Open model (open-weight / open-source)

An open model is one whose learned weights — the parameter values that determine how it behaves — have been released publicly, most often under a license that permits downloading, running and further fine-tuning. The Llama family is one example. This is the opposite of closed models, which a vendor makes available only through an API, with no access to the weights themselves.

The difference from full open-source matters here — it is worth separating two degrees of openness. The term open-weight means that the weights of the finished model are publicly available, but not necessarily the training data or the code used to train it. Full open-source goes further and also covers the dataset, training code and documentation, so that the process of building the model can be reproduced. In practice, most popular "open" models today are open-weight rather than fully open-source.

For a company, openness has concrete deployment implications. The model can be run locally or on-premise, which improves data privacy and independence from a single vendor; in the case of smaller variants it leads toward a small language model running on your own hardware. Large open models often play the role of a foundation model — a base that a team tunes to its own tasks — in contrast to a closed large language model, which is available only as a service.