Soumith Chintala
@soumithchintala
@TaikoToeda We will release NVFP4 and MXFP8 checkpoint. The MXFP8 checkpoint has a very marginal 1-2 percentage points gain over the NVFP4 checkpoint over evals (even though model native training is NVFP4), but we don't think its meaningful for any practical use-cases.