Two dozen companies have signed an open letter to US policymakers, urging them to protect open-weight AI models from additional restrictions.
Who Signed the Letter
The signatories include both competitors and collaborators. Meta, Microsoft, Nvidia, and IBM signed the letter, along with Dell Technologies, CrowdStrike, Palantir, ServiceNow, Hugging Face, Perplexity, Mistral, Andreessen Horowitz, Y Combinator, the Linux Foundation, and Mozilla.
The letter draws a comparison to the 1980s open-source software movement. It argues that AI now faces the same fight: should model weights stay open, or be locked away behind paid APIs?
What “Open-Weight” Actually Means
Open-weight models publish their trained parameters for anyone to download. People can inspect them, modify them, and run them on their own hardware. Closed models work differently. OpenAI and Anthropic, for example, keep their weights locked inside their own servers. Users only reach these models through an API.
The signatories say open weights let AI capability spread beyond a few big labs. They point to factories, hospitals, farms, classrooms, and small businesses as the real beneficiaries.
Three Core Arguments
The letter makes its case on three fronts:
- Lower costs. Open weights let startups and public institutions skip the huge expense of training their own frontier models.
- More competition. Open access pushes competition across chips, cloud services, and applications. This keeps prices in check, the letter argues.
- Less vendor lock-in. Companies that run open-weight models keep control of their own data. They don’t need to rely on a single vendor’s pricing or roadmap.
A Surprising Take on Security
The letter tackles the security question head-on, and it flips the usual argument on its head.
Once a company releases model weights, it loses control over them. People can modify these models, strip out safety features, and share them further. Nobody can trace or recall these changes.
The signatories don’t propose a ban, though. They compare the situation to cybersecurity. Defenders need models just as capable as the ones attackers use. Closed systems don’t easily offer that kind of access.
The letter pushes this argument further. Closed models aren’t automatically safer, it claims. Hackers can still breach them, people can still misuse them, and outside researchers can’t always verify how they behave. When a handful of closed providers control the technology, this creates single points of failure, not fewer risks.
Open models work differently, the letter argues. Outside teams can test them, run red-team exercises, and catch problems that one company’s internal testing might miss. This echoes an old software security debate: many eyes catch more bugs than one gatekeeper does. The letter doesn’t back this claim with hard data on AI vulnerabilities specifically, though.
A Defense of Distillation
The letter also defends a specific and controversial technique: distillation. This process trains one model using another model’s outputs. Researchers use it constantly, for testing, validation, and transferring skills between models of different sizes.
The signatories separate legitimate distillation from what they call unlawful attempts to steal value from closed models. They argue that new rules shouldn’t punish the whole technique just to stop a few bad actors.
This section responds directly to recent disputes. Chinese models like DeepSeek and Kimi sparked accusations from US labs, who claimed rivals had distilled their outputs without permission. The letter’s answer: use legal and commercial tools to fight misappropriation. Don’t restrict a technique the whole field relies on.
What Comes Next
The letter doesn’t attach itself to any specific bill or regulation. Instead, it stakes out a position before Washington moves on AI policy. It asks lawmakers to expand compute access for startups, fund shared datasets and evaluation tools, and avoid rushing into restrictions on open models.
Companies like Nvidia, IBM, and Dell have clear business reasons to want open models to thrive. A wider ecosystem of deployable models sells more chips and cloud services, no matter which lab builds the underlying technology.
For companies weighing open versus closed AI deployments, the policy picture still isn’t settled. Any future rules on distillation or open releases could reshape the economics of running AI in-house, and they could do it fast.
