Open weights
A model released so that anyone can download its trained parameters, the numbers that come out of training, and run it themselves. In practice that means you can host it on your own infrastructure, keep patient data inside your own network, fine tune it on local data, and test it against your own population without sending anything to a vendor API. That control is usually the real reason a health system wants open weights. What you do not get is the recipe: the training data, the training code and the decisions that shaped the model stay with the developer, so you can measure how the model behaves but not inspect why it behaves that way. Open weights also says nothing about the licence. Most releases, Llama and Gemma included, come with vendor terms and use restrictions, and some are research use only, which is a procurement question before it is a technical one. Open weights is not open source.
An informatics team downloads an open weights language model, runs it on servers inside the hospital network and tests it on a few hundred de-identified discharge summaries from its own wards. Because the weights sit on the hospital's own hardware, the version they validated stays the version clinicians use until the team decides to change it, with no silent update from a provider. Before go-live, have procurement read the license for field-of-use limits, because a release restricted to research use cannot back a clinical service.
Terms like this come up in real clinical scenarios across the HelloAI courses: bite-sized modules with verifiable certificates. An account takes one minute, no password needed.
Sign in →