Beijing Is Moving to Lock Advanced Chinese AI Models Inside Its Borders
China is preparing to restrict overseas access to its strongest AI models, according to Reuters reporting published July 7. Beijing has held discussions with Alibaba, ByteDance, and Z.ai about keeping advanced models domestic — including systems not yet released. The Ministry of Commerce led the meetings, with China’s state planning agency also present.
The involvement of the state planning agency alongside Commerce is the key signal: this is being framed as export control, not routine platform regulation.
What Gets Restricted
The targets are both closed models — API access to systems like DeepSeek V4 Pro, Qwen 3.6+, and GLM-5.2 — and open-weight systems, where the downloadable weights themselves would be treated as controlled technology. That distinction matters. The open-weight question is harder to enforce but more consequential: Chinese open-weight models currently represent the primary alternative to US lab pricing at the frontier tier.
Officials also discussed treating leaks or theft of proprietary AI as a national security offence, not merely an IP dispute. Foreign funding restrictions for Chinese AI startups were part of the same conversation, effectively tightening control over capital, talent, and model access in a single policy sweep.
The likely implementation path, per Reuters sources, is tiered: basic open tools require filings, stronger systems face formal reviews, and frontier models remain domestic only.
Timing and Trigger
The timing maps directly onto the US moves against Anthropic and OpenAI. Washington restricted Claude Mythos access in late June, triggering emergency sessions at the Bank of England and US Federal Reserve and a 19-day blackout for most users. Beijing has watched that episode as evidence that AI models are now strategic infrastructure — and drawn its own conclusion about who controls the chokepoints.
China also fears Mythos specifically could be used to find software vulnerabilities in Chinese systems. That concern runs in the other direction too: US officials have restricted Chinese model access partly on the basis that DeepSeek and its equivalents could identify weaknesses in US infrastructure or be used for intelligence purposes.
Beijing has separately investigated Chinese AI startups that relocated abroad, and ordered Meta to unwind its $2 billion acquisition of Manus. The pattern suggests a comprehensive posture shift — not a narrow reaction to the Mythos episode.
What It Means for Open-Source AI
The open-weights question is the hardest part. GLM-5.2 from Z.ai was described as the first Chinese model to match US frontier labs; Marc Andreessen called it a “step change.” Kimi K2.7-Code is already in GitHub Copilot’s model picker. DeepSeek V4 Pro holds an MIT licence that has made it the most cost-effective coding model in production for thousands of developers outside China.
If Beijing imposes formal controls on open-weight release, the flow of high-quality, low-cost model weights from Chinese labs to global users slows or stops. Open-source alternatives that have compressed the frontier gap from 13 to 6 intelligence index points over the past year would cease arriving.
Washington has already built a two-tier frontier access system through its export control machinery. If Beijing mirrors it, the global AI stack fragments into geographically distinct model pools — with price and capability implications for every developer, enterprise, and government operating between them.
No formal regulations have been published. Reuters reports the discussions are ongoing. The timeline is unclear, but the fact that the meetings are happening at ministerial level suggests the decision has already moved past the question of whether and into how.