An open-weight model ships its trained parameters, the numbers that encode what the model learned, as a downloadable file anyone can run on their own hardware. That is a narrower claim than “open source”: the training data, training code, and methodology usually stay private. What is public is enough to run and fine-tune the model without ever calling the original provider’s servers.
The distinction matters for governance because a closed, API-hosted model can be switched off, rate-limited, or have its usage policy enforced by the company running it. An open-weight model, once released, cannot be recalled. Roughly two million open-weight models already circulate publicly across repositories like Hugging Face, several months behind the capability frontier at any given time, runnable on consumer hardware with no guardrail attached unless the person running it adds one.
This is the practical limit on any single country’s AI safety legislation: a national ban can stop labs within its own jurisdiction from releasing a model, but it cannot un-release a model another country’s lab has already published as open weights, or compel a hobbyist running one locally to comply with rules written for hosted APIs.