China Moves to Control Exports of 23 AI Technologies: Thresholds, Weights and What Changes for Anyone Downloading Models
On October 16 China proposed adding 23 AI technologies to its export catalogue —model weights, agents and chip design— and switched on a registration regime already in force, with public comment open until November 15. What the draft says and what changes if you download models.
What China Published on October 16
The Ministry of Commerce (MOFCOM) and the Ministry of Science and Technology (MOST) published Joint Announcement No. 48 of 2026 on Friday, October 16, at 16:00 Beijing time. The text proposes revising the Catalogue of Technologies Prohibited or Restricted from Export, the list defining which technical knowledge may leave the country and under what conditions. At the same time, and this is what headlines usually miss, it introduced a pre-export registration requirement already in force, including for deals under way and not yet completed.
| Sector | Proposed | Prohibited | Restricted |
|---|---|---|---|
| Artificial intelligence | 23 | 6 | 17 |
| Biomedicine | 5 | 3 | 2 |
| Photovoltaics | 12 | 5 | 7 |
| Total | 40 | 14 | 26 |
AI carries the weight: 17 of the 26 restricted entries are artificial intelligence. The catalogue dates from 2001 and was revised in 2008, 2020, 2023 and 2025; this would be the first revision with a block devoted to large models. The comment period runs from October 16 to November 15, 2026 and accepts input on scope, technical thresholds and the facilitation arrangements.
Prohibited and Restricted Are Not the Same Thing
Prohibited means the technology cannot be exported at all. Restricted means it can be exported with a licence from the Commerce Ministry. A restricted chip design can leave with approval; the weights of a prohibited model cannot. Registration is triggered by four kinds of transaction: transferring ownership to an overseas entity, including intellectual property; licensing Chinese technology for use abroad; a change in ownership or actual control that puts it in foreign hands; and any case the ministry names by notice.
Read also
The practical distinction: the catalogue is still a draft, but registration is already required. Anyone saying China "has already banned AI model exports" is describing a decision that has not been made.
What Goes on the AI List
Weights, Checkpoints, Code and Training Methods
The annex defines "core large-model technology" precisely: weights, key checkpoints, source code, architecture and training or inference methods. It does not control only the final file, but also how that file was produced. A checkpoint is an intermediate copy of the model during training, and their inclusion matters because anyone holding one can keep training without the finished model.
The Thresholds: When a Model Makes the List
The restriction applies if any one of three conditions is met: 2 trillion parameters or more, 10^26 FLOPs or more of training compute, or a global top-10 ranking by capability. A FLOP is a floating-point operation, and counting them is the standard way to measure training compute. The threshold is an administrative line, not a quality measure: it keeps frontier models inside and a small open-source model outside, even when that small model does its job very well.
Agents, Clusters, Chips and Explicit Bans
Agents get thresholds of their own: task chains of 20 steps or more, 60 agents working together, or 200 tools used with 90% success. Compute clusters, AI information processing and voiceprint recognition are also included. On hardware, the draft restricts 2.5D and 3D AI chip design and 3D flash memory: not just the model, but the blueprint used to build the chip that trains it.
On the other side, the explicit ban covers the core technology of large models trained on the country's key sensitive data, and AI agents serving critical national information infrastructure, defence or core government affairs.
What the Draft Leaves Out (and Why It Matters)
API, SaaS and Apps: Controlling the File, Not the Service
According to the announcement itself, using a restricted model through an API, app, web, SaaS or on-device deployment falls outside export control. What triggers registration is taking the asset, not buying the service. The commercial logic is clear: a Chinese company can keep selling access to its model abroad while a foreign buyer is stopped from walking away with the file.
The Open-Source-Already-Published Question
The announcement is explicit: technologies already in the public domain through open-source releases, academic publications or other means are not subject to these controls. Outbound licensing without a transfer of ownership gets facilitation, including automatic licensing. One question the draft does not answer: what happens to what gets published next. A model already uploaded to a public repository is not "un-published"; what changes is the incentive to open the next one.
Why China Is Doing This Now
The proposal lands in the same week as TSMC's record results and a selloff in Asian chip stocks, and it is part of a two-year exchange of controls: Washington limits chip and tool sales to China, Beijing answers with access to critical materials and now with the exit of its own technology. These are separate regimes and worth keeping apart: one limits what comes in, the other what goes out. A catalogue under consultation is also a negotiable document, and having registration already live gives Beijing an immediate tool without approving the final text.
What It Means If You Code or Train Models
If You Consume Chinese Models Through an API
Nothing changes today. The announcement leaves API and SaaS access out of scope, and Chinese companies have every interest in keeping that channel open. The thing to watch here is the provider, not the file: if terms change, you hear it from a contract.
If You Download Weights or Fine-Tune
Here there is something to check. If your pipeline depends on downloading the weights of a large model and training on them, it is worth knowing whether that model falls inside the thresholds and whether redistributing it from China requires registration. A model already released as open source sits outside per the text; its next version may not. And if you work with hardware, 2.5D and 3D chip design is on the restricted list: restricted is not banned, but it stops being automatic.
What to Watch Until November 15
The thresholds in circulation come from an annex under consultation and can change; the definitive list of the 23 AI entries is the document to read when it is published. Today registration is informational: who exports, to whom, which technology and at what estimated value. If the final text turns it into a licence with prior approval, the administrative cost changes by an order of magnitude. And a Chinese export list covering frontier AI invites a symmetrical answer from the US and the EU, which is the part nobody can forecast.
Three facts are verified: the announcement is from October 16, it covers 40 technologies with 23 of them AI, and pre-export registration is already live while the catalogue stays under consultation. What remains open are the final thresholds and how far the open-source exemption reaches. For this week, the sensible move is to read the annex and know where the weights you use come from. If you want the background, we already covered China's ban on buying Nvidia chips and the rise of DeepSeek, the two pieces that explain why this catalogue exists.
