Deploy MiniMax-M2.7

MiniMax-M2.7 is a language model for agents, software engineering, and complex productivity tasks. For a new project that needs native multimodal input or a 1M-token context, use MiniMax-M3. Continue with this page for an existing M2.7 workflow.

Open release and license

Resource	Description
Model repository	`MiniMaxAI/MiniMax-M2.7`
License	MiniMax-M2.7 License
SGLang Cookbook	MiniMax-M2.7

The MiniMax-M2.7 License permits the personal, research, and educational uses listed in the license. Commercial use includes attribution and prior-authorization requirements. Read the complete license before production or commercial use.

Hardware and images

The current SGLang Cookbook lists these reference combinations:

Platform	Reference hardware	SGLang image
NVIDIA	A100 / H100 / H200 / B200	`lmsysorg/sglang:v0.5.10.post1`
NVIDIA	B300 / GB300	`lmsysorg/sglang:v0.5.10.post1-cu130`
AMD	MI300X / MI325X	`lmsysorg/sglang:v0.5.10.post1-rocm720-mi30x`
AMD	MI355X	`lmsysorg/sglang:v0.5.10.post1-rocm720-mi35x`

Common NVIDIA reference topologies use four high-memory GPUs with TP 4 or eight GPUs with TP 8 and EP 8. Other hardware supports different two-, four-, or eight-GPU combinations. Use the SGLang configurator to obtain the matching command.

Reference deployment: four high-memory NVIDIA GPUs

Start the server

docker run --gpus all \
  --shm-size 32g \
  -p 30000:30000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  --ipc=host \
  lmsysorg/sglang:v0.5.10.post1 \
  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 \
  --host 0.0.0.0 \
  --port 30000

On the first start, SGLang downloads the model from Hugging Face. When authentication is required, add this argument to docker run:

--env "HF_TOKEN=<your-hf-token>"

Verify the deployment

curl http://localhost:30000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMaxAI/MiniMax-M2.7",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Write a Python function that tests whether a number is prime."}
    ],
    "max_tokens": 2048
  }'

You can also use the OpenAI SDK:

from openai import OpenAI

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

response = client.chat.completions.create(
    model="MiniMaxAI/MiniMax-M2.7",
    messages=[{"role": "user", "content": "Explain mixture-of-experts models."}],
    max_tokens=2048,
)

print(response.choices[0].message.content)

Key launch parameters

Parameter	Recommended value	Purpose
`--tool-call-parser`	`minimax-m2`	Converts native tool calls to an OpenAI-compatible format
`--reasoning-parser`	`minimax-append-think`	Processes model reasoning content
`--trust-remote-code`	Enabled	Loads the implementation from the model repository
`--mem-fraction-static`	`0.85`	Allocates the static GPU memory fraction for model execution and KV cache
`--tp`	`2`, `4`, or `8`	Sets tensor parallelism for the hardware
`--ep`	Set by the configurator	Expert parallelism is available for eight-GPU NVIDIA or AMD deployments

Parallelism, KV cache capacity, and available context length are interdependent. Do not change only the GPU count. Select a complete configuration from the SGLang configurator whenever the hardware changes.

Reasoning and tool calls

MiniMax-M2.7 emits reasoning content. With minimax-append-think, read or parse the trace according to the response format of your SGLang version. With minimax-m2, functions can be supplied through the OpenAI-compatible tools parameter. Before production integration, validate:

Non-streaming and streaming text generation
Your policy for displaying or hiding reasoning content
Single, parallel, and nested tool arguments
GPU memory headroom at the required context length

Community quantizations and Mac deployment

Community-converted MLX and GGUF versions are available on Hugging Face, but they are maintained by their respective publishers and are outside the MiniMax + SGLang reference deployment on this page. When using a community build, separately verify the conversion method, quantization error, license, and runtime compatibility.

Resources

SGLang MiniMax-M2.7 Cookbook

Open the complete hardware combinations, configuration parameters, and benchmarks.

MiniMax-M2.7 model repository

Open the model card, weights, and license.

Get started

Use Guides

FAQs

Terms & Policy

Open release and license

Hardware and images

Reference deployment: four high-memory NVIDIA GPUs

Start the server

Verify the deployment

Key launch parameters

Reasoning and tool calls

Community quantizations and Mac deployment

Resources

SGLang MiniMax-M2.7 Cookbook

MiniMax-M2.7 model repository

​Open release and license

​Hardware and images

​Reference deployment: four high-memory NVIDIA GPUs

​Start the server

​Verify the deployment

​Key launch parameters

​Reasoning and tool calls

​Community quantizations and Mac deployment

​Resources

SGLang MiniMax-M2.7 Cookbook

MiniMax-M2.7 model repository

Open release and license

Hardware and images

Reference deployment: four high-memory NVIDIA GPUs

Start the server

Verify the deployment

Key launch parameters

Reasoning and tool calls

Community quantizations and Mac deployment

Resources