Right, my interpretation (and I think monocasa's too) is that the training and packaging pipeline is not published therefore the code used to build the model is not open, therefore there is no source available to rebuild the model as Mistral's engineers have done.

In essence the output of the training pipeline is "data" (weights), not code. The way those data are read to make things work is the "model". In the (bad) analogy of the JPEG, the data is the coefficients and the model is the encoding/decoding scheme.

I'm impressed by the patience monocasa has in the face of so many (seemingly intentionally) obtuse responses.

The JPEG analogy makes no sense. A more correct analogy is: they released compiled binaries, not the source code.

The difference is though, here compiling from source is prohibitively expensive for the average person. And the source is massive.

My analogy makes perfect sense.

Also, being able to reverse-engineer the precise build process is an orthogonal concern to having the source code available.