China Proposes Export Curbs on 23 AI Technologies, From Models to Chips

China wants a say over which of its AI models, agents and chip designs leave the country. On Friday, October 16, the Ministry of Commerce and the Ministry of Science and Technology jointly published a draft update to the catalogue of technologies subject to export bans or restrictions. The draft adds 23 AI technologies, and a temporary export registration rule took effect the same day.

The 23 AI items sit inside a larger draft covering 40 technologies, which also includes solar and biomedicine, according to AP, as republished by The Star and China Daily. Bloomberg Tax describes the AI portion in more detail, including restrictions on model weights and source code for products above certain performance thresholds. The draft is not yet in force, and public comments remain open until November 15.

The timing is notable. The proposal arrives while China and the United States advance dialogues on AI safety and governance, along with a channel for reporting AI incidents that could threaten national security. It also lands as Chinese open-weight models press US labs on price.

What the draft covers

The reporting agrees on the broad categories, though it describes individual items at different levels of detail. Across the sources, the list includes:

• Advanced large models, including core technology for the largest systems
• AI agents
• Chip designs, some for AI processing and memory chips
• Computing clusters
• Model weights, checkpoints, source code and architecture for products above the thresholds

The breadth is what stands out. A list that covers models, agents, silicon and the infrastructure used to train them reaches much of the AI stack, not one product class. It also means a single model release could fall under several items at once. That reading is my inference, and the final text will decide how the items overlap.

The 2-trillion-parameter line

The sharpest test in the draft is size. Core technology for models with at least 2 trillion parameters could face export restrictions, according to Taipei Times, citing Bloomberg. The same provision covers models that rank among the world's top 10 on performance, usage or safety measures. The restrictions would reach weights, checkpoints, source code and architecture.

The parameter count is easy to state and easy to check. The ranking test is not. The reporting does not say which leaderboards or metrics would count, or how a model's position would be measured. Yahoo Finance, citing Bloomberg and Taipei Times both leave that question open. Because the ranking test is independent of size, a smaller model that tops a usage chart could fall within scope even if it sits below 2 trillion parameters. That follows from the wording, but no source confirms it.

Two models appear as examples of systems that would clear the threshold: Moonshot AI's Kimi K3 and Alibaba's Qwen3.8-Max. Those are the names US developers will check first. They also suggest the rule is aimed at frontier-scale systems first.

Open weights: where the reports disagree

This is where the sources conflict, and it matters most to developers. Bloomberg reporting, carried by Yahoo Finance, says restrictions would reach model weights and source code for products above certain thresholds. The Next Web reports that open-source technology is exempt and that overseas users keep access to models through apps. Both descriptions could be partly right, but they point to very different planning decisions.

Open release is central to how Chinese labs compete. Chinese companies tend to share model weights so developers can build on them, which is one reason their models have spread quickly beyond China. A rule that reaches weights would cut across that habit. A rule that exempts open-source releases would leave the practice intact but could still restrict the most capable closed systems.

DeepSeek's V4.1-Flash-1017 release, which shipped with open weights and unchanged prices, shows how much of the current Chinese model market depends on that practice. Until the final text is public, no one should assume those weights will stay downloadable under the same license.

Beijing's stated reason and the US mirror image

The Commerce Ministry has framed the measures as a security matter. A spokesperson, unnamed in a statement reported by Bloomberg, said the measures aim at ensuring "the safety and controllability" of AI technology. AP reported a similar statement, saying the government is taking such measures to ensure that AI remains under human control. The stated targets are misuse against critical infrastructure, defense and core government systems.

Those are standard security concerns, but they sit inside a larger contest. US-led restrictions have kept China from the most advanced AI chips and chipmaking machines. Read alongside that record, the draft is less a standalone move than a counterpart to American export policy.

Competition adds another layer. Chinese startups narrowed the gap with US leaders this summer, and models such as Kimi K3 and DeepSeek's latest editions offer better pricing. A government that sees its labs gaining ground has a reason to decide early how those systems can be used abroad. That motive is my reading of the timing, not something a ministry has said.

What US developers and companies should do now

For most readers, nothing changes today. The draft is not in force, and overseas users can still reach models through apps. The people who need to act are those who build products on Chinese models, and a few concrete steps will reduce exposure whichever version of the rule survives.

1. Inventory every Chinese model in production. Record the model name, exact version, license terms and where it is hosted.
2. Save dated copies of model cards and license pages. If terms change, you will have a record of what you agreed to and when.
3. Map dependencies. Identify which products, customers or pipelines would break if a model's weights stopped being available or were re-licensed.
4. Test a fallback model from a non-Chinese provider on your most important tasks, so a switch becomes a configuration change rather than a rebuild.
5. Watch the November 15 deadline. If the rule would affect your business, consider filing a comment before the window closes, since the draft is still open to input.

Chipmakers and infrastructure buyers have less immediate work. The chip-design and computing-cluster items are the ones to watch, but the reporting does not show a direct effect on US companies yet.

What to watch before November 15

Four questions will decide how far the draft reaches. First, whether the final text keeps restrictions on weights for above-threshold products or adopts an open-source exemption. Second, how the top-10 test is measured. Third, how enforcement would work for overseas users and for providers outside China. Fourth, whether US officials or Chinese AI labs respond publicly. None of the reporting I reviewed includes a US government reaction, a market reaction or a statement from a Chinese AI lab.

The next signal will come from the final text, not the draft. If the weights provision survives the comment period, developers outside China will face a harder choice about which Chinese releases to depend on. If the open-source exemption holds, the bigger story becomes how Beijing uses export registration to decide which systems travel abroad. Either way, November 15 is the first point where the rules could move, and teams that treat it as a planning date rather than a formality will be better placed when they do.