I've come across a couple of people on the F-Droid and Trisquel forums, who argue that it's fine to ship a generative model as a component of Free Code software, even a proprietary one it's marked as 'nonfree media'.
I'm pretty sceptical of the claim that a generative model is not software. But if that's true it definitely adds MaSS (Model as Software Substitute) to a project.
@strypey I don't even see how any so-called "AI" models would be software.
If I download a text to speech model like magpie-tts-multilingual-357m-q8_0.gguf then I can't just chmod +x and ./run the thing. I will have to acquire something like audiocpp or koboldai and use those to open or ./run model.gguf the same way I would have to use gimp or krita to edit image.jpeg.
I see models as assets the same way SuperTuxKart is distributed in a git repository with source code and a cvs repository with assets (image files, etc).
It does get a bit murky the moment you ask questions like "What license is the model distributed under?", "Is it possible to reproduce this binary blob using free software tools?", "Is the training set available?", "Is the training data freely available or did they just use/steal a bunch of copyrighted images?" and so on. Chinese models tend to be a huge binary blob under a free license, but that's all you get, and a lot of models have some rather limited license (you can run it on your own computer but you can't use the output for commercial purposes etc).
F-Droid and Trisquel already have no-no policies for non-free assets and I don't remember seeing any special exceptions for non-free vector databases.
One obvious question regarding this whole drama light is of course: Why is bundling models even on the table? It's not the norm. koboldcpp doesn't come with any models, you'll have to get them on your own the same way you have to get the video files you'd like to play in a video player.