How to run MiniMax-M3 with vLLM and SGLang

This guide covers serving the MiniMax-M3 family. Memory sizing and launch commands appear below.

Engine support

The model card points to SGLang, vLLM, Transformers, KTransformers, unsloth and ATOM (AMD ROCm), and lists quantized builds for llama.cpp, Ollama and LM Studio. It states no minimum version. The vLLM recipe says vLLM 0.24.0 or newer is required and notes that support had not yet shipped in a stable release when it was written, so check which build you install.

vLLM flags

• --tool-call-parser minimax_m3 and --reasoning-parser minimax_m3 (earlier MiniMax releases used minimax_m2, so do not reuse old commands)
• --block-size 128, called mandatory on all platforms for the sparse attention indexing
• --trust-remote-code, listed as required for the NVFP4 and MXFP4 variants; the card also requires trust_remote_code=True for Transformers
• --mm-encoder-tp-mode data for data-parallel vision encoding at high tensor parallelism
• --language-model-only for text-only workloads

Thinking modes

Set thinking_mode through chat_template_kwargs:

• enabled: always reasons
• adaptive: the model decides (the recipe's default)
• disabled: no reasoning, lowest latency

Context and inputs

Context is 1M tokens. The card documents no rope or YaRN override. The model accepts text, images and video natively.

Sampling

The card recommends temperature=1.0 and top_p=0.95.

Pitfalls

• Missing --block-size 128.
• Using the old minimax_m2 parsers.
• Loading the vision stack when you only need text.