China Drafts Export Controls on 23 AI Technologies, From Model Weights to Agents

China's Ministry of Commerce and Ministry of Science and Technology on October 16, 2026 proposed adding 23 AI technologies, covering large models, AI agents, computing clusters and information processing, to the country's catalog of technologies banned or restricted from export. The core technology of large models, including their weights, would need an export license once a model passes set thresholds for compute, size or global ranking, while AI technology already released as open source stays outside the rules, the draft says.

The proposal follows months of reporting on Beijing's AI policy plans. In July 2026, the commerce ministry was consulting Alibaba, ByteDance and Zhipu on options that included stopping foreign users from downloading the weights of China's most advanced models (FT via TechRepublic). The published draft is narrower on open models. Public comments run until November 15, 2026.

What does China's draft add to the export catalog?

China's draft adds 40 technologies to its export control catalog, 23 of them in AI, according to Announcement No. 48 of 2026 published by the two ministries on October 16. The rest are 5 biomedicine and 12 solar technologies.

By a count of the draft's annex, 6 of the 23 AI items would be banned from export outright and 17 would be restricted, which means they need a license. The bans cover AI that writes and mutates malicious code on its own to attack critical infrastructure, defense or core government systems; large models trained on China's "core data"; high performance agents built mainly for those sensitive settings; and some deepfake detection systems.

The restricted list reaches beyond models: AI assisted vulnerability discovery, voiceprint recognition, AI chip design, and networking and scheduling technology for clusters of more than 10,000 AI accelerators. Agents are restricted when they pass performance thresholds, such as splitting an instruction into 20 or more subtasks at set accuracy and success rates.

Which AI models would fall under the new rules?

Under the draft, the core technology of a large model becomes restricted when the model was trained with at least 10²⁶ floating point operations, has at least 2 trillion parameters, or ranks among the global top 10. Floating point operations are the basic calculations a chip performs, so the first figure is a measure of how much computing power went into training.

"Core technology" means weights, key checkpoints, underlying source code and architecture designs, plus training and inference optimization methods. A model counts as globally leading if public rankings by major Chinese or foreign evaluators or industry platforms place it in the top 10 at the time of export, on measures such as overall performance, global usage or safety.

A separate clause restricts models trained on China's "important data" under its Data Security Law, with exemptions for lawfully public data and for cleaned trace amounts of 0.1% or less of the training data.

What happens to open weight models?

AI technology that has already entered the public domain through open source releases or academic papers is excluded from the controls, the commerce ministry said on October 16, 2026. Open-sourcing restricted AI technology in the future would fall under "automatic licensing", which the ministry says amounts to clearance on filing.

That matters for China's largest open models. Alibaba's Qwen3.8-2.4T-A95B, whose weights have been on Hugging Face since August 2026, has 2.4 trillion parameters according to its model card, above the draft's 2 trillion threshold. Because it is already public, it stays outside the rules; under the draft's text, a successor of that size would need an automatic license filing before release. In Meta's own table for Muse Spark Open, which Meta calls the most capable US open weight model, Qwen3.8, GLM 5.3 and Kimi K3 all score higher on the Artificial Analysis Intelligence Index.

The ministry also drew a line on deployment. Serving a restricted model abroad through APIs, apps, web pages, SaaS or lightweight on-device versions is not a technology export, while installing restricted technology on local servers abroad is, though a company's own overseas cloud can use the faster process.

In the same post, Ball called the framework a mirror image of model weight controls the US government has already used. The US Framework for Artificial Intelligence Diffusion required a license to export the weights of closed models trained with 10²⁶ operations or more, the same compute threshold, and left open weight models alone; the Commerce Department announced its rescission on May 13, 2025.

What do Chinese exporters have to do now?

Chinese exporters must register exports of the listed technologies on a commerce ministry platform from October 16, 2026 until the revised catalog takes effect, including exports already under way, according to the announcement.

Registration covers transfers of ownership abroad, licenses for foreign use and a change to a foreign owner or controller, and exporters report the recipient, destination, use and expected value.

In its spokesperson's answers, the ministry said the revision does not target any country, region or entity, and that most AI technology will remain freely exportable.

What we don't know yet

• When the revised catalog will take effect, and how much the final text will change after the comment period.
• Which rankings the ministry will use to decide whether a model is in the global top 10.
• How long an automatic license for an open source release will take in practice, and whether a filing can be refused.
• Whether the 2 trillion parameter threshold refers to total or active parameters in mixture of experts models.