China Mulls Tiered Limits on Overseas AI Model Exports

China’s Ministry of Commerce held meetings this month with Alibaba, ByteDance and startup Z.ai to discuss restricting overseas access to advanced and unreleased AI models. Officials outlined a three-tier framework: basic open-source tools would require a government filing; more advanced models would face security reviews before release; and the most sensitive frontier systems could be barred from public release or limited to domestic use.

The talks addressed both closed-source models and open-weight models that developers can download and run locally. Participants also considered making unauthorized disclosure or theft of proprietary AI technology a national security offense and tightening which investors may fund domestic AI startups.

No formal timeline or final rules were announced. Some participants indicated any new limits might apply only to models created after the rules take effect rather than to systems already deployed. Government officials and company representatives continue to negotiate the scope of restrictions and enforcement mechanisms.

Alibaba, ByteDance and Z.ai are among China’s largest AI developers. Alibaba’s Qwen models have been widely distributed on public repositories. ByteDance’s Doubao is commonly used inside China. Z.ai’s GLM-5.2 has drawn attention for strong benchmark results and low-cost API access. Proposed limits would reduce cross-border availability of models some developers use as lower-cost alternatives to leading U.S. systems.

U.S. authorities applied export controls to deployed AI models in June and asked developers to preview new frontier models with vetted partners. Chinese regulators monitored those actions and raised concerns that certain cybersecurity models could be reverse-engineered to target Chinese infrastructure. Allegations that some models contained code collecting information about Chinese users also intensified scrutiny.

Chinese authorities have already tightened other technology controls, including requiring government approval for some foreign investments in AI firms and blocking deals that raised security concerns. Several domestic platforms removed or limited humanlike agent features ahead of rules scheduled to take effect in mid-July.

Analysts note that nationality-based access controls can exclude foreign-born engineers who work at leading labs and who help fix vulnerabilities and maintain systems. They also note that restricting open-weight models would shrink the global supply of accessible AI models used by businesses seeking lower-cost options.

OpenAI commented that “government-gated access shouldn’t become the long-term default.” At present, officials are still weighing technical criteria for each tier, how to treat already released models and how investor rules would be enforced. No final regulation has been published, and it remains uncertain whether the proposed framework will be adopted.