China drafts export licensing for the weights of 2T-parameter AI models, with open-source releases fast-tracked

China's commerce and science ministries want the weights, checkpoints and source code of the largest AI models to need a license before they leave the country. The proposal is a draft, MOFCOM and MOST Announcement No. 48 of 2026, published on October 16 and open for public comment until November 15. We read the announcement and its annex directly.

WHAT THE DRAFT COVERS

The annex adds 40 items to China's list of technologies whose export is banned or restricted, 23 of them in AI: large models, agents, computing clusters, chip design and information processing. Under the restricted item for large-model technology, a model's core technology (its weights, key checkpoints, underlying source code and architecture design, plus its training-optimization and inference-acceleration methods) would need an export license if the model meets any one of three tests:

1. Training compute of at least 10^26 floating-point operations. 2. At least 2 trillion parameters. 3. "World-leading" capability, which the annex defines as a top-10 place, at the time of export, in a public global ranking by a major Chinese or foreign evaluator or a large industry platform, on overall performance, specific skills, global usage, safety or training efficiency.

The same item restricts large models trained on China's "important data" that are offered as a service. A separate item on the banned list covers large models trained on its "core data". Both terms come from China's Data Security Law.

WHAT IS LEFT OUT

Technology that is already public through open source or academic papers is outside the draft's scope. On that wording, weights that are already out, such as Kimi K3's 2.8-trillion-parameter release in July, are not caught.

Open-sourcing covered AI technology gets "automatic licensing and other facilitation measures", the same treatment the draft gives Chinese companies running their own compute centers abroad. So a new open-weight release of a model above the line would go through a license that is granted automatically rather than being blocked. The draft does not spell out how that process works.

MOFCOM's explainer, published by Xinhua, says offering a restricted model through an API, an app or web chat, a SaaS product or a lightweight on-device deployment is not a technology export. Deploying the model on servers abroad is, though it too gets automatic licensing when the servers are the company's own cloud.

REGISTRATION STARTS NOW

From the day of the announcement until the revised list takes effect, anyone exporting a listed technology has to register it on MOFCOM's platform first, and that includes exports already under way.

WHAT CHINA SAYS

MOFCOM calls the revision routine technology-trade management that does not target any country, region or company. It says the aim is to hold a security baseline, prevent misuse and keep AI under human control, while encouraging open source and cooperation.

WHAT WE DO NOT KNOW

Whether the three thresholds survive the comment period, how "top 10" will be counted when the major boards rank models differently, and how long an automatic license takes in practice. Reuters reported the draft on October 16; it picked up attention on X on October 18 after @kimmonismus posted about it.