Deploy MiniMax M3 with vLLM OpenAI API

Run MiniMax M3 locally with vLLM and expose an OpenAI-compatible API.

This guide is for developers who want to serve MiniMax M3 with vLLM and keep the interface OpenAI-compatible. By the end, you will have a running model server, tool-calling and reasoning parsers enabled, and a quick way to verify that requests reach the endpoint.

You will also know which runtime pieces matter most: GPU access, Hugging Face cache mounting, tensor parallelism, and the exact flags used by the MiniMax M3 recipe in vLLM.

Before you start

• Docker installed, version 24+.
• NVIDIA GPU with CUDA-capable drivers installed.
• At least 1 GPU; 8 GPUs recommended for the sample tensor parallel setting.
• Hugging Face account and access to the MiniMaxAI/MiniMax-M3-MXFP8 model.
• Hugging Face token configured locally with huggingface-cli login or an equivalent secret mount.
• Linux host with --privileged and --ipc=host support for the container run command.
• Enough disk space for model weights and cache, ideally 100 GB+ free.

Step 1: Pull the vLLM OpenAI image

Your first outcome is a ready-to-run container image that already includes the OpenAI-compatible server entrypoint used by the MiniMax M3 recipe.

docker pull vllm/vllm-openai:minimax-m3

After the pull completes, you should see Docker report the image as downloaded locally. If you run docker images, you should see vllm/vllm-openai with the minimax-m3 tag.

Step 2: Mount the Hugging Face cache

Your next outcome is persistent model caching, so the weights do not download again every time you restart the server.

mkdir -p ~/.cache/huggingface

Then make sure your Hugging Face credentials are available to the runtime. A common path is to log in once on the host and mount the cache into the container, as shown in the final run command. You should be able to list files under ~/.cache/huggingface and see token and model cache directories after the first download.

Step 3: Start the MiniMax M3 server

Your main outcome is a live API server on port 8000 that loads MiniMax M3 with the recipe settings from the source guide.

docker run --gpus all --privileged --ipc=host -p 8000:8000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  vllm/vllm-openai:minimax-m3 MiniMaxAI/MiniMax-M3-MXFP8 \
  --block-size 128 \
  --tensor-parallel-size 8 \
  --tool-call-parser minimax_m3 \
  --enable-auto-tool-choice \
  --reasoning-parser minimax_m3

When the container starts correctly, you should see vLLM logs that mention model loading, tokenizer setup, and the OpenAI-compatible server binding to 0.0.0.0:8000. If the model is downloading, expect extra progress output before the server becomes ready.

Step 4: Verify the OpenAI-compatible endpoint

Your outcome here is proof that the server is reachable and responding to API calls, not just running in the background.

curl http://localhost:8000/v1/models

You should see a JSON response that lists the loaded model or available model entry. If that request returns model metadata, the server is healthy and the OpenAI-style route is working.

Step 5: Confirm tool calling and reasoning parsers

Your final outcome is a server configured for agentic workflows, with MiniMax M3-specific tool-call and reasoning parsing enabled.

curl http://localhost:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "MiniMaxAI/MiniMax-M3-MXFP8",
    "messages": [{"role": "user", "content": "List two tools you would use to inspect a repo."}],
    "max_tokens": 64
  }'

You should see a chat-completions response rather than an error, and the server logs should show the request passing through the MiniMax M3 parser path. If you later connect an agent framework, this is the endpoint you will point it at.

Common mistakes

• Using the wrong model name. Fix: keep MiniMaxAI/MiniMax-M3-MXFP8 exactly as shown in the recipe unless the vLLM docs say otherwise.
• Forgetting GPU support in Docker. Fix: install the NVIDIA Container Toolkit and rerun with --gpus all.
• Setting tensor parallelism higher than available GPUs. Fix: match --tensor-parallel-size to the number of visible GPUs, or reduce it for a smaller machine.

What's next

Once the server is stable, the next step is to connect an agent framework or client SDK to http://localhost:8000/v1, then tune context length, batching, and GPU memory settings using the vLLM recipe and the MiniMax M3 source notes.