Foundations & Models
Model Weights
The weights are the model — a huge file full of numbers, the result of training. Whoever has them can run, modify and pass on the model. That is why “weights available or not” is the real line between open and closed, more important than any marketing label.
An example
The weights are the result of months of compute time and sums in the hundreds of millions — and at the same time nothing more than a large file of numbers. Whoever has them has the model: they can run it, modify it, keep training it and copy it, without the creator ever finding out. That is why the decision to release weights is irreversible. A file that has been shared once cannot be recalled.
A closer look
The weights are the result of training, technically a file with billions of numbers. Whoever has them can run the model without the provider, keep training it and examine it. Releasing them is therefore a strategic decision, not a technical one.
Why it matters
For procurement this is the most important contractual distinction there is. With API access you rent access: the provider can change prices, retire models, adjust terms of use and see your requests. With the weights in hand you own a capability that nobody can switch off — but you pay for operation, hardware and updates yourself. For regulated areas where data must not leave the building, this is often not a cost comparison but a precondition.
Commonly misunderstood
Open weights are not the same as open source code. Training data and training procedures almost always remain under lock and key. The result can be used, not reconstructed.