Meta, Microsoft and Nvidia Back Open-Weight AI

Twenty-four companies and organisations, including Meta, Microsoft and Nvidia, signed an open letter published as a PDF urging US policymakers to protect the ability to publish trained AI model weights.

The letter defines open-weight models as systems whose trained parameters, or weights, are published so anyone can download, inspect, modify and run them on local hardware. That contrasts with closed models delivered only through vendor-controlled application programming interfaces, where the underlying weights remain on the provider’s servers.

Signatories include chip and cloud providers, enterprise software firms, startups and research groups. Named backers in the document include Nvidia, Microsoft, Meta, IBM, Dell Technologies, CrowdStrike, Palantir, ServiceNow, Hugging Face, Perplexity, Mistral, Andreessen Horowitz, Y Combinator, the Linux Foundation and Mozilla.

The letter lists three primary benefits of open weights. It says publishing weights lowers the cost of entry for startups, public institutions and other organisations that cannot afford to train large models from scratch or pay per-use fees. The letter says open weights increase competition across chips, cloud services and applications. It also says hosting models locally gives organisations control over their data and reduces dependence on a single vendor for updates and pricing.

On security, the document acknowledges that released weights can be modified or redistributed without the original developer’s control. The letter argues against broad prohibitions, saying defenders and researchers need access to comparable models to detect threats and simulate attacks. The signatories add that concentrating advanced models behind a small number of providers creates single points of failure that external teams cannot inspect.

The letter defends distillation, a technique that uses one model’s outputs to train another, as a standard method in research and product development. The authors ask that alleged unlawful extraction be addressed through targeted legal and commercial remedies rather than broad limits on distillation practices.

The group did not propose specific legislation. The document calls on lawmakers to expand access to compute for startups and researchers, fund shared datasets and evaluation frameworks, and avoid premature restrictions on releasing weights or on distillation. The letter notes that wider distribution of weights would affect demand for hardware and services used to run models locally.