China Proposes Export Restrictions on AI Models Over 2 Trillion Parameters, Exempting Open Source and APIs
A draft regulation from Chinese authorities targets core tech of massive AI models for export control but explicitly exempts open-weight releases and API services.
ImportanceMaterialEvidenceE3 inspectableWrite-upQuick
According to a report by ZDNet Korea, China's Ministry of Commerce and Ministry of Science and Technology published a draft revision on October 16 proposing to add 40 technologies in AI, biopharma, and solar power to the export prohibition or restriction list. Twenty-three of these are AI-related.
The draft sets a specific threshold: core technologies of AI models with more than 2 trillion parameters or those possessing world-leading performance in specific capabilities are classified as subject to export restrictions. Additionally, AI agent technologies used in defense and critical information infrastructure are proposed for outright bans.
Crucially, the draft clarifies exemptions. Technologies disclosed via open-source software or academic papers are generally excluded from management. Furthermore, Chinese companies providing AI services to overseas users via APIs or websites are not considered to be engaging in restricted 'technology exports.' This implies that the current usage and download of widely adopted open-weight models like DeepSeek and Qwen will not be immediately restricted.
This move is viewed as a response to previous US restrictions on advanced AI chip exports to China. For global developers relying on the Chinese open-source ecosystem, while existing models remain accessible, the compliance path for acquiring the underlying architecture of future frontier models may become complex.