Grafting vision onto text models for fun and profit.

So as we know.. llama.cpp separates the vision or other multimedia from the main weights. Conversely, trained model capabilities might be removed at release. What if there was a way to put them back? Mistral has now rel