An open-weights model is a model whose trained parameters are publicly available for download, so it can be run on your own infrastructure, fine-tuned on your own data, quantized and examined from the inside. The term is deliberately distinguished from “open source”: with software, open source means the availability of the source code, but here the real “source” is the training corpus and the training procedure, and those tend to be kept secret. What is published is only the resulting weights, which is closer to shipping a compiled binary. Licences matter too – some are genuinely permissive, others restrict commercial use, the number of users of a service, or the use of outputs to train competing models, and therefore fail the standard definition of open software. The practical advantages are substantial: data never leaves the organization’s infrastructure, which resolves a good part of compliance requirements; there is no risk of the vendor suddenly retiring or changing the model version; and at high call volumes running it can be considerably cheaper. The price is hardware cost, operational expertise and usually a gap behind the best closed models.
It is the difference between having a car in your garage and calling a taxi. The taxi is convenient, you look after nothing and you always ride in the latest model – but the driver knows where you went, the fares can change, and if the company folds you have no transport. Your own car costs you a garage, servicing and time, yet you drive it where you like and when you like, you answer to nobody and nobody can take it away. And mind one thing: getting the keys to a car is not the same as getting the manufacturing documentation. You can download the weights, but the recipe for the data they came from usually stays behind.