China’s Ministry of Commerce has held a series of meetings with the country’s top AI companies over the past month to discuss restricting overseas access to China’s most advanced artificial intelligence models, including models that have not yet been publicly released. The discussions, confirmed to Reuters by three people familiar with the matter, mark a significant shift in how Beijing is thinking about its AI exports, and signal that the government is beginning to treat frontier AI models as a strategic national asset rather than a commercial product.
The meetings were led by the Ministry of Commerce, with officials from the National Development and Reform Commission also in attendance. Companies present at the discussions included Alibaba, ByteDance, and Z.ai, the Hong Kong-listed firm behind the GLM-5.2 model that has drawn intense interest from US developers in recent weeks. The scope of potential restrictions remains under discussion, and it is unclear when, or even whether, any formal measures will take effect.
Three Pillars of the Proposed Framework
According to the sources, officials discussed three distinct areas of potential action. The first concerns placing limits on the most advanced AI models — both closed-source proprietary systems and open-weight models that can be downloaded and run independently. The second involves making the leaking or theft of AI model weights an offense under China’s national security law, a step that would criminalize the unauthorized extraction of AI model weights that US officials have accused Chinese actors of conducting against American labs. The third area covers new restrictions on who can fund domestic AI startups, a measure that would give Beijing tighter control over the capital flows shaping China’s AI ecosystem.
The proposal to restrict open-weight models is particularly significant. China’s open-source AI strategy has been central to its global expansion: DeepSeek’s R1 model, released under an open license in January 2025, triggered a rout in US technology stocks and established Chinese AI as a credible alternative to OpenAI and Anthropic. Z.ai’s GLM-5.2, released in June 2026 under an MIT license, has since become the fastest-adopted model on Vercel’s developer platform and now ranks fifth on OpenRouter by usage. Restricting the overseas release of future open-weight models would fundamentally change that dynamic.
The Mythos Trigger and the Security Calculus
The discussions appear to have been partly triggered by US actions in the AI space. Chinese authorities are deeply concerned about the Mythos model developed by Anthropic, which US officials placed under export controls in June. While export controls on Anthropic’s Fable model were subsequently lifted after new safeguards were put in place, Mythos remains restricted to a limited set of trusted US organizations. Chinese officials have reportedly expressed worry that Mythos-class models could be used to exploit software vulnerabilities in Chinese systems. Zhou Hongyi, founder of cybersecurity firm 360, publicly called for China to develop its own equivalent of Mythos.
The security framing extends beyond Mythos. In April 2026, China ordered Meta to unwind its $2 billion acquisition of the Manus AI startup, and launched investigations into Manus and other local AI startups that had moved operations abroad. The Supreme People’s Court’s academic journal has also proposed a tiered regulatory system for AI models: basic open-source models would require only a filing, advanced models would undergo security review, and frontier models would be barred from public release or restricted to domestic use only. That proposal now appears to be influencing the Ministry of Commerce discussions.
What Is at Stake for the Global AI Market
The timing is notable. GLM-5.2 has, in the words of one Reuters source, “set Silicon Valley abuzz” with capabilities that approach those of leading US offerings at a fraction of the cost. On OpenRouter, the six most popular AI models are now Chinese. US companies including Lindy, Coinbase, and LaunchLemonade have publicly shifted workloads to Chinese models to manage costs. Any restriction on future Chinese model releases would remove a significant source of competitive pressure on US AI pricing.
The potential measures also raise questions about the future of China’s open-source AI strategy. Beijing has used open-weight model releases as a form of soft power, building developer ecosystems in Southeast Asia, Africa, and Latin America that are increasingly dependent on Chinese AI infrastructure. Restricting that flow would protect Chinese frontier capabilities from foreign exploitation but could also slow the global adoption of Chinese AI standards.
(Related: US Lifts Export Controls on Anthropic’s Fable 5 and Mythos 5 — and China AI Filled the Gap)
The Ministry of Commerce and the companies named in the Reuters report did not respond to requests for comment. The scope of any final measures and the timeline for their implementation remain uncertain. But the fact that Beijing is now treating its most advanced AI models as assets requiring export controls, mirroring the logic of the US chip restrictions that have defined the tech war for the past four years, represents a structural change in how the two countries are approaching the AI competition.
(Related: Why Ad Hoc AI Export Controls Will Always Help China — A PIIE Analysis of the Fable-Mythos Saga)