Image: Government of China

China wants to put its most powerful AI models under export control. The Ministry of Commerce and the Ministry of Science and Technology have published draft changes to the country's list of technologies that are banned or restricted from export. The draft adds 23 AI items covering large models, agents, computing clusters and information processing.

完蛋了…..以后中国开源模型只能开源小参数的模型了….

商务部科技部就修订《中国禁止出口限制出口技术目录》征求意见，为期一个月，涉及人工智能（大模型、智能体、算力集群、信息处理等23项）生物医药、光伏3个领域。 pic.twitter.com/Brc8usMjVx

— Max For AI (@MaxForAI) October 16, 2026

The news spread fast after AI commentator MaxForAI posted the key thresholds on X with a blunt take: Chinese labs may soon only be able to open-source small models. The post named Kimi's K4, MiniMax's M3 Pro, Zhipu's GLM 5 and DeepSeek's next model as releases that might not be open-sourced. The official documents tell a more nuanced story.

Image: Ministry of Commerce of China

Under the draft, a large model's core technology becomes "restricted" if it meets any one of three tests. It was trained with 10^26 or more floating-point operations, it has 2 trillion or more parameters, or its capability is "globally leading." Core technology here means weights, key checkpoints, source code, architecture, and training and inference tricks.

The first figure may sound familiar. The US used the same 10^26 line in its 2023 AI executive order to decide which models needed reporting to the government.

Image: Ministry of Commerce of China

The third test is the one Chinese labs should watch most closely. The draft defines "globally leading" as ranking in the global top 10 on any public evaluation by major testing bodies or large industry platforms. That covers overall performance, specialist skills, global usage, safety and training efficiency, judged at the moment of export. A model that tops just one popular leaderboard could qualify, even if it's nowhere near 2 trillion parameters.

Models trained on China's "core data," and agents built mainly for critical infrastructure, defence or core government use, go further: they'd be banned from export outright. The draft also sets numeric bars for agents, such as handling 200 or more third-party tools with at least 90% accuracy.

So is open source over? Not according to the government's own wording. A footnote in the draft says open-sourcing restricted AI technology qualifies for "automatic licensing." In its official Q&A, a ministry spokesperson said that should mean exports are basically approved as soon as they're reported. Technology that is already public through open-source releases or research papers falls outside the rules entirely.

That still adds a government checkpoint before a big open release. The documents don't say how long automatic licensing takes in practice, or whether it could be withheld. MaxForAI's warning is one reading of the draft, not a confirmed ban.

For regular users, the spokesperson said offering restricted models through APIs, apps, websites, SaaS products or lightweight on-device versions is not a technology export. Chatting with DeepSeek or Kimi from abroad shouldn't change. Deploying a restricted model on servers overseas would count, though running it on a Chinese company's own overseas cloud can use the fast-track licence.

The rules aren't final yet. Comments are open until mid-November. In the meantime, companies must register any export of the listed technologies with the Commerce Ministry before it happens, and that requirement is already in force. The ministry insists the move doesn't target any specific country and that "the vast majority" of AI technology will remain freely exportable.