Fifteen cents per million input tokens. Sixty cents out. For a model that reads text, images and video across a million-token window, that is the price point that makes MiniMax's general-availability release worth a look before anything else about it.
This is the finished version of the M3.1 Flash Preview that appeared in late September. Before launch it also ran anonymously on OpenRouter under the name "Space Bunny". The API went live on 19 October 2026 and weights reached Hugging Face on 23 October.
Quick reference
| Field |
Value |
| API ID |
MiniMax-M3.1-Flash |
| Architecture |
Native multimodal mixture-of-experts |
| Parameters |
About 428B total, 23B active (model card) |
| Context |
Up to 1,000,000 tokens |
| Input / output |
Text, image, video / text |
| Reasoning |
Always on; effort low, medium, high, xhigh, max |
| Recommended sampling |
temperature 1.0, top_p 0.95 |
| Licence |
MiniMax Community License |
Why it is cheap to run
Two engineering choices do most of the work. First, MiniMax Sparse Attention now runs through the whole stack; the three full-attention layers at the start of the previous design were replaced with sparse ones. Second, a DSpark speculative-decoding draft head ships with the weights, replacing the earlier multi-token-prediction approach. Quantisation-aware training covers the routed experts and attention paths too.
MiniMax claims about 70% lower cost on agent tasks and roughly three times faster inference than its prior model. Those are vendor claims. They are also exactly the kind of claim you can check in an afternoon.
Publisher-reported scores
- DeepSWE v1.1: 73.7
- BankerToolBench: 76.2
- OSWorld 2.0: 61.7
- Terminal Bench 4.0: 38.1
- CyberGym: 82.2
MiniMax notes these were measured on the official weights and that third-party quantisations may score differently.
The parameter mismatch
The model card describes about 428B total and 23B active. The Hugging Face file metadata, however, reads 220B. We cannot tell from the outside which tensors account for the difference. If you are planning GPU memory, measure the downloaded checkpoint rather than trusting either figure.
Pricing and terms
Pay-as-you-go, indicative USD per 1M tokens: $0.15 input, $0.60 output, $0.015 for cached prompt reads. No highspeed tier is listed for this model, which is a change from older lines such as MiniMax-M2.7-highspeed. Confirm current rates on your own MiniMax account.
The licence is MiniMax's own community licence, not an OSI-approved one. Commercial use needs attribution or notification, and written authorisation is said to be required above roughly $20M in annual revenue. Read the licence file yourself if you are near that line.
For self-hosting, SGLang and vLLM are the recommended engines, with TokenSpeed also supported. Transformers, KTransformers and others were listed as in progress.
A test plan that fits the price
Take one agent workflow you already run on MiniMax-M3 or a competitor. Run it at medium effort, then at high. Track three numbers: tasks completed, total tokens, and wall-clock time. Because reasoning cannot be switched off, short classification calls will always carry some thinking overhead; if that matters, measure it explicitly.
The rest of the lineup sits on the MiniMax provider page.