China proposes AI export curbs to manage security

The draft proposal covers specific AI technologies like large language models and computing clusters. It includes restrictions on model weights and source code for products exceeding certain performance thresholds. For example, core technology for models with at least 2 trillion parameters, such as Kimi K3 or Alibaba's Qwen3.8-Max, or those ranking among the world's top 10 on measures including performance, usage, and safety, would be subject to these curbs.

The restrictions would also apply to AI models trained on China's "core data," defined by legislation like the Data Security Law as data concerning national security, the national economy, key public livelihoods, and major public interests. However, open-source technology and research already in the public domain would not be covered, with the government planning streamlined procedures for their export.

This move comes as Chinese AI startups have recently made progress, with models like Moonshot AI’s Kimi K3 and DeepSeek’s latest editions offering competitive pricing. These companies have also looked to markets beyond China, with some making their offerings available on US cloud service providers. The new guidelines, including a temporary registration requirement for exports, are likely to introduce further uncertainty in an industry already navigating content controls and a licensing regime.