# China wants a say over which of its AI models leave the country > Source: > Published: 2026-10-18 15:19:43+00:00 On October 16, China's Ministry of Commerce and Ministry of Science and Technology issued a draft revision of the country's catalogue of technologies that are banned or restricted from export. The revision would add 40 technologies, 23 of them in artificial intelligence. Nothing already public is being withdrawn. What the draft builds is a decision: the power to rule, one model at a time, on whether an unreleased frontier system may leave China. ## The draft adds 40 items, 23 of them in AI The [announcement](https://www.mofcom.gov.cn/zcfb/blgg/gg/2026/art/2026/art_05e2643a8ca04eb9adf9dd0542991fe4.html) is a consultation draft, and it holds two things apart. The proposed additions to the catalogue are open for public comment until November 15. A temporary registration requirement took effect on the day of publication. The Commerce Ministry's [own breakdown](https://www.mofcom.gov.cn/syxwfb/art/2026/art_a4acff07e134447fbe7bbeb486df3ea6.html) places the 40 items across three sectors: 23 AI technologies, 5 in biomedicine and 12 in photovoltaics. Of the 40, 14 would be prohibited from export outright and 26 restricted. The AI group covers large models, AI agents, computing clusters and information processing. Chip design sits inside that group rather than in a category of its own: design technology for AI processors made on 14-nanometre-or-finer processes, or using advanced packaging, would be restricted, along with certain advanced memory-chip designs. One entry would be banned rather than restricted: AI-driven generation and mutation of malicious code, which the ministry says could be used to attack critical information infrastructure, defence systems or government networks. AI-assisted discovery of software vulnerabilities would be restricted. ## What counts as leaving the country The proposal follows the technology, not the service. Exporting a model's weights, its key checkpoints, its underlying source code, its architecture design, and the methods used to train it or speed up its answers all fall inside the rules. Letting customers abroad reach a model through an API, an app, a website, a SaaS tool or a lightweight version on their own device does not, per the [ministry's policy explanation](https://www.mofcom.gov.cn/syxwfb/art/2026/art_a4acff07e134447fbe7bbeb486df3ea6.html). Installing a restricted model on servers overseas does fall inside the rules, although the ministry says a company running its own overseas cloud can qualify for an automatic licence. Technologies already in the public domain are excluded. An open-source release or a published paper takes a model's design out of the managed scope, the ministry says, so nothing previously released is caught by the draft. ## The thresholds that would sort big models from small If a model counts as restricted, the draft annex defines its protected core technology and sets three marks. Any one of them brings the technology under restriction: at least 10^26 floating-point operations of training compute (a 1 followed by 26 zeros), at least 2 trillion parameters (roughly, the adjustable settings inside a model), or a place in the global top 10 on public evaluations of performance, business use, call volume, safety or training efficiency. These figures sit in a draft annex and are not in force; they appear in a [published rendering of the draft catalogue](https://the-decoder.com/china-plans-to-tighten-export-controls-on-open-weight-models-and-23-key-ai-technologies/), not in the announcement text itself. The 2 trillion mark is not a distant ceiling. Moonshot's Kimi K3 and Alibaba's Qwen 3.8-Max already have more than 2 trillion parameters, [AP reported](https://www.news4jax.com/business/2026/10/16/china-moves-to-curb-tech-exports-including-ai-citing-national-security-risks/), though the draft names no model. Alibaba's [own roadmap](https://www.alibabagroup.com/en-US/document-2041268239081668608) projects its coming Qwen 4.5 and Qwen 5 series at 5 to 10 trillion. The annex goes further than size. A model trained on China's core data and then offered as a service would be proposed for outright prohibition, according to the same rendering, and one trained on important data would be restricted, with a carve-out where only a trace amount, 0.1 percent, of that data is used and cannot substantially improve the model. Agent technology is pinned to specific abilities: planning that breaks a plain instruction into a chain of 20 steps, say, or tool calling that links 200 outside tools. A large computing cluster is defined as more than 10,000 AI accelerator cards. ## Open releases get a fast lane The draft does not only exempt what is already public; it accelerates what comes next. Releasing restricted AI technology as open source would qualify for automatic licensing, the ministry says, meaning the export clears almost as soon as the paperwork is filed. The same fast lane would cover computing centres Chinese firms build and run abroad, overseas chip manufacturing, joint research using overseas resources and work on international standards. ## What is in force now Only one part of the package is effective. Since October 16, a company must register before it transfers ownership of listed technology abroad, licenses it for use overseas, or changes the owner or actual controller of a Chinese holder to a foreign entity, per the [announcement's temporary registration rule](https://www.mofcom.gov.cn/zcfb/blgg/gg/2026/art/2026/art_05e2643a8ca04eb9adf9dd0542991fe4.html). Deals already under way are included. The requirement runs until the revised catalogue formally takes effect. Public comment on the rest runs from October 16 to November 15, by email, fax or post, as the same announcement sets out. ## How to read the intent The ministry describes the revision as routine technology-trade management that [does not target any particular country, region or entity](https://www.mofcom.gov.cn/syxwfb/art/2026/art_a4acff07e134447fbe7bbeb486df3ea6.html), and its explanation of the change says the aim is to [prevent technologies from becoming uncontrollable or spreading inappropriately](https://www.mofcom.gov.cn/zwgk/zcjd/art/2026/art_b11d5c1e9374421e93e85dc87eba6e38.html), while keeping normal international cooperation open. A sharper reading of the timing comes from outside the government. Dean W. Ball, head of strategic futures at OpenAI, published [his own analysis, as quoted by The Decoder](https://the-decoder.com/china-plans-to-tighten-export-controls-on-open-weight-models-and-23-key-ai-technologies/), which is not an OpenAI statement. In it he wrote that the move "does not mean 'China is shutting down open-weight models.'" He wrote that it establishes "the optionality to shut down future open-weight models at will," a "mirror image" of the model-weight export controls the United States has already imposed. That reading fits the documents. The draft leaves released models alone, invites open releases onto a fast track, and still leaves the government holding the choice over what comes next. The registration duty is the one fact on the ground, and it reaches deals already under way. Whether the thresholds, the data rules and the fast lane survive the consultation is a question with a date: the window closes on November 15.