Y
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.