China is studying limiting foreign access to its most advanced artificial intelligence models. The information is published by Reuters, which cites several sources familiar with meetings held over the last month between Chinese authorities and some of the country’s large technology companies. On the table is the possibility of impose controls on the most powerful systems developed in China, including models that have not yet been publicly presented.
According to the media, Beijing is beginning to treat artificial intelligence as a critical national asset that needs controlsa logic very similar to that already applied USA with its most sensitive models. Until now, much of the international advance of Chinese AI has been supported precisely in its availability and low costs. Since the emergence of DeepSeek R1 Early last year, Chinese models have gained ground outside the country because they offer competitive features at prices much lower than many American alternatives.
According to Reuters, the meetings included Alibaba, ByteDance and Z.ai. Alibaba develops qwenone of the most used families of AI models in China and with a growing presence abroad. ByteDance, the company that owns TikTokis behind Doubaoanother of the great Chinese systems. Z.ai has attracted attention in Silicon Valley with GLM-5.2a model that is close to the capabilities of the best American proposals, but at a much lower cost.
The conversations were led by Chinese Ministry of Commercewith also the presence of the National Development and Reform Commissionthe state economic planning agency. Reuters notes that neither these organizations nor the companies involved responded to requests for comment.
Possible restrictions are still being debated. The sources consulted explain that they could affect both closed models and more open versionsthat is, systems whose technology can be downloaded, run and adapted in other environments. Controlling a website or an API is relatively simple, but limiting the circulation of downloadable or modifiable models outside of China would be a much more delicate measure with more implications given that many foreign companies could lose a cheap alternative to American models and the AI market would again concentrate in fewer hands.
Among the measures discussed are Toughen penalties for leaks or theft of AI technology. One of the sources maintains that officials proposed treating those leaks as crimes under China’s strict national security law. There was also talk of restrict who can fund domestic AI startupswhich would expand control to the money that enters the sector and not only to the technology that leaves the country.
Reuters has not been able to confirm when those measures would come into force or if they will ultimately be approved. One possibility is that they only affect future models, not those already available.
The United States has already begun to apply restrictions to some of its most advanced models, the so-called frontier. In June, the administration trump ordered to restrict the access of foreigners to Fables and Mythosboth of Anthropic. The company ended up disabling them for all users because it could not verify their nationality. Fable later returned after incorporating new safeguards. Mythos, aimed at cybersecurity professionals, continues reserved for some US organizations considered trustworthy.
According to Reuters, Chinese authorities fear that Mythos could be used to find software vulnerabilities and act against Chinese interests. This concern has been expressed by state media and by Zhou Hongyifounder of cybersecurity company 360which has argued that China needs to develop its own Mythos.
Although the outlet has not been able to determine how potential new restrictions on foreign access to Chinese AI models would work, it does mention what was proposed at a roundtable held in May by Chinese legal experts on the regulation of open source AI. There they talked about a tiered system whereby basic open source tools would be subject to simple registration, more advanced technologies subject to security reviews and the frontier models, more critical, vetoed for public launch or restricted to national use.