Why it matters

FP4 has hardware + quality advantages. Understanding shapes future quant.

Advertisement

The architecture

FP4 format: sign + exponent + mantissa.

NF4: quantile-based.

MXFP4: micro-scaled.

FP4 quant typesFP4 formatsNF4 / MXFP4 / E2M1Per-group scalingblock scalesHardware supportBlackwell tensor coresNF4 in QLoRA for training; MXFP4 for Blackwell inference
FP4.
Advertisement

How it works end to end

NF4: information-theoretic optimal for normal.

MXFP4: OCP standard.

Blackwell FP4 tensor cores 2x FP8.