Large language models store their weights in four bits to fit on a laptop, but the fastest matrix hardware multiplies eight-bit integers, so the weights must be converted and prompt processing slows down.
MXFP4 is a four-bit format that gives each small block of weights one shared power-of-two scale; existing kernels apply that scale to every partial result. If a row's block scales fall within a narrow range, the whole row equals one exact integer vector with a single scale, so the layer runs as one integer matrix multiply.
Real weights already meet this: at least 97.9% of rows across seven models. Enforcing it barely perturbs rounding, files stay standard, and prompt processing runs up to 1.7 times faster.