NVFP4 format enables cheaper LLM inference without quality loss

NVFP4 is a 4-bit floating-point format that reduces GPU memory usage for LLMs with minimal quality loss. NVIDIA demonstrates creating an NVFP4-quantized Nemotron 3 Ultra checkpoint using Model Optimizer.