NVFP4 stores each model value in 4 bits (one sign, two exponent and one mantissa bit) and groups values into blocks of 16 that share an 8-bit scale factor, plus one 32-bit scale per tensor. It was introduced with Nvidia’s Blackwell GPUs.
Nvidia says it cuts memory use by about 3.5 times compared with 16-bit weights. You meet it on model download pages, where releases such as Meta’s Muse Spark and FLUX.3 Image Klein offer an NVFP4 file for smaller hardware.