Anthropic open weight models are now at the center of a policy fight the company says it does not want to turn into a ban. CEO Dario Amodei wrote that Anthropic opposes a blanket prohibition on open weight models, including Chinese ones, while also calling for measures that would make some of the strongest open alternatives harder to build and distribute.
The fight follows an open letter organized by Nvidia that urged the Trump administration not to restrict open weight AI models. The letter argued that open weights increase competition, reduce customer lock-in, and let organizations run and adapt models on their own infrastructure instead of relying only on hosted systems from a few large providers.
OpenAI and Google were not initial signers, according to the campaign around the letter, but later added their names. Anthropic remained off the list. The difference is not subtle: Anthropic has not released its own open weight model, while several prominent Chinese systems, including Kimi and Qwen, are available in forms that developers can download or run outside the vendor’s hosted service.
What is an open weight AI model?
An open weight model makes the trained parameters available so others can run, inspect, fine-tune, or deploy the model themselves. That differs from a closed hosted model, where users send prompts to a company’s servers and get responses back through an app or API. For readers who want the machinery under the hood, Kernel has a plain explanation of how large language models answer prompts.
The policy stakes are practical. Open weights can let a company keep data inside its own systems, reduce dependence on one vendor, and customize models for narrow tasks. They can also push down prices by giving customers something other than a closed API bill and a prayer.
Why is Anthropic worried about open weight models?
Amodei wrote that banning Chinese open weight models would not address the security risks officials claim to fear, because malicious users would not need to be legitimate U.S. businesses. He also acknowledged that such a ban would shield U.S. AI companies from competition, while saying that protectionism is not his goal.
Anthropic’s preferred approach is narrower but still aggressive. Amodei called for cracking down on “industrial-scale distillation operations,” tightening chip export controls, and requiring pre-release safety testing for sufficiently capable models, whether open or closed, foreign or domestic.
Distillation is a training technique in which one model learns from the outputs of another, stronger model. Treasury Secretary Scott Bessent told Fox Business, according to CNBC, that the United States could sanction overseas models if officials conclude they are stealing from American companies. CNBC also reported that Anthropic sent a letter to the Senate Banking Committee alleging that Alibaba carried out what Anthropic called the largest known distillation attack against it.
Amodei’s argument is that distillation lets Chinese companies narrow the gap with U.S. frontier models despite chip limits. He wrote that distillation may bring Chinese systems within months of U.S. capabilities, even if it does not make them equivalent or superior.
The Nvidia-led letter took the opposite policy lesson from the same moment. It said relying only on closed models creates concentrated failure points and limits outside scrutiny. Open weight models, the letter argued, allow researchers and developers to test behavior, find vulnerabilities, build safeguards, and benchmark systems against real harms.
Amodei said he agrees that open weights can expand access, strengthen competition in some uses, and give customers more control. He disputed the claim that broad access necessarily helps defenders more than attackers, writing that the reverse may be just as likely.
That leaves Anthropic in an awkward position. The company says it does not want open weight models banned. It also wants legal and policy limits on the training methods and supply chains that help make competitive open weight models possible. For developers and customers, the outcome will shape whether advanced AI remains mostly a rented service from a few labs or becomes infrastructure they can actually run themselves.
This story draws on original reporting from Techdirt.
