China AI Export Controls Draft: Model Weights Curbed, APIs Spared

China has proposed export controls on 23 artificial intelligence technologies, and the draft would put the weights of its most powerful AI models behind a government licence. The commerce and science ministries published the plan, Announcement No. 48 of 2026, at 4 p.m. Beijing time on Friday, October 16 (4 a.m. in New York). It would restrict exports of the weights and source code of any model that crosses one of three lines, including 2 trillion parameters. Selling access through an API, app or website would stay outside the rules, according to China's Ministry of Commerce (MOFCOM). Comments are open until November 15, but one part already applies: exporters must now register before sending covered technology abroad.

What happened

The draft revises China's Catalogue of Technologies Prohibited or Restricted from Export, a list Beijing has updated many times since 2001. It adds 40 technologies: 23 in AI, five in biomedicine and 12 in solar manufacturing, according to the MOFCOM announcement. In a question-and-answer statement, a ministry spokesperson said 14 of the 40 would be banned from export outright and 26 would need a licence. The AI items span large models, agents, computing clusters and information processing, and the draft annex also lists AI chip design.

The spokesperson called the revision routine and said it does not target any specific country, region or entity, as Xinhua and China Daily also reported. For AI, the ministry says the goal is to keep the technology "safe and controllable," and it cites the risk of AI-generated malicious code being used against critical infrastructure. Reuters called it Beijing's latest move to limit transfers of tools it sees as vital to national security and its rivalry with Washington. Reuters reported in July that officials had discussed limiting overseas access to top models with Alibaba, ByteDance and Z.AI, among others.

Key details

Which AI models would be covered

The draft's restricted list covers a large model's weights, key checkpoints, source code and architecture, plus training and inference-speedup methods. A model is caught if it meets any one of these tests:

Compute: trained with at least 10^26 floating-point operations. The South China Morning Post calls this a widely used threshold for frontier AI.
Size: at least 2 trillion parameters.
Ranking: "globally leading" capability, which the annex defines as a top-10 place in public global rankings at the time of export.

Models trained on China's "important data" would also be restricted, and those trained on its "core data" would be banned outright. The draft doesn't name any model. The Next Web, citing Bloomberg and the Associated Press, says Moonshot AI's Kimi K3 and Alibaba's Qwen3.8-Max are among the models above the parameter line.

What counts as an export, and what doesn't

Not an export: offering a restricted model through API calls, an app or web page, a SaaS product or a lightweight on-device version, according to the MOFCOM Q&A.
An export: deploying a restricted model locally on servers abroad. If it runs on the Chinese company's own overseas cloud, automatic licensing can apply, the ministry says.
Exempt: technology already in the public domain through open-source releases or published papers.
Fast-tracked: new open-source releases, Chinese firms' own data centres abroad, chip manufacturing runs abroad, joint research and international standards work. These would get automatic licensing, meaning "immediate clearance upon filing," as SCMP translates the ministry's phrase.

Agents, chips and clusters

Restricted agent technologies include systems that can split an instruction into 20 or more executable steps, coordinate at least 60 agents at once, or work with at least 200 third-party tools, each paired with accuracy or success-rate targets. AI chip design know-how is covered if, among other triggers, it needs a 14-nanometre or smaller manufacturing process or uses 2.5D or 3D advanced packaging. Cluster rules apply to setups with more than 10,000 AI accelerator cards. Products made with the listed technologies aren't covered, MOFCOM says.

Cyber tools face outright bans

The banned list includes AI-driven malicious code generation that could be used against critical infrastructure, with thresholds such as a 60% success rate against multi-engine malware detectors. AI-assisted vulnerability hunting would be restricted rather than banned, for example if a system can find five or more undisclosed high-severity flaws in one major open-source project or operating system in a single run.

The rule that already applies

Since October 16, exporters have had to register on a MOFCOM online platform before transferring ownership of a listed technology abroad, licensing it for use abroad, or letting a foreign entity become its owner or actual controller. Unfinished deals that began earlier are included.

Why it matters

API access keeps flowing. As written, the draft leaves US, UK and EU developers who call Chinese models through an API or app untouched. Reuters notes this could let DeepSeek and Alibaba keep growing abroad. For DeepSeek's latest release, see DeepSeek V4.1 Flash 1017 Is Out: 77% DeepSWE at $0.20 a Task.
Self-hosting future frontier models could get harder. Running a restricted model on a customer's own servers abroad would count as an export. SCMP reports that Moonshot AI and Z.ai are signing revenue-sharing deals to host their open-weight models on US clouds such as Amazon Web Services. Alex Colville of the Australian Strategic Policy Institute told the paper the rules might delay or prevent such deals for the most powerful models.
Open weights are the big unknown. Published weights are exempt and new open releases would be fast-tracked, but "There's still a lot of things we don't know," Colville told SCMP.
Ownership deals are in scope. Shanghai lawyer You Yunting told SCMP the registration rule looks like a response to Meta's aborted $2 billion purchase of agent developer Manus, which Beijing halted.
Export controls now run both ways. Reuters notes that US limits on China's access to advanced AI chips have helped keep American labs ahead. Chinese labs were already in the spotlight, as in OpenAI Disrupts Moonshot-Linked Reasoning Theft.

What's next

Comments are open until November 15, 2026, by email, fax or letter to MOFCOM. A final catalogue will follow, with no start date yet, and registration stays in force until then.

Watch the final wording on open-source releases and foreign cloud deployments. AI is also expected to be a major topic when Chinese President Xi Jinping and US President Donald Trump meet at the APEC summit in Shenzhen next month and the G20 summit in Miami in December, according to SCMP. The two leaders agreed last month to set up a bilateral dialogue on AI risks.

FAQ

Can I still use DeepSeek, Qwen or Kimi outside China?

Yes, based on the draft. MOFCOM says offering a restricted model through an API, app, website or SaaS product doesn't count as a technology export. The rules target transfers of weights and source code, and local deployments abroad.

Which AI models fall under China's export controls?

Under the draft, exporting the core technology of any large model trained with at least 10^26 operations, with 2 trillion or more parameters, or ranked in the global top 10 at the time of export would need a licence. The draft names no models, and weights already released as open source are exempt.

When do China's AI export controls take effect?

The licensing rules aren't final. Comments close on November 15, 2026, and no start date has been set for the revised catalogue. A temporary pre-export registration rule has applied since October 16.

Search This Blog

The Frontier Brief