How to Run MiniMax M2.7 Locally: Hardware, GGUF Setup, and Performance Guide (2026)
MiniMax M2.7 is a 230B parameter MoE model that activates only 10B parameters per token. It scores 56.22% on SWE-Pro (rivaling Claude Opus 4.6) while running at 100 tok/s via API. Open weights dropped on HuggingFace in April 2026, meaning you can now self-host it.
The 4.3% activation ratio is what makes M2.7 interesting for local deployment. Despite 230B total parameters, inference cost scales with the 10B active portion. This puts it in reach of multi-GPU consumer setups.
This guide covers hardware requirements, quantization options, and step-by-step setup. If you want API access instead, see our MiniMax M2.7 complete guide.
Hardware requirements
Memory by quantization
| Quantization | Disk size | RAM/VRAM needed | Hardware example |
|---|---|---|---|
| BF16 (full) | ~427 GB | ~450 GB | 6x A100 80GB |
| Q8_0 | ~243 GB | ~260 GB | 4x RTX 4090 + 128GB RAM |
| Q6_K | ~188 GB | ~200 GB | 2x A100 80GB or Mac Studio 192GB |
| Q5_K_M | ~162 GB | ~175 GB | Mac Studio 192GB or 3x RTX 4090 |
| Q4_K_M | ~138 GB | ~150 GB | Mac Studio 192GB or 2x RTX 4090 + 64GB RAM |
| IQ3_XXS | ~95 GB | ~110 GB | M4 Max 128GB or 2x RTX 3090 |
| IQ2_XXS | ~75 GB | ~90 GB | M3/M4 Pro 96GB or 1x A100 80GB |
What can actually run this?
Apple Silicon (best consumer option):
- M3/M4 Max 96GB: IQ2_XXS or IQ3_XXS (~8-15 tok/s)
- M2/M3/M4 Ultra 192GB: Q4_K_M comfortably (~15-25 tok/s)
- Mac Studio M4 Ultra 256GB: Q6_K with room for context (~20-30 tok/s)
NVIDIA GPUs:
- 2x RTX 4090 (48GB VRAM) + 128GB system RAM: Q4_K_M with CPU offload (~15-20 tok/s)
- 4x RTX 4090 (96GB VRAM): Q6_K fully on GPU (~30-40 tok/s)
- 2x A100 80GB: Q5_K_M fully on GPU (~50-60 tok/s)
Not viable:
- Single RTX 4090 (24GB) — model does not fit even at lowest quant without heavy CPU offload
- Laptops with 32GB or less
- CPU-only systems under 128GB RAM (technically possible but <2 tok/s)
Setup with Ollama
The fastest path to running M2.7 locally:
# Pull the Unsloth GGUF (recommended - calibrated quantization)
# First download the specific quant you want
pip install huggingface_hub hf_transfer
HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download \
unsloth/MiniMax-M2.7-GGUF \
--include "*Q4_K_M*" \
--local-dir ./models/minimax-m2.7/
# Create Modelfile for Ollama
cat > Modelfile << 'EOF'
FROM ./models/minimax-m2.7/MiniMax-M2.7-Q4_K_M.gguf
PARAMETER temperature 0.7
PARAMETER num_ctx 16384
PARAMETER top_p 0.9
EOF
# Import and run
ollama create minimax-m2.7 -f Modelfile
ollama run minimax-m2.7 "Write a Python async web scraper with rate limiting"Alternative: use bartowski’s pre-built quants
# bartowski's GGUF is also well-regarded
huggingface-cli download bartowski/MiniMaxAI_MiniMax-M2.7-GGUF \
--include "*Q4_K_M*" \
--local-dir ./models/minimax-m2.7-bartowski/Setup with llama.cpp
For multi-GPU or when you need fine-grained layer control:
# Build with CUDA
git clone https://github.com/ggml-org/llama.cpp
cmake llama.cpp -B llama.cpp/build \
-DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON
cmake --build llama.cpp/build --config Release -j \
--clean-first --target llama-cli llama-server
# Run on 2x RTX 4090 with layer split
./llama.cpp/build/bin/llama-cli \
--model ./models/minimax-m2.7/MiniMax-M2.7-Q4_K_M.gguf \
--threads 16 \
--ctx-size 16384 \
--n-gpu-layers 62 \
--tensor-split 50,50 \
--temp 0.7 \
--top-p 0.9 \
-no-cnv \
--prompt "Your prompt here"MoE CPU offload trick (single GPU)
If you only have one GPU, offload the expert layers to CPU:
./llama.cpp/build/bin/llama-cli \
--model ./models/minimax-m2.7/MiniMax-M2.7-IQ3_XXS.gguf \
--threads 32 \
--ctx-size 8192 \
--n-gpu-layers 99 \
-ot ".ffn_.*_exps.=CPU" \
--temp 0.7 \
-no-cnv \
--prompt "Your prompt here"This keeps attention layers on GPU (fast) while routing expert computation through CPU (slower but fits). Expect ~5-10 tok/s on a single 24GB GPU with IQ3_XXS.
Serving as a local API
To use M2.7 with coding tools that expect an OpenAI-compatible API:
# Start llama.cpp server
./llama.cpp/build/bin/llama-server \
--model ./models/minimax-m2.7/MiniMax-M2.7-Q4_K_M.gguf \
--ctx-size 16384 \
--n-gpu-layers 62 \
--host 0.0.0.0 \
--port 8080
# Now use with any OpenAI-compatible client
curl http://localhost:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "minimax-m2.7",
"messages": [{"role": "user", "content": "Write tests for this function"}],
"temperature": 0.7
}'This works with Aider, Continue.dev, and any tool that supports custom OpenAI endpoints.
Performance benchmarks
Community-reported numbers:
| Setup | Quantization | Generation speed | Context |
|---|---|---|---|
| M3 Max 96GB | IQ3_XXS | ~10-14 tok/s | 8K |
| M4 Ultra 192GB | Q4_K_M | ~18-25 tok/s | 16K |
| 2x RTX 4090 | Q4_K_M | ~20-28 tok/s | 16K |
| 4x RTX 4090 | Q6_K | ~35-45 tok/s | 32K |
| 2x A100 80GB | Q5_K_M | ~50-65 tok/s | 32K |
For comparison, the MiniMax API runs at ~100 tok/s. Self-hosting is 2-5x slower but free after hardware costs.
Quality at different quantizations
MoE models handle quantization differently than dense models. Because only 10B parameters are active per token, quantizing the inactive experts has less impact on output quality:
| Quantization | Quality retention | Practical impact |
|---|---|---|
| Q6_K | ~98% | Negligible quality loss |
| Q5_K_M | ~96% | Very slight loss on complex reasoning |
| Q4_K_M | ~93% | Good for most coding tasks |
| IQ3_XXS | ~88% | Noticeable on nuanced tasks, fine for code generation |
| IQ2_XXS | ~82% | Useful for testing, not recommended for production |
For coding tasks, Q4_K_M is the sweet spot. The 7% quality loss vs full precision is rarely noticeable in practice for code generation, review, and debugging.
M2.7 vs API: cost comparison
| Usage level | Self-host cost (2x 4090) | API cost ($0.30/$1.20 per 1M) |
|---|---|---|
| 1M tokens/day | $0 (after hardware) | ~$45/month |
| 10M tokens/day | $0 (after hardware) | ~$450/month |
| 100M tokens/day | $0 (after hardware) | ~$4,500/month |
| Hardware cost | ~$4,000 upfront | $0 |
Break-even: At 10M tokens/day, self-hosting pays for itself in ~9 months. Below that, the API is more practical unless you need data privacy.
When to use M2.7 locally vs alternatives
| Need | Best option |
|---|---|
| Best coding quality (local) | DeepSeek V4 Pro (80.6% SWE-bench) |
| Best quality/size ratio | MiniMax M2.7 (56% SWE-Pro at 10B active) |
| Fits on consumer GPU | Qwen 3.6-27B (16GB RAM) |
| Cheapest API | DeepSeek V4 Flash ($0.28/M) |
| Agentic workflows | MiniMax M2.7 (designed for agents, 256 experts) |
| 1M context locally | GLM-5.2 (1M context, MIT) |
M2.7 excels at agentic tasks because of its 256-expert architecture. If you are building multi-step agent workflows that need to run locally (tool use, code execution loops), M2.7 is purpose-built for this.
FAQ
Can I run MiniMax M2.7 on a Mac?
Yes. M3/M4 Max with 96GB unified memory runs IQ3_XXS at ~10-14 tok/s. M2/M3/M4 Ultra with 192GB runs Q4_K_M comfortably at ~18-25 tok/s. Apple Silicon’s unified memory architecture is ideal for large MoE models.
What is the minimum hardware for M2.7?
The absolute minimum is 96GB unified memory (Apple) or 2x RTX 3090 (48GB VRAM) + 64GB system RAM, using IQ3_XXS quantization. Expect ~8-12 tok/s. Below this, the model either does not fit or runs too slowly to be useful.
Is M2.7 better than DeepSeek V4 for local use?
Different strengths. DeepSeek V4 Pro is better at coding (80.6% vs 56.2% SWE-bench). M2.7 is better at agentic tasks and has a much smaller active parameter count (10B vs 49B), meaning it needs less compute per token despite larger total size. For agent workflows, pick M2.7. For raw coding, pick DeepSeek.
Does MiniMax M2.7 support tool calling locally?
Yes. M2.7 was trained with native tool calling support. When running via llama.cpp server or Ollama, you can use function calling through the OpenAI-compatible API endpoint. Format tool definitions in your system prompt using the MiniMax chat template.
Should I wait for MiniMax M3?
MiniMax M3 is already available via API (59% SWE-bench Pro, 1M context). However, M3 open weights have not been released yet. If you need local deployment today, M2.7 is your option. When M3 weights drop, it will be the better choice.