Open-weight model

An open-weight model is an AI model whose trained parameters — its weights — are made publicly available for download, so that anyone can run the model on their own infrastructure. This is distinct from a closed, hosted model, which is only reachable through the provider’s API and whose weights are never released. Access to the weights is what makes local, on-premises and air-gapped deployment possible.

It is worth distinguishing open-weight from fully open-source. An open-weight release provides the parameters needed to run the model, but does not necessarily include the training data, the training code or full documentation of how the model was built, and it may come with a licence that restricts certain uses. "Open weights" therefore describes what you can run, not always how the model was made or what you are permitted to do with it — the licence terms need reading in each case.

For organisations with strict data-handling requirements, open-weight models are significant because they break the dependence on sending data to a third-party service. The model can be deployed inside the organisation’s own security perimeter, kept in a private environment, or fully air-gapped, so the capability is obtained without the data ever leaving the boundary. That shifts operational responsibility onto the organisation, but it is what enables AI adoption where a hosted service is not acceptable.