China allows DeepSeek and MiniMax to export their models to the world

An account named "Yuyuan Tantian," affiliated with China Central Television (CCTV), stated that DeepSeek and MiniMax have completed the registration process and obtained open-source licensing and overseas export rights, weeks after Chinese authorities announced new rules for exporting AI technologies. The post, reported by Reuters, added that most AI technologies will remain allowed for export.

On October 16, China's Ministries of Commerce and Science proposed adding 23 AI technologies to the list of technologies prohibited or restricted for export, including large models, agents, and computing elements. The rules require companies to register before exporting any listed technology, with the public comment period running until November 15.

According to the post, the rules target three areas. First is data: exporting models trained on Chinese "core data" is prohibited, and exporting models trained on "important data" is restricted regardless of size. Second is sensitive uses: agents exceeding a certain performance threshold are subject to restrictions, and those dedicated to critical infrastructure, defense, or essential government affairs are prohibited. Third is computing: incorporating AI chip design and high-speed interconnects for large clusters.

Providing a restricted model via API, applications, websites, SaaS services, or lightweight local deployment remains outside the scope of "technology export," meaning companies can continue their commercial operations as usual. Open-source technologies are automatically granted a license, while technologies previously published in academic papers or open sources fall outside the scope of the rules. DeepSeek and MiniMax did not respond to requests for comment, and TNW stated that it had not independently verified the post.

What do these terms mean?

Technology Export: The transfer of knowledge, software, or designs outside the country, which in some nations is subject to prior licensing if deemed sensitive to national security.

Open-Source Model: An AI model published with its code or weights for everyone to use and modify, facilitating its spread beyond the developing company's borders.

Training Compute: The amount of processing power used to train a model, measured by the number of computing operations, and used as a benchmark for classifying large models.