jkyamog/MiniMax-M3-mmproj-GGUF

MiniMax M3 mmproj GGUF

This repository contains the MiniMax M3 vision projector GGUF for use with ik_llama.cpp MiniMax M3 vision support.

Derived from MiniMaxAI/MiniMax-M3; follow the upstream MiniMax-M3 license terms (minimax-community on the upstream model card).

Files

• mmproj-minimax-m3-vl-f16.gguf

Usage

Use this file as the --mmproj argument when serving a MiniMax M3 GGUF model with an ik_llama.cpp build that includes MiniMax M3 vision projector support.

Example:

./llama-server \
  -m /path/to/MiniMax-M3-UD-Q6_K-00001-of-00009.gguf \
  --mmproj /path/to/mmproj-minimax-m3-vl-f16.gguf

Notes

The projector was converted from MiniMaxAI/MiniMax-M3. MiniMax M3 vision uses Qwen2.5VL-compatible image preprocessing, two temporal patch embedding convolutions, 3-axis NEOX vision RoPE, and the MiniMax multimodal projector plus patch merge MLP.

The tested artifact was used with MiniMax M3 Q6 through the OpenAI-compatible server path and passed smoke tests on general object recognition and receipt OCR.