Kimi K3
Moonshot AIMXFP4 weights (Microscaling FP4): Each weight is stored in 4-bit floating point with per-block scaling factors.
1 more from this model
using MXFP4 weights
specific method · filed under inference & serving
Weights are quantized to 4-bit floating point with microscaling block scales.
The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See inference quantization for the methods filed beside it.
Also called Microscaling FP4 (MXFP4) weights, MXFP4 weights, native MXFP4 quantization.
3 quotes from 3 sources, one card per model, each opening with its strongest.
MXFP4 weights (Microscaling FP4): Each weight is stored in 4-bit floating point with per-block scaling factors.
using MXFP4 weights
is optimized to run on a single H100 GPU with native MXFP4 quantization
Picked by hand, not extracted: where to read more, not evidence for anything on this page.