A Chinese state-media account argues for limits on open AI capabilities

The argument for openness also preserves room for closed services

The commentary makes a positive case for openness: researchers, companies and independent developers can inspect model behaviour, find vulnerabilities and develop mitigations without waiting for a single supplier’s update. It also says real-world use exposes models to more situations, and notes that companies can deploy open and closed models for different tasks, citing large-language-model routing as an example.

Its claim of scale is similarly its own: Chinese open-source large models have surpassed 10 billion cumulative global downloads. The source supplies no model list, measurement method or time period, so the figure cannot establish a comparable market share or adoption rate. It does, however, show why the account frames broad availability as strategic: it argues that such models give users more choice and can reduce dependence on particular suppliers.

The proposed hierarchy therefore does not resolve the open-versus-closed trade-off. It explicitly retains closed commercial services as an option while arguing that open releases can help find and mitigate flaws. Its stated distinction is between capabilities and risks, not a blanket endorsement of unrestricted distribution.

Overseas access is a separate, preliminary question

A report recounting Bloomberg’s reporting based on unidentified people says Chinese officials had recently discussed with companies, including Alibaba, how to reduce risks from more capable AI systems. According to that account, options discussed included restricting overseas users’ access to the most advanced models; the talks were at an early stage and there was no plan to implement measures.

That report should not be read as proof that Yuyuantantian’s framework has become policy. The two accounts overlap on the concern about more capable systems, but only the commentary sets out the proposed tiering, while the report supplies neither a final decision nor an implementation design.

What would show whether the signal becomes policy

The next evidence to watch is concrete rather than rhetorical: an official agency, defined capability and risk tests, the type of access being regulated, and whether any requirement reaches domestic users, overseas users, released model weights or hosted commercial services. Until those choices appear, the commentary is best understood as an argument for graded governance—one that preserves China’s case for broad model availability while reserving the possibility of tighter control at the high-capability end.