China’s Ministry of Commerce and Ministry of Science and Technology have published a draft revision of the country’s Catalogue of Technologies Prohibited or Restricted from Export that would add 23 artificial intelligence technologies to the controlled list, Bloomberg reported and the ministries’ own explanation confirms. The draft, issued as Announcement 2026 No. 48 on October 16, pulls model weights and source code into an export regime that until now reached mostly the hardware beneath them.
The AI items sit inside a wider revision covering 40 technologies across AI, biomedicine and solar power: 14 of the 40 would be prohibited from export and 26 restricted, with the AI set split between 6 prohibited and 17 restricted items. Public consultation runs until November 15, 2026, and a temporary pre-export registration requirement for the listed technologies is already in force.
The Thresholds: 10^26 FLOPs, Two Trillion Parameters, Or A Top-10 Ranking
The model rules bite at frontier scale. The underlying technologies of a large model — its weights, checkpoints, source code, architecture and training or inference methods — become controlled once a model reaches 10^26 FLOPs of training compute, carries at least 2 trillion parameters, or ranks in the global top ten on public evaluator rankings. On the Associated Press’s reading, that threshold already captures named Chinese systems such as Kimi K3 and Qwen 3.8-Max, though the draft names no models itself.
A second trigger covers data. Core technology for models trained on China’s “important data”, as defined under the Data Security Law, faces an outright export ban, with narrow exceptions where the data is public, cleaned below 0.1% of the training set, never appears in outputs and does not materially boost core-technology research. MOFCOM also cited a specific security risk: AI technologies able to autonomously generate or modify malicious code aimed at critical infrastructure, defence systems or government networks.
What Stays Free: APIs, SaaS And Open Source
The draft is careful about the boundary between transferring technology and selling access. Using a restricted model through an API call, a website, an app, a SaaS product or a lightweight local version is not treated as an export. Deploying the same restricted technology on servers overseas is, and requires registration or a licence — though Chinese firms operating their own data centres abroad may qualify for simplified procedures.
Open source gets its own lane. Technologies already in the public domain through open-source releases or academic publications sit outside the catalogue’s scope, and new open-source releases qualify for a fast-track “automatic licensing” route. That design protects the strategy that produced DeepSeek, Qwen and Kimi — the open-weight ecosystem our explainers on DeepSeek and Qwen trace — while giving Beijing visibility over frontier-scale transfers that stay closed.
Why It Matters
The central entity is China’s technology-export catalogue, and its new attribute is reach: for the first time it covers the trained artefacts of AI — weights, checkpoints and methods — and not only chips and equipment. Its neighbours are the Chinese labs whose global expansion now runs through a registration desk in Beijing, the foreign cloud providers that host Chinese models overseas, and Washington, whose own export rules have used compute thresholds for years; the draft mirrors that logic back. Read alongside the global map of AI laws in 2026 and the EU AI Act, it is another sign that model weights are becoming a controlled strategic asset in every major jurisdiction, even as the services built on them keep flowing freely.
Background
China’s export catalogue dates from a system built for an earlier technology era and has been revised in rounds as Beijing’s strategic priorities shifted; the commerce ministry has been consulting domestic firms including Alibaba, ByteDance and Zhipu on the shape of AI controls since the summer. The October draft is the first to put trained AI systems themselves squarely inside the perimeter, extending a global pattern in which the US restricts advanced chips and compute, and the EU regulates models by capability tier.
Source: Digital Watch Observatory, The Business Standard (AP) and AI Weekly, reporting the MOFCOM and MOST draft published October 16, 2026.
0 Comments