DeepSeek and MiniMax register under China's draft AI export controls

Compliance by major developers

DeepSeek and MiniMax have officially signed up under Beijing’s draft AI export rules. According to a CCTV-affiliated account, Yuyuan Tantian, these two developers have also finished the process of open-sourcing and securing licenses for overseas use of their models.

The move comes as the Chinese government seeks to formalize how artificial intelligence technology moves across borders. By registering, these companies are aligning their international operations with the requirements set out by the Ministry of Commerce and the Ministry of Science and Technology.

Overseas internet users have reportedly encouraged other Chinese open-source models to follow this same path. This voluntary alignment suggests a proactive approach by developers to maintain international access while complying with domestic regulations.

Defining the restricted list

The revised Catalogue of Technologies Prohibited and Restricted from Export identifies 23 specific AI technologies. These include large models, agents, compute clusters, and various forms of information processing. The list was created to provide a clearer framework for what constitutes a sensitive export.

The draft delineates three primary areas of concern for the government. One category involves models trained on core national data; these are barred from export regardless of their size or number of parameters. Another category focuses on agent technology that exceeds certain performance levels.

The third category covers compute infrastructure, specifically the design of AI chips and high-speed interconnects. The rules set numeric thresholds, such as training compute exceeding 10^26 floating-point operations or models with more than 2 trillion parameters.

Licensing and exceptions

While the rules create restrictions, they are not intended to shut down the entire AI industry. Yuyuan Tantian noted that most artificial intelligence technology will continue to be exported freely once the measures are officially in effect. The focus is on the act of exporting technology rather than the technology itself.

Many activities are expected to qualify for automatic licensing, which is described as a nearly seamless file-and-go process. These include open-sourcing AI, running self-built compute centers abroad, and joint research and development with foreign resources. International standards work also fall into this category.

Companies can still seek licenses for restricted items without having to cancel existing overseas plans. Lawyer Liu Hongmei stated that restricted items can still win licenses, providing a pathway for companies to continue their international work under government oversight.

What is not considered an export

The draft clarifies that simply providing access to a model does not always count as an export of the technology itself. Providing API access, interacting with a model via an app or web interface, or offering SaaS services are not classified as exports under these rules.

Lightweight on-device deployment of a restricted model is also excluded from the export definition. This distinction is important for how developers can distribute their tools to global users without triggering the same level of scrutiny as hardware or core model transfers.

The government's Q&A session spent more time explaining what the rules do not cover than what they do. This suggests a desire to maintain the flow of standard software services while tightening control over high-end infrastructure and sensitive data.

The global context

The current revision is considered routine work under existing laws, as China has revised the catalogue several times since 2001. However, the draft has still triggered significant questions in foreign policy circles regarding the future of Chinese AI exports.

In contrast, Washington has taken different steps, such as instructing federal agencies to discontinue the use of DeepSeek products. This move was signed by Secretary Markwayne Mullin in July 2026, following a framework established in early 2025.

China's approach involves categorizing restrictions by country and alliance. By establishing these clear thresholds, Beijing aims to manage its technological footprint while navigating the complex landscape of international AI competition.

Next steps for the industry

The primary question remaining for companies is whether there will be clearer operating guidance as the draft moves toward finalization. Many concerns raised by companies during the consultation period were incorporated into the current draft to address practical hurdles.

Members of the World Artificial Intelligence Cooperation Organization are expected to receive stronger support for building capabilities. This organization will play a role in making rules together and fostering international cooperation on AI standards.

Developers will need to monitor the specific numeric thresholds for compute and parameters as they may shift with the industry. Staying updated on these specific requirements will be essential for any company planning to deploy large-scale models globally.