That's why you don't run the model at low quantization