What are open weights

A model's weights are the file that actually is the model — billions of numbers produced by training. Open weights means that file is published: you can download it, run it on your own hardware, and nobody meters your usage or sees your data. The alternative is a model that exists only behind its maker's API, where every use is a request to someone else's computer.

This is the single biggest fork in every decision on this site, which is why every model here wears one of three labels:

OPEN — downloadable under a permissive licence. Run it, modify it, build a product on it.

OPEN* — downloadable, but the licence restricts what you may do: some ban commercial use outright, others add conditions above a revenue threshold. The licence page linked from each model spells out what we know; where terms are unparsed we say "review terms" rather than guessing.

HOSTED ONLY — no downloadable copy. Where that classification comes from evidence, the page says so plainly; where it is only that we found no repository, the page says that instead — the gap is ours, not proof the lab published nothing.

One distinction worth keeping: open weights is not the same as open source. The weights being downloadable does not mean the training data or recipe are public, and most "open" models publish neither. For choosing what to run, the licence on the weights is what matters — and note that a downloaded model never phones home, which makes open weights the strongest privacy position available: the prompt that is never sent cannot be logged.

Where next: Licences, in plain language · When to consider running models locally