Home/Format
Format reference
The MXFP4 Format
MXFP4 is not one thing. It is a 4-bit element format, a shared 8-bit scale, a block size, and a set of conventions around packing that the specification deliberately leaves open.
The microscaling family
The microscaling family defines several element types — MXFP8 (E4M3 and E5M2), MXFP6 (E2M3 and E3M2), MXFP4 (E2M1) and MXINT8 — all sharing the same block structure: k elements, one E8M0 scale. MXFP4 is the most aggressive of them and the one where format details matter most, because there is almost no mantissa left to absorb a bad decision.
The four pages below cover the encoding itself, then the two things that decide whether it works in practice: how the scale is chosen, and where the resulting error lands. The two ecosystem pages cover what runs it.
Pages in this section
- Bit Layoutspec
- /format/bit-layout/The E2M1 element encoding value by value, the E8M0 scale, and how a block is packed to bytes.
- Block Scalingspec
- /format/block-scaling/Choosing the shared exponent, the amax rule, saturation and the outlier problem.
- Quantization Erroranalysis
- /format/quantization-error/Relative error bounds, what block size changes, and how error accumulates through a network.
- Hardware Supportecosystem
- /format/hardware-support/Native tensor-core paths, emulated paths, and how to tell which one you are on.
- Kernel Supportecosystem
- /format/kernel-support/The shape of a real MXFP4 matmul and the conventions checkpoints disagree about.