Open weights
The model file itself, downloadable and runnable on your own hardware.
A model's weights are the numbers produced by training — the model itself, as a file. When a lab releases weights, anyone can download them, run them on their own machines, and modify them. When a lab offers an API, the weights stay with the lab and you send your data to them.
Open weights is not the same as open source. Most released weights come with licences restricting use, and the training data and code are usually not published, so the thing cannot be reproduced or fully audited. The phrase open model covers a wide range of positions, and the licence is where the meaning is.
The gap between the best open-weight models and the best closed ones has narrowed considerably, which changes the calculation for anyone who cannot send data outside their own environment.
Why it matters here
This is the whole sovereignty and data-residency conversation in one concept. If the material cannot leave the jurisdiction or the network, open weights running on controlled infrastructure is the only route — and the cost moves from a per-use fee to hardware, hosting and the people who run it, which is a different budget line and a different capability requirement.
The question to ask
If this data cannot leave our environment, what does running it ourselves actually cost, staff included?
Go deeper
Reviewed 2026-09-20 · All decoders