Meta, Microsoft, Nvidia and 20 Others Back Open AI Weights

Twenty-four companies and organisations, including Meta, Microsoft, Nvidia and IBM, signed a public letter this week asking U.S. policymakers to protect the distribution of trained AI model weights. The letter urges officials to avoid broad restrictions that would limit the free circulation of those parameters.

The signatories describe open-weight models as systems where the trained parameters are published for anyone to download, inspect, modify and run on local hardware. That contrasts with closed, API-only models where the underlying weights remain on a vendor’s servers and users access capabilities through calls to the provider.

The coalition includes cloud and chip suppliers, cybersecurity and enterprise software firms, AI startups and foundations. Named backers include Dell Technologies, CrowdStrike, Palantir, ServiceNow, Hugging Face, Perplexity, Mistral, Andreessen Horowitz, Y Combinator, the Linux Foundation and Mozilla.

The letter lists three main arguments for protecting open weights. First, published weights lower costs for startups, public institutions and smaller firms that cannot train large models from scratch or pay high per-call fees to closed services. Second, open weights are presented as promoting competition across the technology stack — from chips and servers to cloud services and applications — which the letter says prevents concentration of value among a few providers. Third, organisations that run open-weight models can avoid vendor lock-in, keep control of internal data and adapt models to meet operational or regulatory needs.

On security, the letter acknowledges that once weights are released they cannot be fully controlled and that modified versions may circulate with safety controls removed. The signatories argue against bans, stating that defenders — security teams, researchers and institutions — need access to comparable model capabilities to detect threats, evaluate risks and conduct red-team exercises. The document contends that concentrating advanced models behind closed APIs can create single points of failure and limit independent review, while published weights permit wider inspection of behaviour and vulnerabilities.

The letter defends model distillation — using one model’s outputs to train or evaluate another — as a common research and development technique. It calls for misappropriation claims to be handled through targeted legal and commercial measures rather than blanket bans that would affect routine research practices.

The document does not lay out specific legislative text. It asks lawmakers to increase compute access for startups and researchers, fund shared training datasets, and develop common evaluation frameworks while avoiding what it describes as premature restrictions on open models. The filing was submitted ahead of expected debate in Washington on AI governance and frames the signatories’ preferred policy approach to that discussion.