China may restrict access to its most advanced AI models

Beijing is considering a move that would upend the strategy that made Chinese AI companies global players: restricting who outside the country can use their most advanced models.

China's Ministry of Commerce has held a month of meetings with Alibaba, ByteDance and startup Z.ai about limiting foreign access to top-tier systems, including ones not yet released, Reuters reported this week, citing three people familiar with the talks.

The majority of Chinese AI labs currently release their models “open weight", meaning that anyone can access and use the model without restrictions. This trend, which has been gaining speed ever since the release of DeepSeek’s R1 last year, is what made Alibaba’s Qwen and ByteDance’s Doubao popular worldwide despite lagging behind the American models by about seven months.

Z.ai’s GLM-5.2 attracted special interest from the Silicon Valley community due to its closeness to the leading American models while costing significantly less, and some American companies have even opted for the Chinese models to save money.

Blocking access would be to sacrifice the key to China’s AI expansion, an action that, according to experts, can only be justified when security considerations trump commercial considerations.

"China will have to face up to the fact that some models will eventually become dangerous," explains Scott Singer, a fellow with the technology and international affairs programme at the Carnegie Endowment for International Peace, pointing out that Beijing is in the same position as Washington was in the past months.

The US government has already imposed restrictions on foreign access to Anthropic’s Mythos model because of its capability to autonomously discover software vulnerabilities and has requested OpenAI to release a preview version of GPT-5.6 for evaluation by select partners for the same reason.

The talks follow weeks of escalating friction between the two countries' AI sectors. Alibaba has banned employees from using Claude Code effective July 10 after a developer found code in the tool designed to detect whether a user was in China or linked to a Chinese AI lab.

Anthropic said the code was added in March to combat "distillation", training a weaker model on a stronger one's outputs, and that it would be removed.