China's Ministry of Commerce and Ministry of Science have published a draft update to the country's export control list. The new regulations cover 23 key AI technologies, including large models, agents, compute clusters, and data processing tools.
Export bans apply to AI tools capable of generating or modifying malware for attacks on critical infrastructure or government agencies. Other software for identifying vulnerabilities will face strict export limitations. Models trained on sensitive data are also restricted, though exceptions exist for public data and models with minimal reliance on restricted datasets.
Controls on core model technologies, such as weights and architecture, trigger if a model reaches 10^26 FLOPs of training compute, contains at least 2 trillion parameters, or ranks in the global top 10. Officials determine this status based on performance, safety, and usage metrics from major evaluators.
To support the domestic open-source ecosystem, China is introducing a fast-track process. Companies publishing restricted AI tech as open source qualify for automatic licensing. Similar rules apply to firms operating overseas data centers or conducting research with foreign partners, allowing for immediate clearance upon filing paperwork.
While API access, SaaS tools, and lightweight local versions remain exempt, installing restricted AI technology on foreign servers requires a standard license. China also intends to facilitate AI capacity building within the World Artificial Intelligence Cooperation Organization. Until final rules are enacted, companies must register affected exports via the commerce ministry's portal.
