dennny123/MiniMax-M2.7-GGUF

🔥 GGUF Quantizations of MiniMax-M2.7

Quantized using llama.cpp from BF16 source weights.

Original model: MiniMaxAI/MiniMax-M2.7

Run them in LM Studio or directly with llama.cpp.

Download a file from below

Filename	Quant type	File Size	Description
MiniMax-M2.7-BF16.gguf	BF16	~427 GB	Full BF16 weights. Use for re-quantizing or max quality.
MiniMax-M2.7-Q8_0.gguf	Q8_0	~243 GB	Extremely high quality, generally unneeded but max available.
MiniMax-M2.7-Q6_K.gguf	Q6_K	~188 GB	Very high quality, near perfect, recommended.
MiniMax-M2.7-Q5_K_M.gguf	Q5_K_M	~162 GB	High quality, recommended.
MiniMax-M2.7-Q4_K_M.gguf	Q4_K_M	~138 GB	Good quality, default size for most use cases, recommended.
MiniMax-M2.7-Q3_K_M.gguf	Q3_K_M	~109 GB	Lower quality but usable, good for tight hardware.
MiniMax-M2.7-Q2_K.gguf	Q2_K	~83 GB	Low quality, only for extreme memory constraints.

Downloading

pip install -U "huggingface_hub[cli]"

huggingface-cli download dennny123/MiniMax-M2.7-GGUF --include "MiniMax-M2.7-Q4_K_M*" --local-dir ./

For split files (>50GB):

huggingface-cli download dennny123/MiniMax-M2.7-GGUF --include "MiniMax-M2.7-Q8_0/*" --local-dir ./

Running the model

llama.cpp

./llama-cli -m MiniMax-M2.7-Q4_K_M.gguf -ngl 99 -cnv -p "You are a helpful assistant."

Ollama

ollama run hf.co/dennny123/MiniMax-M2.7-GGUF:Q4_K_M

LM Studio

Search for dennny123/MiniMax-M2.7-GGUF in the model browser.

Which file should I choose?

Have this much memory	Use this quant
256GB+	Q8_0 or Q6_K
192GB	Q5_K_M
144GB	Q4_K_M (most popular)
112GB	Q3_K_M
96GB	Q2_K

MiniMax-M2.7 is a Mixture-of-Experts model (229B total, ~10B active per token). All 229B parameters must be loaded into memory even though only a fraction are active per token. Size your hardware by total parameter count.

Quantization details

• llama.cpp: Latest main branch
• Conversion: BF16 GGUF intermediate, quantized in second pass
• Hardware: NVIDIA GH200 96GB + 525GB RAM

About MiniMax-M2.7

MiniMax-M2.7 is a 229B parameter MoE model (10B active) built for coding and agentic workflows.

• SWE-Pro: 56.22% (matches GPT-5.3-Codex)
• VIBE-Pro: 55.6%
• Terminal Bench 2: 57.0%
• GDPval-AA: ELO 1495 (highest open-source, surpasses GPT-5.3)
• MLE Bench Lite: 66.6% medal rate

Recommended inference parameters: temperature=1.0, top_p=0.95, top_k=40

See the official model card for full details.