For a simple classification task you generally want to prioritize regularization over more sophisticated behavior, so fewer parameters with larger quantization makes sense. For more generic chat-like purposes, Q2 of a larger model may often be preferable to Q4 of a smaller one.