China Drafts Tight Export Limits on Frontier AI Models

China's ministry of commerce and ministry of science and technology have published draft rules introducing strict export controls across 23 critical artificial intelligence sectors. Under the proposed framework, restrictions trigger automatically for systems trained using sensitive Chinese data or frontier systems exceeding specific technical thresholds. Specifically, the rules target models built with 10^26 FLOPs of training compute, those containing at least 2 trillion parameters, or systems that secure a spot in global top-10 evaluations covering performance, safety, or deployment scale. Furthermore, software capable of autonomously generating or modifying malware to breach government, military, or critical infrastructure targets will face an absolute export ban.

To prevent domestic research from stalling, the regulatory proposal incorporates streamlined exemptions. Technology and model weights that are already in the public domain are completely excluded from the restrictions. For newly developed open-weight releases, Chinese institutions can obtain automatic licensing upon filing registration paperwork. This accelerated pathway also covers domestic companies running data centers abroad, fabricating chips internationally, or sharing technical standards through regional initiatives like the World Artificial Intelligence Cooperation Organization. While lightweight local implementations and API-based SaaS services remain unhindered, deploying covered AI technologies directly onto foreign local servers will demand standard export licenses.

For developers and enterprise practitioners, this regulatory shift reshapes the global open-source ecosystem. The framework provides Beijing with discretionary power to block future open-weight distributions, mirroring similar export restrictions established by the United States. Organizations outside China that depend on imported open-weight architectures may face supply uncertainty, forcing a pivot toward domestic alternatives. In particular, European cloud providers relying on foreign open-source weights to establish sovereign infrastructure could see their strategy disrupted if Beijing chooses to block future model iterations.