Hacker News new | ask | show | jobs
by gpugreg 1 day ago
The model is already natively MXFP4-quantized during training, so there is no quality loss.