What it is. NVFP4 is a 4-bit number format from Nvidia for storing and running AI models, introduced with its Blackwell generation of GPUs.
An analogy. Imagine writing down the height of everyone in a school. You could record each one to the millimetre, which takes a lot of digits. Or you could split the pupils into small groups, write one “ruler size” for each group, and then record each pupil as a single rough mark on that group’s ruler. You lose a little precision but save a lot of space. NVFP4 does the same with a model’s parameters.
How it works. Each value is stored in just 4 bits: one for the sign, two for the exponent and one for the mantissa (a layout written as E2M1). On its own, 4 bits can only represent a handful of values. So NVFP4 groups values into blocks of 16, and each block gets its own scale factor stored in an 8-bit format (FP8 E4M3). A second scale, in 32-bit precision, covers the whole tensor. Nvidia says this cuts a model’s memory footprint by about 3.5 times compared with 16-bit weights and about 1.8 times compared with 8-bit.
Why it was in the news this week. Meta’s Muse Spark open-weights release came in BF16 and NVFP4 versions, and Black Forest Labs published FLUX.3 Image Klein in bf16, FP8 and NVFP4 “for smaller cards”. The 4-bit versions are how people with less memory get to run these models at all.
One thing to remember. Smaller number formats are a form of quantisation: they trade a little accuracy for a lot of memory and speed. Whether the trade is worth it depends on the model and the task, so check any quality figures the publisher gives for the 4-bit version.