NVFP4 will be better if the model provider actually post-trained properly after quantizing.