MiniMax-M3.1-Flash: open weights at a fraction of the price
New models on Complete
MiniMax-M3.1-Flash is out with open weights, and it's available for enterprise teams on Complete. It's a low-cost model for coding and automated workflows that companies can run on their own infrastructure or through US cloud providers.
MiniMax published the full weights on Hugging Face on October 17, 2026. The model reads text, images and video, and handles up to 1 million tokens in a single request.
The numbers
Model | Intelligence Index (v4.3.2) | Cost to run the full index |
|---|---|---|
GPT-6.1 Terra | 48.1 | $909 |
MiMo-V2.6-Pro | 46.3 | $207 |
DeepSeek V4.1 Flash 1017 | 45.1 | $445 |
Claude Haiku 5.5 | 43.4 | $330 |
MiniMax-M3.1-Flash | 42.7 | $283 |
GPT-6 Luna | 38.1 | $122 |
Source: Artificial Analysis, October 2026.
MiniMax-M3.1-Flash scores within a point of Claude Haiku 5.5 and costs about 14% less to run the same tests. It costs under a third of GPT-6.1 Terra. Xiaomi's MiMo-V2.6-Pro scores higher for less on the same index, so MiniMax's case rests on open weights, video input and broad hosting rather than raw score per dollar.
What you get
Open weights under the MiniMax Community License, a custom license set by MiniMax rather than a standard open-source license.
Text, image and video input with text output.
A context window of up to 1 million tokens.
A mixture-of-experts design with about 428 billion parameters, of which about 23 billion are active for each token, so it runs on less hardware than its total size suggests.
Broad hosting: Artificial Analysis tracks 13 providers, and OpenRouter lists US providers including Together AI, Nebius, DigitalOcean, SambaNova and Parasail.
What it costs
Artificial Analysis and OpenRouter list MiniMax-M3.1-Flash at $0.15 per million input tokens and $0.60 per million output tokens. The model writes long answers, producing about twice as much output per task as GPT-6.1 Terra on the Artificial Analysis tests, so the cost per finished task is higher than the per-token price suggests. The $283 figure above already accounts for that.
Artificial Analysis measures a median output speed of 118.6 tokens per second, about half the speed of Claude Haiku 5.5.
Where it performs
MiniMax reports strong results on its own agent and coding tests, including 73.7% on DeepSWE v1.1 and 61.7 on OSWorld V2. These are vendor-reported figures.
On independent tests from Artificial Analysis, it scores 38.6% on FrontierCode and 25.8% on Terminal-Bench 4.0, compared with 55.6% on Terminal-Bench 4.0 for GPT-6.1 Terra.
What to watch
On the Artificial Analysis knowledge test, MiniMax-M3.1-Flash guessed wrong instead of declining 92.8% of the time it didn't know the answer. It fits work where answers come from your own documents and data, such as extraction, classification, code changes and grounded search. It is a poor fit for open-ended factual questions with no source attached.
Run it on your own workflows
MiniMax-M3.1-Flash is available for enterprise teams on Complete. Bring a high-volume workload and compare it against the models you run today.
