Beijing Weighs Restricting Foreign Access to Top Chinese AI Models

China considers reversing its AI openness strategy

Chinese authorities have initiated discussions with major AI companies including Alibaba, ByteDance, and Z.ai about potentially restricting foreign access to their most advanced models, according to Reuters, which cited three people familiar with the talks. Officials have outlined options ranging from blocking public releases entirely to limiting distribution to domestic users only.

The deliberations represent a dramatic potential shift for Chinese AI firms, which have built their international presence by releasing models as open-weight systems. This approach allows anyone worldwide to download model weights and run the systems independently—a strategy that has enabled Chinese companies to gain global traction despite their models trailing American counterparts by an average of seven months on key benchmarks.

Why it matters

This potential policy reversal would eliminate the primary competitive advantage Chinese AI companies have used to penetrate global markets. By offering free, downloadable models, firms like Alibaba and ByteDance have attracted cost-conscious businesses seeking alternatives to proprietary American systems. A restriction would signal that Beijing now views national security concerns as outweighing the commercial benefits of open distribution—mirroring Washington's recent decisions to limit access to Anthropic's Mythos and OpenAI's GPT-5.6 models.

Security concerns drive both superpowers

Scott Singer, a fellow in the technology and international affairs program at the Carnegie Endowment for International Peace, notes that China faces the same security calculus the White House has confronted. "China will need to reckon with the reality that models that reach certain capabilities are unsafe," Singer said, adding that Beijing must balance global market access against controlling technology central to national security.

The timing follows Z.ai's recent release of GLM 5.2, which the company claimed matched Mythos' capability to identify software vulnerabilities. Once model weights are published online, they cannot be recalled or retrofitted with safeguards, creating persistent security challenges that proprietary models avoid.

Escalating tensions over model distillation

The potential restrictions emerge amid intensifying disputes over "distillation"—the practice of using outputs from advanced AI models to improve weaker ones. Anthropic published research in February alleging that Chinese firms DeepSeek, Moonshot, and MiniMax generated 16 million exchanges with Claude through approximately 24,000 fraudulent accounts. In June, Anthropic reportedly sent a letter to U.S. officials accusing Alibaba of attempting to distill Claude's capabilities.

The controversy escalated when a developer discovered code in Anthropic's Claude Code designed to detect Chinese users by analyzing time zones and network addresses. The Information reported Friday that Alibaba responded by banning Claude Code internally and requiring employees to remove Claude from work computers. An Anthropic engineer stated the detection code was added in March to combat distillation and would be removed.

Any restrictions Beijing implements would only affect future model releases, as published weights cannot be retracted. No final decisions have been made, and the involved ministries have issued no official statements.

These details were first reported by Reuters and the Information.