China: Export Controls on 23 AI Technologies and Model Thresholds

China formalizes an export framework for 23 AI technologies, combining prohibitions, technical thresholds, and exemptions. Open source technologies already public, uses via API, and expedited procedures are among the relaxations, while very large models and offensive tools are targeted.

Automatic Licenses and Exemptions for API Uses

The open-source publication of technologies or restricted weights grants an automatic license. This regime also applies to Chinese companies that operate data centers abroad, manufacture chips outside the country, conduct research with foreign partners, or participate in the development of international standards. Exports that meet the required conditions receive approval as soon as the necessary documents are submitted, thanks to this expedited mechanism. Additionally, access to restricted models through APIs, applications, websites, SaaS tools, or locally installed lightweight versions is not subject to export controls. However, deploying these technologies on servers located abroad requires obtaining a standard export license, unless the software is hosted on the company's internal cloud, in which case the simplified procedure remains possible. Until the final rules come into effect, targeted exports must be declared on the Ministry of Commerce's platform.

Prohibitions and Limits: Autonomous Malware Blocked

The updated catalog from the Chinese ministries of commerce and science covers 23 key AI technologies, including large models, agents, computing clusters, and data processing. AI tools capable of autonomously generating or modifying malware to target critical infrastructure, the military, or government agencies are completely prohibited from export. Some AI tools intended for detecting software vulnerabilities are not entirely banned but are subject to new restrictions. This update to the national catalog of prohibited and restricted technologies has been published by the relevant authorities.

Technical Thresholds Triggering Control of Large Models

The essential components of large models, namely weights, checkpoints, source code, architecture, as well as training or inference methods, are subject to control. These limitations come into play when a model reaches a training power of 10^26 FLOPs, has at least 2 trillion parameters, or ranks in the global top 10. These requirements are in addition to those related to models trained on sensitive Chinese data. Membership in the top 10 is assessed by export authorities based on public rankings from national or international evaluators or industrial platforms, taking into account overall performance, certain specific tasks, security, or global usage.

"Important" Chinese Data: Exemptions and 0.1% Quota

Models trained on data considered important under Chinese law and providing services based on this data are subject to export controls. Legally public data is excluded from these restrictions. A quota of up to 0.1% of this data may be tolerated if it has been cleaned, never appears in the model's results, and does not provide a substantial improvement to R&D capabilities for key sector technologies.

International Access: Open Source Exemptions and European Vulnerabilities

The proposed controls could lead American companies using open-weight models to favor domestic alternatives over Chinese models. The situation becomes more complicated for Europe, which relies on the use of open models in local clouds and the on-site hosting of high-level closed models. This sovereignty approach weakens if the main open models come from the United States or China and risk being blocked by export restrictions. AI systems already released as open source or published in scientific articles are not affected by these new provisions. Furthermore, Beijing claims it wants to support AI development in the Global South and facilitate certain exports within the framework of the World Organization for Cooperation in Artificial Intelligence.