Open-weights model
Also known as: open-weight model
What is Open-weights model?
A model whose trained weights are published, so you can download it and run it on your own hardware. Many of the best-known open models work this way. It is not the same as fully open source: you get the finished model, rarely the training data or the code to rebuild it. The appeal is control, privacy and no per-call fee.
Why it matters
The appeal of running your own model is easy to state and easy to underestimate. Skipping the per-call fee and keeping data in-house sounds like pure saving, until you count what the vendor used to do for you. Now your team owns the servers, the GPU capacity, the updates and the security patching, and there is no provider steadily improving the model or filtering abuse on your behalf. The licences vary too; “open” can still forbid certain uses or commercial scale, so the terms need reading, not assuming. For an organisation with sensitive data, strict residency rules or very high volume, the control is worth the operational load. For a light workload, a hosted model is usually cheaper once the running costs are counted honestly.
In practice
A healthcare provider rules out sending patient records to an external API and runs an open-weights model on its own servers instead. That solves the privacy problem and creates a staffing one: someone now has to keep the hardware, the model version and the security current. They budget for that role from the start, treating the model as infrastructure they operate rather than a feature they switched on.
