vultr/MiniMax-M3-GGUF

MiniMax M3, GGUF

MiniMax M3 is a native multimodal Mixture-of-Experts model designed for long-context reasoning, coding, and agentic workflows across text, image, and video. Powered by MiniMax Sparse Attention (MSA), it is delivering highly efficient long-context processing and frontier coding performance.

This repository holds Vultr's GGUF quantizations of it. They are hybrid quantizations: rather than applying one precision to the whole network, each layer's precision is chosen from a measured importance score, so the bits go where they change the output most.

Model Overview

Attribute	Value
Base model	MiniMaxAI/MiniMax-M3
Architecture	minimax-m3, Mixture-of-Experts
Parameters	approximately 428B
Active parameters	23B
Layers	60
Hidden size	6,144
Attention heads	64
Experts	128 routed, 4 selected per token, plus a shared expert
Context length	up to 1M tokens
Vocabulary	200,064
MTP modules	7
License	MiniMax Community

Attribute	Value
Base model	MiniMaxAI/MiniMax-M3
Architecture	minimax-m3, Mixture-of-Experts
Parameters	approximately 428B
Active parameters	23B
Layers	60
Hidden size	6,144
Attention heads	64
Experts	128 routed, 4 selected per token, plus a shared expert
Context length	up to 1M tokens
Vocabulary	200,064
MTP modules	7
License	MiniMax Community

Quantizations

File	Type	Size	KLD vs bf16
MiniMax-M3-Q5_K_M-*.gguf (3 parts)	Q5_K_M	332.87 GB	0.0406
MiniMax-M3-Q4_K_M-*.gguf (2 parts)	Q4_K_M	283.57 GB	0.0589
MiniMax-M3-IQ3_M-*.gguf (2 parts)	IQ3_M	204.85 GB	0.2048
MiniMax-M3-IQ3_S-*.gguf (2 parts)	IQ3_S	202.95 GB	0.2166

File	Type	Size	KLD vs bf16
MiniMax-M3-Q5_K_M-*.gguf (3 parts)	Q5_K_M	332.87 GB	0.0406
MiniMax-M3-Q4_K_M-*.gguf (2 parts)	Q4_K_M	283.57 GB	0.0589
MiniMax-M3-IQ3_M-*.gguf (2 parts)	IQ3_M	204.85 GB	0.2048
MiniMax-M3-IQ3_S-*.gguf (2 parts)	IQ3_S	202.95 GB	0.2166

The vision encoder is in MiniMax-M3-mmproj-f16.gguf (1.73 GB). llama.cpp downloads it automatically with the command under Run, so image input works without extra steps.

KL divergence measures how far a quantized model's next-token probabilities drift from those of the original bf16 model, averaged over every token of the evaluation text. The lower the value, the closer the quantized model is to the original; zero would mean identical output.

Evaluation

Every variant in this repository was evaluated against the original bf16 model for perplexity, KL divergence, top-token agreement, HellaSwag and WinoGrande, under identical conditions. The full results are maintained on the model's page in the Vultr Model Library:

https://docs.vultr.com/models/minimax-m3/minimax-m3?tab=quantized

That page is the reference for these files. The summary above is an excerpt from it.

Run

llama-server -hf vultr/MiniMax-M3-GGUF:Q4_K_M

The model's chat template is embedded in every file, so no extra prompt formatting is needed.

License

These files are derived from MiniMaxAI/MiniMax-M3 and are distributed under the same license, MiniMax Community. The model is the work of MiniMax; Vultr produced the quantizations.