dense small models do not like quantization. i find 27b fp8 to be smarter albeit less knowledgable versus the 122B