Model Weights
Model weights are the billions of numbers that training produced. They are the model: an architecture without weights is useless, and weights without it are just a file.
Open and closed weights
Some developers publish their weights: the model can be downloaded, run locally, fine-tuned and inspected. Others keep them in-house and give access only through an API. That choice decides who holds your data and what happens if the provider changes terms.
"Open weights" is not the same as open source. The training data and the training process are almost never published, and licences often restrict commercial use. Weights alone do not let anyone reproduce the model.
The size of the weights file sets the hardware requirement. Smaller models run on an ordinary laptop; large ones need server-grade GPUs, which is why most people use an API rather than their own machine.
What follows from it
- Your weights, your dataRunning locally means requests never leave your infrastructure. For sensitive data that is often the deciding argument.
- A closed model can changeA provider updates the model on their side and behaviour shifts without your involvement. Downloaded weights do not.
- Read the licenceBeing downloadable is not the same as being licensed for commercial use. Terms differ substantially between models.