Deploy MiniMax-M3

MiniMax-M3 is a native multimodal MoE model that accepts text, images, and video and provides a 1M-token context window. It has approximately 428B total parameters, activates approximately 23B parameters per token, and uses MiniMax Sparse Attention (MSA) to reduce long-context inference cost.

The complete weights must still be loaded or sharded. “23B activated parameters” does not mean that the deployment only needs memory for 23B parameters.

Open release and license

Resource	Description
BF16 model	`MiniMaxAI/MiniMax-M3`
MXFP8 model	`MiniMaxAI/MiniMax-M3-MXFP8`
MSA kernel	MiniMax-AI/MSA
License	MiniMax Community License

The MiniMax Community License includes product attribution, commercial notice or authorization, and prohibited-use terms. Read the full license and complete the applicable process before production or commercial use.

Choose hardware and weights

The SGLang Cookbook currently selects different weights and execution backends by hardware:

Hardware	Weights	Reference topology	Notes
NVIDIA B200	MXFP8	8 GPUs, TP 8	Reference configuration on this page; uses FA4, DeepGEMM, and the MSA fast path
NVIDIA B300 / GB200 / GB300	MXFP8	4 GPUs, TP 4	Uses the corresponding CUDA 13 image
NVIDIA H200	BF16	8 GPUs, TP 8	Hopper does not run the MXFP8 kernel and uses the Triton path
AMD Instinct	MXFP8	See the configurator	Different CDNA generations use different ROCm images and weight handling

The SGLang validation matrix and images evolve. For other hardware, generate a command with the MiniMax-M3 configurator.

Reference deployment: 8 × B200, MXFP8

The following command corresponds to the B200, MXFP8, Balanced, Single Node configuration in the SGLang Cookbook.

Start the server

docker run --gpus all \
  --shm-size 32g \
  -p 30000:30000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  --ipc=host \
  lmsysorg/sglang:dev-minimax-m3 \
  sglang serve \
  --trust-remote-code \
  --model-path MiniMaxAI/MiniMax-M3-MXFP8 \
  --reasoning-parser auto \
  --tool-call-parser auto \
  --tp 8 \
  --attention-backend fa4 \
  --moe-runner-backend deep_gemm \
  --chunked-prefill-size 8192 \
  --mem-fraction-static 0.65 \
  --host 0.0.0.0 \
  --port 30000

The current SGLang Cookbook uses a dedicated MiniMax-M3 development image. It includes the recommended MSA kernel for Blackwell. After support is available in a tagged SGLang release, prefer the stable image and command published by the Cookbook.

For a repository that requires authentication, pass HF_TOKEN into the container:

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

Verify text inference

curl http://localhost:30000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMaxAI/MiniMax-M3-MXFP8",
    "messages": [
      {"role": "user", "content": "Explain sparse attention in one sentence."}
    ],
    "max_tokens": 1024
  }'

With --reasoning-parser auto, the model trace is returned in message.reasoning_content and the final answer in message.content. With --tool-call-parser auto, native tool calls are converted to OpenAI-compatible tool_calls.

Verify image input

curl http://localhost:30000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMaxAI/MiniMax-M3-MXFP8",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "text", "text": "Describe this image."},
        {"type": "image_url", "image_url": {"url": "https://example.com/image.jpg"}}
      ]
    }],
    "max_tokens": 1024
  }'

Replace the example URL with an image URL accessible from the server. The SGLang Cookbook validates image URL and Base64 inputs. Check the current Cookbook notes to determine whether video input has been validated for your selected SGLang configuration.

Recommended inference parameters

Parameter	Recommended value
`temperature`	`1.0`
`top_p`	`0.95`
`top_k`	`40`, supplied per request

SGLang reads generation_config.json and automatically applies temperature=1.0 and top_p=0.95. top_k is not in that file and must be supplied as a request parameter when needed.

Production considerations

The MSA fast path for Blackwell requires an SM100-family GPU; other platforms use their corresponding SGLang backends.
Longer contexts consume more KV cache and prefill resources. A 1M-token context is the model limit, not a recommended default on every hardware configuration.
Low-concurrency latency and high-concurrency throughput require different tuning. Do not reuse a parallelism strategy from different hardware.
Add authentication, rate limiting, monitoring, content safeguards, and resource isolation before exposing the service.

Get started

Use Guides

FAQs

Terms & Policy

Open release and license

Choose hardware and weights

Reference deployment: 8 × B200, MXFP8

Start the server

Verify text inference

Verify image input

Recommended inference parameters

Production considerations

Resources

SGLang MiniMax-M3 Cookbook

MiniMax-M3 model repository

​Open release and license

​Choose hardware and weights

​Reference deployment: 8 × B200, MXFP8

​Start the server

​Verify text inference

​Verify image input

​Recommended inference parameters

​Production considerations

​Resources

SGLang MiniMax-M3 Cookbook

MiniMax-M3 model repository

Open release and license

Choose hardware and weights

Reference deployment: 8 × B200, MXFP8

Start the server

Verify text inference

Verify image input

Recommended inference parameters

Production considerations

Resources