for each one write down how it is hosted: your own GPU, a third-party inference provider, or the lab's own API. The API case is the least exposed, according to the draft. The self-hosted case outside China is the one the draft mentions by name.
Then keep a local copy of the weights you actually run, as long as the license allows it, because already-published weights are exempt and a copy is cheap insurance against a future takedown. Finally, put the model name in config, not in code, so swapping to a Western or other open model is a one-line change. If you have not built that switch yet, build it now. It pays off whether or not this draft ever becomes final.
I would not rewrite anything until the final text is out.
Author
Lukas
@lukcombinator