Recent work highlighted how quantization for recent LLaMa 3 models can lead to non-negligible decay in model performance. Does reduced precision hurt model performance?