What happened
Reuters reports: Beijing is considering restricting foreign access to China's top AI models.
According to three sources familiar with the discussions, Chinese authorities have spent the past month meeting with leading tech companies about potentially limiting overseas access to advanced Chinese models — including ones not yet released.
The meetings were convened by China's Ministry of Commerce; Alibaba, ByteDance, and the creators of GLM-5.2, Z.ai, were among the participants. The scope of any potential restrictions is still being debated, and they may apply only to future models. Whether and when they'll actually take effect remains unclear.
Another concern driving the talks
Two sources also say Chinese authorities are deeply concerned about the possibility that software vulnerabilities discovered by the Claude Mythos model could be exploited, and that Washington might use the model against Chinese interests. The threat is being treated as real.
Where this leads
I've said many times here and in other channels that I think cheerleading for Chinese open source is misguided, and that I don't see much of a future for open models. Almost every model with serious economic potential will be proprietary — nobody is going to release a model that could plausibly bring in, let's say, $10–30 billion by automating economically valuable tasks. Even if companies themselves wanted to release those models, they'd almost certainly not be allowed to. Maybe this doesn't happen in 2026. Maybe not even in 2027 — but the further into the race we go, the more restrictions pile on.
The choice, essentially, will come down to two options: proprietary US models or proprietary Chinese models — and that's if there's any choice at all. Non-citizens or countries outside the allied coalition might not get even that, which is entirely plausible. "Weak" local models will still exist, but there's no guarantee they'll advance much beyond what's publicly available right now (roughly Mythos-class).