Nvidia compresses Nemotron 3.5 Super VL using 4-bit NVFP4 precision, allowing the model to run entirely in 4-bit precision on Blackwell GPUs