NVFP4 is a 4-bit floating-point number format introduced by NVIDIA for its newer GPU generations. With only four bits per value, it can represent very few distinct numbers, so it relies on scaling: small blocks of values share a scale factor, which lets the format adapt to the local range of the data. That keeps accuracy higher than a plain 4-bit format would.
Its purpose is efficiency. Storing a model in 4 bits takes about a quarter of the memory of 16-bit weights, and hardware designed for the format can perform the arithmetic faster and with less energy. That can reduce the cost of serving large models and allow bigger ones to fit on a given system.
The trade-off is precision. Quality can drop compared with higher-precision formats, particularly on demanding tasks, so it is applied with careful calibration, and some parts of a model may stay at higher precision. It also needs supported hardware and software.
Test the model in this format on your own workload, compare accuracy and speed against 8-bit and 16-bit versions, and confirm hardware support.