AIBRIEF
Two dozen rivals and open source bodies argue that once open weight AI models (the trained numerical parameters that produce a model's outputs) are released they cannot be recalled—and that this irreversibility is not, on its own, a reason to lock
Two dozen companies and organisations, including direct rivals Meta, Microsoft, Nvidia, and IBM, signed one open letter to US policymakers asking Washington to extend 1980s open-source software protections to AI model weights.
Open-weight models are AI systems whose trained parameters, the numerical settings that produce the model's outputs, are published for anyone to download, inspect, modify, and run on their own hardware. That is distinct from closed frontier products such as OpenAI's or Anthropic's, which are sold only through APIs, meaning paid remote access rather than a downloadable copy.
The letter makes three positive arguments. First, lower entry cost for startups and public institutions that cannot afford to train frontier models or pay per-token API fees. Second, more competition across chips, cloud, and applications, which keeps costs down. Third, an escape from vendor lock-in for enterprises that need to control their own data and adapt the model internally.
In the security section, the letter concedes that model weights, once released, are beyond the developer's control: modified versions are hard to trace, safety guardrails can be stripped, and there is no recall mechanism. The signatories' stated answer is that this irreversibility is not a reason to lock weights down.
The coalition has to overcome a public counter-position. Major closed-model vendors and a meaningful share of safety researchers argue the same irreversibility is exactly why frontier weights should stay restricted.