Modelpublic

Serpen/Minimax-M3-MSA-GGUF

sourceHugging Faceupdated 2mo agoView on Hugging Face

1likes104downloads

Model Card

Model

This repo contains specialized MoE-quants for MiniMax-M3 with the Indexer Tensors preserved at FP32. A BF16 MMPROJ file for image vision input has also been provided.

The text model files should load on mainline as this PR got merged. The MMPROJ requires this one, which is a superset of the first PR.

Quant	Size	Mixture	PPL	1-(Mean PPL(Q)/PPL(base))	KLD
Q4KM	246.11 GiB (4.96 BPW)	Q80 / Q4K / Q4K / Q5K	5.287961 ± 0.035451	+1.9814%	0.069890 ± 0.000978

[image] [image]