How you get them
Whether the weights are published, and on what terms.
Open weights, permissive licence
Open-weights models with permissive licences publish their trained parameters for anyone to download, run, fine-tune and build into products, under terms such as Apache 2.0 or MIT that ask for little more than keeping the copyright notice. The training data and full training code are usually not released, which is why many people prefer the term open weights to open source
Open weights with conditions
Some downloadable models come with licences that add conditions. The weights are free to use, but the terms may require large companies to seek a separate deal, ban certain uses, require attribution or forbid using the outputs to train rival models
Fully open
Fully open models publish everything needed to study or rebuild them: the weights, the training data or a precise recipe for it, the training and evaluation code, and often intermediate checkpoints saved during training. They are rarer than open-weights releases because gathering lawful data and documenting it takes work
API and app only
API-only models are offered as a service. Developers send text, images or audio to the maker's servers through an application programming interface (API) and pay by the token; the public uses the same models in apps such as ChatGPT, Claude and Gemini
Research access only
Some models were never offered to the public at all. They were described in a paper, shown in a demonstration, or shared with selected researchers, testers or partners
Retired
Retired models are no longer offered by their makers. Hosted models are switched off on a published timetable when newer, cheaper versions arrive; chatbots are closed or renamed; a few were pulled within days after misbehaving