3 ms·The model is already natively MXFP4-quantized during training, so there is no quality loss.by gpugreg 2mo agoThe model is already natively MXFP4-quantized during training, so there is no quality loss.