Thirty-five companies signed a letter telling Washington not to restrict open-weight AI models

Thirty-five names sit at the foot of the letter. Nvidia, Meta, Microsoft, OpenAI, IBM, Cisco, Dell Technologies, Palantir, Palo Alto Networks, ServiceNow, Andreessen Horowitz, Y Combinator, Hugging Face, Mistral, Mozilla, GitHub, Perplexity, Cohere, CrowdStrike and the Linux Foundation are among them, along with smaller research outfits including Arcee AI, Black Forest Labs, Nous Research, Prime Intellect and Reflection. Anthropic is not a signatory. The document is titled Open Weights and American AI Leadership, it is published on Microsoft's corporate responsibility site, and it asks policymakers to keep the frontier plural by avoiding premature restrictions on open models.

The asks themselves are short. Expand access to compute for startups and researchers. Invest in shared training assets, by which the letter means datasets, tools and evaluation frameworks. Avoid restrictions that the signatories say would stifle competition or drive innovation overseas.

Most of the specific language goes to distillation, the practice of using one model's outputs to help train or improve another. The letter asks policymakers not to conflate legitimate model development techniques with misappropriation, calls distillation a widely used method for improvement, evaluation and validation, and says that unlawful efforts to extract value from closed models should be addressed through targeted legal and commercial frameworks rather than sweeping restrictions on techniques the signatories describe as important to innovation. Rebecca Bellan at TechCrunch, who reported the letter on Friday, wrote that it arrives while the administration considers how to respond to Chinese open-weight models.

The letter does not claim the risks are imaginary. It states that once weights are released they are beyond the original developer's control, and that modified versions are difficult to trace or reverse. Its answer is that prohibition is the wrong response to that, because defenders facing attackers who use advanced AI need models of comparable capability, and because concentrating capability behind a small number of closed models creates a small number of single points of failure. Openness, it argues, may be one of the most important paths to AI safety and security.

One thing the document is not is a filing. It names no agency, no docket and no comment deadline, and it is addressed to policymakers generally rather than to a proceeding. Nothing in it obliges anyone to respond.