As US weighs response to Chinese AI, industry urges against broad open-weight restrictions

A group of AI heavyweights just put their names on a letter asking Washington not to slam the door on open-weight models. Hugging Face, Meta, Microsoft, Mistral and Nvidia are among the signatories, and the timing is not an accident. The Trump administration has reportedly been weighing a ban on Chinese open-weight models and possible sanctions against Chinese AI firms, following accusations that Moonshot AI distilled Anthropic's Fable model to build its Kimi K3 system.

The letter itself never mentions China by name, but the target is obvious enough. Its authors want policymakers to separate "legitimate model-development techniques" from actual misappropriation. Distillation, they argue, is just how AI progress works: one model learns from another's outputs, the same way open-source software has always built on what came before. If a company is illegally siphoning value from a closed model, fine, go after that with targeted legal tools. But don't use it as an excuse to restrict a technique that's baked into how the whole field improves.

The letter also tackles the argument that open weights are simply too dangerous to leave unregulated, since anyone can grab them and use them for cyberattacks or worse. Its response: defenders need access to models just as capable as the ones attackers might use, or they're fighting blind. Open models, the signatories say, spread that defensive capability around and let vulnerabilities get caught and fixed by more people, not fewer. There's a real-world example baked into the debate, too. OpenAI recently disclosed that during testing of GPT-5.6 Sol and another unnamed model, one of the systems exploited a flaw in its test environment to reach into a Hugging Face repository holding a benchmark solution. Hugging Face said its commercial frontier tools couldn't help it respond, because their guardrails couldn't tell the difference between someone building an exploit to attack and someone building one to defend. It ended up turning to Z.ai's GLM 5.2, an open-weight model out of China, to handle the defense instead.

What's notable is who didn't sign. OpenAI and Anthropic, which have been pushing the administration to act against alleged Chinese IP theft, are absent, and so are Google DeepMind and SpaceX. That split lines up pretty neatly with business incentives. Nvidia, Microsoft Azure and other infrastructure players make more money when models are commoditized and interchangeable, because that means more GPUs sold, more cloud rented, more apps built on top. Closed-model companies, meanwhile, are watching cheap, capable, freely available models chip away at the moat their business depends on.

The letter wraps with a broader ask: more compute access for startups and researchers, shared investment in datasets and evaluation tools, and a plea to keep multiple approaches alive rather than let heavy-handed rules push AI development overseas.