MiniMax M3 review: an open model for agentic coding and vision

MiniMax M3 (minimax-m3) is the model to look at when you want agentic coding and image or video understanding from one open-weight model. Compared with MiniMax M2.7, multimodal input is no longer bolted on; it is part of the design.

Architecture: 428B MoE, about 23B active

MiniMax released M3 as open weights at the turn of May and June 2026. It is a Mixture-of-Experts model with 428B total parameters and roughly 23B active per token, which keeps inference cost low for its size.

• Context: about 1M tokens as published by MiniMax, with up to 512K tokens of output
• Input: text, images and video
• Output: text

The notable part is that images and video were mixed into training from the first stage, rather than a vision encoder grafted onto a finished text model. MiniMax pitches it for work like reading screenshots and driving a desktop.

On FastMetal, M3 runs on DeepInfra with a 524,288-token context window, half of MiniMax's published figure. Image input, tool calls and strict JSON schema output were verified on our route with real calls; video input is MiniMax's claim and we have not tested it through the gateway.

Benchmarks

MiniMax describes M3 as the first open-weight model to combine reasoning, agentic coding and multimodal input in one model. The reported figures:

The 59.0% on SWE-Bench Pro was reported ahead of GPT-5.5 and Gemini 3.1 Pro. M3 was also reported ahead of Gemini 3.1 Pro on OmniDocBench (document parsing) and ahead of Claude Opus 4.7 on SVG-Bench.

The speed story is MSA (MiniMax Sparse Attention). At a 1M-token context, MiniMax reports more than 9x faster prefill and more than 15x faster decoding than M2, at a twentieth of the compute per token. If you actually fill a long context, that is the difference you notice.

Cost

Reports put M3 well below GPT-5.5 or Gemini 3.1 Pro in cost for comparable headline benchmark scores, which is the usual open-weight MoE trade.

On FastMetal, M3 is billed per million input and output tokens; the current prices are on the model page. Images count as input tokens, and visual input adds up fast, so measure a few real requests before you estimate a batch job.

Non-English use

M3 is multilingual and handles general tasks in other languages, Japanese included. We found no language-specific benchmarks for it. Multimodal prompts in particular are where instruction-following tends to wobble outside English, so test on your own prompts before production.

Good fit, poor fit

Good fit

• Agent workflows whose inputs include screenshots or other images
• UI automation and computer use driven by screenshots
• Coding and vision work you would rather run on one model
• Long-context jobs that actually use hundreds of thousands of tokens

Poor fit

• Text-only work, where lighter and cheaper models will do
• Writing where nuance in a specific language decides the quality
• Audio input, which it does not support

FAQ

Is it worth moving from MiniMax M2.7?

If you work with images, yes. For text-only use, it comes down to whether the long-context speedup matters for your workload.

Does MiniMax M3 really take images and video?

MiniMax trained it on text, images and video from the start, with text output. On FastMetal, image input is verified; video has not been tested through the gateway.

What context window does FastMetal serve for MiniMax M3?

524,288 tokens, from DeepInfra. MiniMax's own published figure is about 1M.

Try it on FastMetal

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax-m3",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

M3 sits behind the same API key and balance as closed models like Claude, so you can change model and compare the two on your own tasks. The full list is in the model catalog. Your own evals, not public leaderboards, are what should decide whether it goes into production.