Open Models Almanac
Methodsoptimizationquantization-aware training

specific method · filed under optimization

MXFP4 Quantization

MXFP4 is the Open Compute Project's four-bit format: E2M1 elements with one power-of-two scale shared by each block of 32 values, 4.25 bits per value. Since the scale can only be a power of two, a block's largest values are either clamped or use only part of the range, and implementations differ on which.

Written for the explainer Quantization for serving: which bits to remove, and what each removal saves.

Also called MXFP4 quantization of MoE weights.

sources
3
model
1
lab adopt it
1
strongest
default

Documented in

Evidence

3 quotes from 3 sources, one card per model, each opening with its strongest.

How sources treat it
used 2default 1

gpt-oss-120b

OpenAIadopted3 sources

The models were post-trained with MXFP4 quantization of the MoE weights

defaultpost trainingin gpt-ossOpenAI
2 more from this model

MXFP4 quantization: The models were post-trained with MXFP4 quantization of the MoE weights, making gpt-oss-120b run on a single 80GB GPU

usedunclearin gpt-oss-120bOpenAI

We post-trained the models with quantization of the MoE weights to MXFP4 format

usedinference servingin gpt-oss-120b & gpt-oss-20bOpenAI
↑