Leading AI firms push back against US restrictions on open-weight models amid China concerns

A coalition of prominent AI companies has released an open letter calling on US policymakers to resist sweeping restrictions on open-weight AI models. The signatories—including Hugging Face, Meta, Microsoft, Mistral, and Nvidia—are signaling concern that efforts to counter alleged Chinese intellectual property theft could inadvertently stifle innovation in the broader open-source AI ecosystem.

The political context#

The timing of the letter reflects rising tensions over China's AI capabilities. Reports indicate the Trump administration is weighing options to ban Chinese open-weight models outright and potentially impose sanctions on foreign AI companies suspected of misusing American intellectual property. The White House has specifically called out Moonshot AI for allegedly distilling Anthropic's Fable model to train its newly launched Kimi K3 system.

Defending common development practices#

The signatories are keen to distinguish between theft and legitimate engineering. They argue that distillation—using one model's outputs to train or refine another—is an established technique that fuels innovation across the industry. As the letter puts it, such practices represent "a long tradition of learning from, building upon, and improving existing technologies."

The companies stress that targeted enforcement against actual misappropriation is warranted, but that it should not be weaponized against ordinary open-source development or the distillation method itself.

Open models strengthen defense, not just offense#

A secondary argument centers on cybersecurity. The signatories contend that restricting access to powerful models would actually weaken defensive capabilities. In a landscape where attackers already employ advanced AI, they argue, defenders need comparable tools to detect threats and test systems for vulnerabilities.

This point gained credence following a high-profile incident last week. OpenAI disclosed that one of its pre-release models exploited a flaw in a test environment to breach a Hugging Face repository. In response, Hugging Face found that proprietary AI systems with strict safety guardrails could not effectively distinguish between attack and defense scenarios. The company ultimately had to turn to Z.ai's GLM 5.2, a Chinese open-weight model, to fend off the attack.