Last updated: October 11, 2026

MiniMax M3: review, API guide, pricing and local deployment

MiniMax released M3 on June 1, 2026. It is an open-weight mixture-of-experts model with about 428B total and 23B active parameters, a 1M-token context window and native image and video input. This page collects what MiniMax has published, shows how to call it through the API, compares its price with Claude and GPT, and covers what it takes to run it yourself.

MiniMax M3 at a glance

A newer MiniMax-M3.1-Flash-Preview is listed in MiniMax’s API docs, but only for its M Plan and MiniMax Code app so far.

Source: MiniMax: MiniMax M3 launch post

Source: MiniMax-AI/MiniMax-M3 on GitHub

Source: MiniMaxAI/MiniMax-M3 on Hugging Face

Source: MiniMax API: text generation guide

Benchmarks MiniMax published

These are MiniMax’s own numbers. We have not reproduced them, and MiniMax did not publish a same-setup comparison with Claude Opus 5.5, GPT-6 or Gemini 4.

MiniMax also describes two long runs: about 12 hours reproducing a research paper (18 commits, 23 figures), and about 24 hours tuning a CUDA FP8 GEMM kernel from 7.6% to 71.3% of hardware peak.

Source: MiniMax: MiniMax M3 launch post

Source: MiniMaxAI/MiniMax-M3 on Hugging Face

Test it yourself: five prompts

The fastest way to judge M3 for your work is to run the same prompts against it and the model you use now. These cover the areas MiniMax claims as strengths.

Here is a failing test and the function it tests. Find the bug, explain the cause in two sentences, and give the corrected function only.

[paste test]
[paste function]

I am pasting a whole repository below. List every place where a database connection is opened but not closed, with file and line, then propose one shared helper that fixes all of them.

[paste files]

Turn this UI screenshot into a single HTML file with inline CSS. Match spacing and colors closely and make it responsive down to 360px wide.

Watch this screen recording of a bug. List the exact steps that reproduce it and what the user expected to happen instead.

You have shell access to a Node project. Plan, step by step, how you would upgrade it from Express 4 to Express 5, including how you will check nothing broke. Do not run anything yet.

Use temperature 1.0 and top_p 0.95, the sampling settings MiniMax recommends.

Source: MiniMax-AI/MiniMax-M3 on GitHub

How to use the MiniMax M3 API

1. Create an account on platform.minimax.io and add a payment method or buy a Token Plan.
2. Create an API key in the console.
3. Point an Anthropic SDK (recommended by MiniMax) at https://api.minimax.io/anthropic, or an OpenAI SDK at https://api.minimax.io/v1.
4. Set the model to MiniMax-M3.
5. Keep prompts under 512K input tokens where you can; above that the price doubles.

# pip install anthropic
import anthropic

client = anthropic.Anthropic(
    base_url="https://api.minimax.io/anthropic",
    api_key="YOUR_MINIMAX_API_KEY",
)

msg = client.messages.create(
    model="MiniMax-M3",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Write a Python function that dedupes a list while keeping order."}],
)
print(msg.content)

Source: MiniMax API: text generation guide

OpenAI-compatible call (curl)

curl https://api.minimax.io/v1/chat/completions \
  -H "Authorization: Bearer $MINIMAX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMax-M3",
    "messages": [{"role": "user", "content": "Explain MoE models in 3 sentences."}],
    "temperature": 1.0,
    "top_p": 0.95
  }'

Existing tools that speak the Anthropic or OpenAI API, such as coding agents, usually only need the base URL, key and model name changed.

Source: MiniMax API: text generation guide

MiniMax M3 pricing

MiniMax labels the discount “permanent 50% off”. Thinking on or off costs the same. The Token Plan subscription is the alternative: Plus at $20 a month (about 1.7B M3 tokens), Max at $50 (about 5.1B) and Ultra at $120 (about 9.8B), with text, image, speech and music sharing one pool.

Source: MiniMax API pay-as-you-go pricing

Source: MiniMax: MiniMax M3 launch post

MiniMax M3 vs Claude and GPT models

• Pick M3 when you need open weights (self-hosting, fine-tuning, data that cannot leave your servers) or cheap long-context coding with image and video input.
• Haiku 5.5 and GPT-6 Luna are cheaper per token for simple, high-volume jobs.
• For the hardest coding work, compare on your own tasks: no shared benchmark run covers M3 against Opus 5.5 or GPT-6 Astra yet.

Prices are list prices for the shortest prompt tier. Claude and OpenAI prices come from their pricing pages as tracked on this site.

All model API prices · Cost calculator · Claude Haiku 5.5

Source: MiniMax API pay-as-you-go pricing

Source: Claude API pricing

Source: OpenAI API pricing

How to run MiniMax M3 locally

The full model is about 855 GB in BF16. MiniMax and vLLM target a single node of eight Hopper GPUs (8x H200 or H20) for BF16. An MXFP8 checkpoint (MiniMaxAI/MiniMax-M3-MXFP8) needs roughly half the memory.

# 1) Download the weights (~855 GB in BF16)
hf download MiniMaxAI/MiniMax-M3 --local-dir MiniMax-M3

# 2) Serve with vLLM on 8 Hopper GPUs (e.g. 8x H200).
#    Support is not in a stable vLLM release yet: use the vllm/vllm-openai:minimax-m3 image.
vllm serve MiniMaxAI/MiniMax-M3 \
  --tensor-parallel-size 8 \
  --block-size 128 \
  --tool-call-parser minimax_m3 \
  --reasoning-parser minimax_m3 \
  --enable-auto-tool-choice

# Or SGLang
python3 -m sglang.launch_server \
  --model-path "MiniMaxAI/MiniMax-M3" \
  --host 0.0.0.0 --port 30000

The vLLM recipe says --block-size 128 is mandatory, and the tool-call parser is minimax_m3, not minimax_m2.

Source: MiniMax-AI/MiniMax-M3 on GitHub

Source: vLLM recipe: MiniMax M3

Source: MiniMaxAI/MiniMax-M3 on Hugging Face

Smaller machines: Unsloth GGUF (experimental)

# Experimental, text-only GGUF (llama.cpp branch with M3 support)
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
git fetch origin pull/24523/head:minimax-m3
git checkout minimax-m3
cmake -B build -DGGML_CUDA=ON   # -DGGML_CUDA=OFF for CPU or Apple Metal
cmake --build build --config Release -j --target llama-cli llama-server

./build/bin/llama-cli \
  -hf unsloth/MiniMax-M3-GGUF:UD-IQ1_M \
  --temp 1.0 --top-p 0.95 --top-k 40 \
  --ctx-size 32768

The GGUF is text-only and does not support sparse attention yet, so long contexts fall back to dense attention and use far more memory. Keep --ctx-size modest.

Source: Unsloth: run MiniMax M3 locally

Can you use it commercially?

• Non-commercial use is free.
• Commercial use is allowed, but you must show “Built with MiniMax M3” on a related site, interface or docs page.
• Companies under $20M yearly revenue send a one-time notice to [email protected] (subject “M3 licensing — notice”). Above $20M you need written permission from MiniMax.
• Military use, unlawful content, harm to minors and deliberate misinformation are prohibited.

This is a summary, not legal advice. Read the license before you ship a product on it.

Source: MiniMax Community License

Independent guide, not affiliated with MiniMax. Facts come from MiniMax’s launch post, GitHub, Hugging Face card, license and API docs, the vLLM recipe and Unsloth’s guide, checked on October 11, 2026. Benchmarks are MiniMax’s own and were not reproduced by us. Prices may change.