Seems pretty useful, often hopping between different quants. I wonder if this would work for different "branches" of a model.
E.g. qwen 3.6 35b regular vs abliterated.
It probably needs to store the full F16 version for each model variant, even if the variants share the same base model