China AI policy

Models exceeding 2T parameters or 10^26 FLOPS training compute will require export approval, and public open-source releases could be restricted — release strategies must be re-evaluated

Transfers of AI technology, models or related technical materials abroad (including to ASEAN) will require prior export licences, raising deal timelines and compliance costs

Exporting non-public large model, agent, or computing cluster technologies via licensing, sale, or weight delivery requires interim registration with MOFCOM during the transition; in-flight transactions are also covered, with no grandfathering.

Export of large models, AI agents and computing cluster technologies would become restricted, so cross-border licensing, partnerships and deployments need export-control review

Export of model weights, source code, and training methods above the 10^26 FLOPs or 2 trillion parameter thresholds will require approval before release, delaying or blocking cross-border collaborations

Cross-border export or transfer of AI, biopharma or solar technology will require approval, so companies must plan licensing ahead

Chinese companies transferring AI models, algorithms or related technology abroad will face new export licensing requirements.

Companies exporting large model, agent, or computing cluster technologies must complete export licensing or pre-export registration for the 23 AI items

Overseas localised deployment of restricted large models falls under export control and requires a licence (automatic licence available for deployment on overseas self-owned cloud); API calls, SaaS apps and lightweight on-device deployment are outside scope.

Obligations: Companies in the electronics manufacturing sector must align their R&D and production strategies with the plan's goals, particularly in developing advanced chips and promoting domestic operating systems.

Clarifies the judgment rules for infringement behaviors such as AI face change and big data discrimination.

Obligations: Companies must comply with the classification and grading system, ensure content is properly labeled, and strictly follow platform requirements for portrait and voice authorization.