Deploy the Open‑Source MiniMax‑M2.7 Model Locally: Step‑by‑Step Guide

Model Highlights

Self‑evolution : after >100 autonomous optimization rounds, performance improves ~30%.

Software engineering : SWE‑Pro score 56.22%, comparable to GPT‑5.3‑Codex; production‑grade incident recovery <3 minutes.

Office abilities : GDPval‑AA ELO 1495 (highest among open‑source models); high‑fidelity multi‑turn editing of Word/Excel/PPT.

Native Agent Teams : supports multi‑agent collaboration with stable roles and autonomous decision‑making.

Code ability claimed comparable to GPT‑5.3‑Codex.

Deployment Ecosystem

Base GPU memory requirement ~230 GB. Two H200 GPUs are marginal; official recommendation is at least four H200 GPUs.

Quantized versions are under urgent development; only folder structure prepared, compression to a few dozen GB expected based on prior unSloth experience.

Ollama

Ollama 0.19 includes minimax-m2.7:cloud for free cloud inference (required because the 230‑billion‑parameter MoE model exceeds local GPU memory).

# Use with OpenClaw
ollama launch openclaw --model minimax-m2.7:cloud

# Direct chat
ollama run minimax-m2.7:cloud

vLLM

vLLM v0.19.0 adds HuggingFace v5 support, multimodal optimizations, and CPU KV‑cache offloading. Provides Day‑0 support.

# Basic deployment (4 × H200/H100/A100)
vllm serve MiniMaxAI/MiniMax-M2.7 \
  --tensor-parallel-size 4 \
  --tool-call-parser minimax_m2 \
  --reasoning-parser minimax_m2 \
  --compilation-config '{"mode":3,"pass_config":{"fuse_minimax_qk_norm":true}}' \
  --enable-auto-tool-choice \
  --trust-remote-code

# 8‑GPU deployment (DP+EP mode)
vllm serve MiniMaxAI/MiniMax-M2.7 \
  --data-parallel-size 8 \
  --enable-expert-parallel \
  --tool-call-parser minimax_m2 \
  --reasoning-parser minimax_m2 \
  --enable-auto-tool-choice

Docker one‑click start :

docker run --gpus all \
  -p 8000:8000 \
  --ipc=host \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  vllm/vllm-openai:minimax27 MiniMaxAI/MiniMax-M2.7 \
  --tensor-parallel-size 4 \
  --tool-call-parser minimax_m2 \
  --reasoning-parser minimax_m2 \
  --enable-auto-tool-choice \
  --trust-remote-code

Supported platforms:

NVIDIA : 4 × H200/H100/A100 tensor parallel, or 8‑GPU DP+EP/TP+EP modes.

AMD : 2 × or 4 × MI300X/MI325X/MI350X/MI355X with AITER acceleration.

System requirements: ~220 GB GPU memory for model weights plus ~240 GB for 1 M context tokens.

SGLang

SGLang provides Day‑0 support.

sglang serve \
  --model-path MiniMaxAI/MiniMax-M2.7 \
  --tp 4 \
  --tool-call-parser minimax-m2 \
  --reasoning-parser minimax-append-think \
  --trust-remote-code \
  --mem-fraction-static 0.85

The minimax-append-think parser separates reasoning from final output, enabling a “Thinking” mode.

Quick test:

curl http://localhost:30000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMaxAI/MiniMax-M2.7",
    "messages": [
      {"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]},
      {"role": "user", "content": [{"type": "text", "text": "Who won the world series in 2020?"}]}
    ]
  }'

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

NVIDIA Support

Free trial endpoint: https://build.nvidia.com/minimaxai/minimax-m2.7

Inference optimizations contributed by NVIDIA and the open‑source community:

QK RMS Norm Kernel : merges compute and communication into a single kernel, reducing launch overhead and memory traffic.

FP8 MoE : integrates TensorRT‑LLM’s FP8 MoE modular kernel, tuned for MoE models.

Throughput gains on NVIDIA Blackwell Ultra GPUs:

vLLM throughput increased 2.5× (within one month).

SGLang throughput increased 2.7× (within one month).

Additional Ecosystem Components

NemoClaw : NVIDIA open‑source reference stack for continuous operation of OpenClaw.

Fine‑tuning : NeMo AutoModel enables post‑training with EP + PP schemes; NeMo RL provides GRPO reinforcement‑learning recipes for 8K and 16K sequence lengths.

# NeMo AutoModel fine‑tuning recipe
https://github.com/NVIDIA-NeMo/Automodel/blob/main/examples/llm_finetune/minimax_m2/minimax_m2.7_hellaswag_pp.yaml

# Distributed training documentation
https://github.com/NVIDIA-NeMo/Automodel/discussions/1786

Transformers : Model can be loaded via HuggingFace Transformers (see deployment guide in the repository).

ModelScope : Weights available at https://modelscope.cn/models/MiniMax/MiniMax-M2.7

Tool Calling and Thinking Mode

M2.7 supports both tool calling and a Thinking mode.

Tool calling example (Python, OpenAI client) :

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:30000/v1",
    api_key="EMPTY"
)

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Get the current weather for a location",
        "parameters": {
            "type": "object",
            "properties": {
                "location": {"type": "string", "description": "The city name"}
            },
            "required": ["location"]
        }
    }
}]

response = client.chat.completions.create(
    model="MiniMaxAI/MiniMax-M2.7",
    messages=[{"role": "user", "content": "What's the weather in Beijing?"}],
    tools=tools
)

message = response.choices[0].message
if message.tool_calls:
    for tool_call in message.tool_calls:
        print(f"Tool Call: {tool_call.function.name}")
        print(f"  Arguments: {tool_call.function.arguments}")

Thinking mode : Enclose reasoning with <think>...</think> tags so streaming outputs can separate thought from final answer.