Pick a decent quant (4-6KM) then use llama-fit-params and try it yourself to see if it's giving you what you need.