China’s Ministry of Commerce (MOFCOM) and Ministry of Science and Technology (MOST) have issued Announcement No. 48, a significant regulatory shift that adds 40 technologies—including 23 related to artificial intelligence—to the nation's Catalogue of Technologies Prohibited or Restricted from Export. While China has previously regulated AI-related technologies like personalized recommendation algorithms, this new draft represents the first time the government has established explicit technical thresholds for frontier AI models. The controls target core components such as model weights, checkpoints, source code, and training/inference optimization know-how.

The regulations define 'frontier' status using three specific metrics: a model must meet at least one of the following to be restricted: a training compute of at least 10^26 FLOPs, a parameter count of at least 2 trillion, or a global top-10 position on public evaluation rankings or call volume. Beyond the models themselves, the announcement introduces controls on AI agents, specifically targeting planning, multi-agent coordination, and tool-use capabilities. For example, agent planning is restricted if it achieves a 20-step subtask chain with over 80% decomposition accuracy. Furthermore, the rules prohibit the export of AI-enabled offensive cyber capabilities, such as automated malicious code generation and adaptive evasion techniques.

A critical distinction in the new framework is the difference between service access and technology custody. MOFCOM guidance suggests that using a model via API, web interface, or Software-as-a-Service (SaaS) does not count as exporting a large model. However, local deployment of a restricted model on infrastructure outside China—a common choice for organizations prioritizing data sovereignty—does trigger export management requirements. This creates a complex compliance landscape for multinational corporations, as the rules also capture transfers of ownership, licensing deals, and even intra-group code sharing. While the government states these measures aim to prevent 'uncontrolled diffusion' while protecting legitimate cooperation, the 'global top-10' ranking remains a subjective criterion, as the announcement does not specify which ranking bodies or metrics are authoritative for determining a model's status.