What version are you running? Despite good benchmark figures, the coherence and ability to keep on task for large question and structured answers seems to drop

What version are you running? Despite good benchmark figures, the coherence and ability to keep on task for large question and structured answers seems to drop significantly using int4 and int8, also depending on the frontend you may get silent corruption from many of those that had been hastily thrown togheter (I.e. In one you get garbage out if you go over the token limit in the inputs and no warning at all)