Beijing is drafting rules that would block foreign users from China’s most advanced AI models, including ones that haven’t shipped yet. That’s the headline from Reuters’ report on Ministry of Commerce meetings with Alibaba, ByteDance, and Z.ai over the past month.
If you’re running a product on Qwen, Doubao, or GLM, read the fine print here. Officials floated a tiered system: basic open-source tools get a simple filing, “advanced technologies” get a security review, and frontier models get walled off from public release entirely. Nothing’s final and any curbs would likely hit future models only, but the direction is unambiguous. China wants its frontier models treated like a national asset, not an export product.
This isn’t happening in a vacuum. Alibaba just put Anthropic’s Claude Code on a high-risk internal list and told staff to switch to its in-house Qoder assistant, after Anthropic accused Alibaba’s Qwen lab of running the largest known distillation attack to date, per CNBC’s reporting. Both governments are converging on the same conclusion at the same time: model access is a security perimeter now, not a licensing line item.
For any founder building on a Chinese open-weight model, that’s a vendor-lock risk you didn’t price in six months ago. If your roadmap depends on Qwen or GLM staying freely downloadable, get a contingency model sourced before this becomes policy, not after. Separately, China’s new rules on AI companion apps hit July 15, forcing Doubao and Qwen to kill custom-persona features already.
Nathan Zakhary