> So if I take a photo of a pretty sunset, release it under MIT license, you'd say it's "not open source" unless I give you the sun and the atmosphere themselves?

Phorographs are not source code, are not computer code at all, and do not have a close analog to source code. Calling them “open source” is, at best, a poor and distant metaphor in any case, and so they aren't a useful model for discussing what open source means for software.

There's a very reasonable argument that model weights are an IL or object-code like artifact with training data and training source code together as the source code.

That doesn't change that the MIT license is an open source license, but when what you release under that license isn’t the whole source aoplicable to the model, but just inference and maybe training code but not the data needed to produce the weights, and final weights, then it is fair to question whether as a whole the model is open source.