What are you talking about? The model is native NVFP4, why you run it at any precision higher than that?