China is preparing for the first time to turn AI models themselves into a subject of export regulation: on October 16, 2026, the Ministry of Commerce and the Ministry of Science and Technology of the PRC opened a draft for public discussion, expanding the national export control list by 40 technologies, 23 of which relate to artificial intelligence, including large language models, AI agents, chip design solutions, and computing clusters. The weights and source code of the largest models may fall under licenses, with thresholds set at two trillion parameters or entry into the global top 10 in performance and safety. The draft is open for comments until November 15, 2026, however, the mandatory registration of export transactions involving controlled technologies has already been in effect since October 16, 2026, under a temporary rule, while open source code, published research, and the maintenance of foreign users through applications, websites, and APIs remain outside the restrictions.

image

What happened

The Ministry of Commerce and the Ministry of Science and Technology of the PRC published the draft on October 16, 2026, and are accepting comments until November 15, 2026, after which the project may become the final list. The document expands the national export control list by 40 items, 23 of which are related to artificial intelligence: it includes large language models, AI agents, chip design solutions, and computing clusters. The controllability threshold for large models is set numerically: no less than 2 trillion parameters or a place in the world's top 10 in performance or safety, and the restrictions cover model weights, their source code, and other key components. Simultaneously, a temporary rule on the mandatory registration of export transactions came into force, so accounting covers transfers already now, before the final decision is made. Open source code, published research, and active maintenance of foreign users through applications, websites, and APIs are directly excluded from control.

Context

The measures are embedded in the chip confrontation: Washington controls the supply of advanced chips to China, while Beijing formalizes a mirror response, choosing an export lever rather than an import one, i.e., regulating the export of its own AI software and model weights abroad. Methodologically, this is the first case where a major jurisdiction describes the subject of export regulation through the technical metrics of the model itself, from which it follows that the regulatory status of a model may now depend on the methodology of its evaluation. Structurally, the project draws a watershed between service and artifact: restrictions are concentrated on the direct transfer of closed weights and source code, while typical channels of the open ecosystem and public access to models remain free. Thus, the weights and code of top models are for the first time turning into a potentially licensable commodity, whose cross-border circulation is subject to a procedure, not the discretion of companies.

Why this matters for the industry

For the industry, the class of operations with Chinese models is changing: the direct export of weights, source code, and key components of closed models, to which Kimi K3 from Moonshot AI and Qwen3.8-Max from Alibaba are attributed during the discussion, becomes a licensable procedure. The immediate effect is administrative: new agreements on such transfers from October 16, 2026, go through mandatory registration under the temporary rule, which adds legal friction and pauses to transactions. Products on public APIs of Chinese models continue to work without changes, as the channel has been explicitly left permitted, and the risk is concentrated among teams that host others' weights abroad themselves. The strategic signal is more important than the letter of the document: a product's dependence on Chinese weights ceases to be a purely technical choice and becomes a legal risk, so if the project is adopted close to the current text, a licensing cycle for moving the weights of the most advanced models between jurisdictions, "export-clean" versions of models deliberately kept below thresholds, increased demand for tools to track the regulatory status of models, and a strengthening of the trend toward stacks without Chinese dependence are expected.

Why this matters for users

For most readers using Chinese models through public APIs, applications, and websites, nothing changes: the maintenance of foreign users is directly left permitted, and open source code and published research are excluded from control. Changes affect those who work with model artifacts: transferring the weights and code of closed top models abroad becomes an operation requiring a license, and corresponding transactions are registered now under the temporary rule. Practical conclusion: the answer to the question of which Chinese models and in what form will remain available outside the PRC now depends not on the quality of a specific model, but on the access channel, because the service is open, and the artifact is under control. Those planning to host advanced Chinese models abroad on their own should check the applicability of registration in advance and prepare backup channels through public APIs.

What is currently unknown / limitations

The formulation "no less than 2 trillion parameters" does not clarify what to count as parameters for sparse MoE architectures: dense, total, or only active, so the attribution of a specific model, including Kimi K3 and Qwen3.8-Max, to the controlled category remains debatable. The criterion of "top 10 in the world in performance or safety" is not supported in the sources by a specific benchmark or leaderboard, making the definition in its current form non-reproducible. The parameter threshold itself is an administratively measurable, not scientifically meaningful, boundary of capabilities, as distillation and quantization allow obtaining strong models significantly below it. The document remains a draft and may change before adoption, and estimates of long-term consequences, including scenarios with "export-clean" versions of models, represent forecasts, not established facts.

Sources

Author

Look at AI, editorial team