To add: https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RC... IQ3_XXS is within a point of the fully unquantized model and IQ3_S actually beats the unquantized model on many tasks! You lose absolutely nothing. It is quantization magic :)