Meta, Microsoft, and 22 others tell Washington: don't lock up AI model weights

Two dozen companies and organizations signed an open letter today urging US policymakers to protect open-weight AI models. The coalition spans fierce rivals and unlikely allies: Meta, Microsoft, Nvidia, IBM, Dell, CrowdStrike, Palantir, ServiceNow, Hugging Face, Perplexity, Mistral, Andreessen Horowitz, Y Combinator, the Linux Foundation, and Mozilla all put their names on it. Their core argument frames open-weight models—where trained parameters are published for anyone to download and run—as the mechanism that spreads AI capability beyond a handful of well-capitalized labs.

The letter runs on three tracks. It says open weights lower the cost of entry for startups and public institutions that can’t afford per-token fees at frontier prices. It argues they increase competition across chips, cloud, and applications, preventing value capture by a few providers. And it positions open models as the enterprise escape hatch from vendor lock-in, since organizations running them control their own data.

The security section inverts the usual narrative. The signatories concede that once weights are released, modified versions with stripped safety guardrails can circulate freely with no recall mechanism. But they argue prohibition isn’t the answer. Their case borrows from cybersecurity: defenders facing AI-equipped attackers need access to models with comparable capability. Closed systems, they claim, create single points of failure that external researchers can’t observe or verify. “Open models let outside researchers examine behaviour, run red-team exercises, and identify vulnerabilities across many teams rather than relying on one vendor’s internal testing,” the letter argues, drawing a direct parallel to the “open-source is more secure than obscurity” debate that shaped decades of software security.

The letter also defends distillation—using one model’s outputs to train another—as a legitimate technique that shouldn’t get swept up in restrictions aimed at unauthorized extraction from closed models. This reads as a clear response to disputes that flared after the rise of Chinese models like DeepSeek, which several US labs accused of distilling from their closed systems without authorization. The letter’s position: address misappropriation through targeted legal mechanisms, not blanket restrictions on a technique the entire field depends on. It arrives without a specific legislative proposal attached, calling instead for expanded compute access, shared training datasets, and avoiding what it terms “premature restrictions.”