For example, in diffusion, there are some models where a Q8 quant dramatically changes what you can achieve compared to fp16. (I'm thinking of the Wan video models.) The point I'm trying to make is that it's a noticeable model change, and can be make-or-break.