MiniMax-M2.7-GGUF

MiniMax-M2.7-GGUF

imatrix-calibrated GGUF quantizations of MiniMaxAI/MiniMax-M2.7 — the first model in the MiniMax lineup to participate in its own training evolution.

M2.7 matches or exceeds GPT-5.3-Codex on SWE-Pro (56.22%), rivals Opus 4.6 on VIBE-Pro (55.6%), and outperforms Claude Sonnet 4.6 on agentic benchmarks — with only 10B active parameters per token out of 229B total.

Why these quants

• imatrix calibrated — all sub-Q8 quants use importance matrix for optimal quality retention. No blind quantization.
• Q8_0 is effectively lossless — source weights are FP8 (float8_e4m3fn). Q8_0 preserves full precision.
• XL variants — embed and output weights kept at Q8_0 for meaningfully better coherence on MoE routing. Available for Q4_K_M, Q5_K_M, Q3_K_M.
• Correct license metadata — NON-COMMERCIAL license properly set in GGUF headers via --override-kv. Other repos incorrectly list modified-MIT.
• Validated — every quant smoke-tested for coherent output, reasoning accuracy, and code generation before upload.

License

NON-COMMERCIAL USE ONLY. Commercial use requires prior written authorization from MiniMax. See LICENSE.

Quick Start

# Recommended: Q4_K_M for most users
hf download Youssofal/MiniMax-M2.7-GGUF --include "MiniMax-M2.7-Q4_K_M/*" --local-dir ./

# Run with llama-server
llama-server -m MiniMax-M2.7-Q4_K_M.gguf \
  -ngl 999 -c 32768 --jinja --reasoning-format auto -fa \
  --temp 1.0 --top-p 0.95 --top-k 40

Which file should I use?

Sizing rule: Pick a quant 1-2 GB smaller than your total VRAM (GPU) or VRAM+RAM combined (Apple Silicon). A great deep-dive with charts is provided by Artefact2.

All Available Quantizations

Download

pip install -U "huggingface_hub[cli]"

# Most quants are split (>50 GB) -- download the folder
hf download Youssofal/MiniMax-M2.7-GGUF --include "MiniMax-M2.7-Q4_K_M/*" --local-dir ./

# Smaller quants (single file)
hf download Youssofal/MiniMax-M2.7-GGUF --include "MiniMax-M2.7-Q2_K*" --local-dir ./

Running

# llama-server (recommended)
llama-server -m MiniMax-M2.7-Q4_K_M.gguf \
  -ngl 999 -c 32768 --jinja --reasoning-format auto -fa \
  --temp 1.0 --top-p 0.95 --top-k 40

# llama-cli (conversation mode)
llama-cli -m MiniMax-M2.7-Q4_K_M.gguf \
  -ngl 999 --jinja --reasoning-format auto -cnv

Sampling parameters (official MiniMax defaults): temperature 1.0, top_p 0.95, top_k 40.

Apple Silicon / MoE Notes

Despite only 10B active parameters per token, all 229B weights must reside in memory. On Apple Silicon with unified memory, use --cpu-moe for improved expert dispatch.

• Q4_K_M (~138 GB) fits on 192 GB Mac
• IQ4_XS (~122 GB) fits on 128 GB with swap pressure
• Q3_K_M (~109 GB) runs on 128 GB Mac

Architecture

XL Variants

XL variants (Q4_K_M-XL, Q5_K_M-XL, Q3_K_M-XL) keep embedding and output weight tensors at Q8_0 instead of the default lower precision. This preserves vocabulary quality at minimal size cost and meaningfully improves coherence -- particularly important for MoE models where expert routing precision matters.

Chat Template

Two templates are provided:

• chat_template.official.jinja -- exact copy from MiniMax, includes <think> in generation prompt
• chat_template.llama-server-patched.jinja -- removes <think> from generation prompt to prevent duplication in some llama-server versions

Compatibility

Quantization Details

• Quantized with llama.cpp at commit ff5ef82
• imatrix generated from bartowski calibration data
• License metadata corrected via --override-kv (upstream model card incorrectly propagates modified-MIT)
• All quants validated: coherent English output, correct reasoning, valid code generation

Credits

• Original model by MiniMax -- MiniMaxAI/MiniMax-M2.7
• imatrix calibration data by bartowski
• Quantized by Youssofal