When a model’s internal parameters are published for anyone to download and run on their own hardware, rather than being available only through a company’s service. This matters for privacy, cost, control and independent research. It is often loosely called ‘open source’, though the two are not the same.
Technical view
Open weights means the trained parameters are released. Genuine open source in the traditional sense would also require the training data and code, which almost no releases provide, and many carry licences restricting commercial use. The trade-offs are contested: openness enables scrutiny, competition and reproducible research, but also removes the ability to withdraw or restrict a model once released.
Related
Part of AI Terminology Index, section: Extending and Applying Models.