China Tightens Export Controls on AI Technologies and Large Models

The Chinese government has announced plans to tighten export controls on artificial intelligence (AI) technologies, including large models and key areas such as frontier models, AI agents, and autonomous malware tools. The move is part of a draft rule released by the commerce and science ministries, which targets 23 critical AI technologies across various categories.

The restrictions will apply if an AI system was trained on sensitive Chinese data or meets specific performance thresholds: it must have at least 2 trillion parameters, reach 10^26 FLOPs of training compute, or achieve a global top-10 ranking. These conditions are intended to prevent the unauthorized transfer of advanced AI capabilities.

The draft rule also introduces an automatic approval process for open-weight releases, allowing domestic companies to publish restricted AI technologies and weights as open source without further review. This fast-track system aims to support China’s thriving open-source ecosystem while maintaining control over sensitive exports.

However, certain AI tools will face new export limits or a total ban on international transfer. For instance, autonomous malware tools designed to attack critical infrastructure, the military, or government agencies are prohibited from being exported entirely. On the other hand, software vulnerability detection tools won’t be banned outright but will require additional permits for foreign deployment.

The restrictions also apply to large AI models in two scenarios: when they’re trained on China’s ‘important data’ (as defined by the Data Security Law and related regulations) and offer services based on it. Publicly available lawfully obtained data is exempt from these controls, as are tiny amounts of up to 0.1 percent if properly cleaned beforehand and not used in AI outputs.

The second case covers core technologies behind large models, including weights, checkpoints, source code, architecture, and training or inference methods. Export officials will determine the top-10 status using public rankings from major domestic and foreign evaluators or industry platforms, considering overall performance, specific tasks, safety, or global usage numbers.

This move has sparked concerns among US companies that rely on open-weight AI models. Dean W. Ball, head of strategic futures at OpenAI, notes that Beijing’s framework mirrors the model-weight export controls implemented by the US government. This may prompt US firms to focus more on domestic models rather than Chinese alternatives.

Europe is also facing challenges due to these new regulations. The EU has been relying on running open-weight models in local clouds to maintain independence and host top-tier closed models on servers within their borders. However, this strategy relies heavily on access to the best global AI models, which may be restricted by export controls from China or the US.

The Chinese government is also introducing a fast-track process for companies that register affected exports through its commerce ministry’s portal until final rules take effect. This streamlined system will allow eligible exporters to clear their applications automatically upon filing paperwork.