MiniMax M3.1 Flash
MiniMax · China · #4 in our ranking · Best for finance agents
MiniMax M3.1 Flash is an open-weight mixture-of-experts model from MiniMax (China), released Oct 2026, with 428B parameters (23B active per token), a 1M-token context window, and the MiniMax Community License.
MiniMax M3.1 Flash builds on M3 with sparse attention throughout the model, quantization-aware training, and DSpark speculative decoding, so it is cheaper and faster to serve. It reads text, images, and video with a 1M-token context, and MiniMax reports 76.2 on BankerToolBench, level with Opus 5 in its table.
MiniMax M3.1 Flash specs
- Provider
- MiniMax
- Origin
- China
- Architecture
- Mixture-of-experts
- Parameters
- 428B total · 23B active
- Context
- 1M tokens | 786K words
- License
- MiniMax Community License
- Input
- Text, Image, Video
- Released
- Oct 2026
- Weights
- Hugging Face
Sourced from the model card on Hugging Face. Last checked 2026-10-29.
Highlights
- 76.2 on BankerToolBench
- Text, image, and video input
- 1M-token context
Where it fits for payment teams
- Partner-bank RFIs
#3 pick. Multimodal with sparse attention built for million-token contexts, useful for long correspondence threads.
- Chargebacks and disputes
#1 pick. Reads text, images, and video with a 1M-token context, so a whole dispute file fits in one pass, and it posts the top open-weight BankerToolBench score in MiniMax's report.
Harnesses in Runtime
Run MiniMax M3.1 Flash through any of these harnesses on Runtime.
Provider-reported benchmarks
| BankerToolBench | 76.2 |
|---|---|
| OSWorld V2 | 61.7 |
| CyberGym | 82.2 |
Benchmark scores are reported by each model provider on its Hugging Face model card. They are not independently verified by Runtime.
How to run MiniMax M3.1 Flash
Download the weights
Pull MiniMaxAI/MiniMax-M3.1-Flash from Hugging Face. Check the MiniMax Community License.
Serve it in your cloud
vLLM or SGLang on your GPUs, or a managed host.
Put it to work in Runtime
Point an agent or skill at it, then compare with evals.
MiniMax M3.1 Flash FAQ
What is the context window of MiniMax M3.1 Flash?
+
MiniMax M3.1 Flash supports a 1M-token context window (1,048,576 tokens), roughly 786K words of English text.
How many parameters does MiniMax M3.1 Flash have?
+
MiniMax M3.1 Flash is a mixture-of-experts model with 428B total parameters, of which 23B are active per token.
Is MiniMax M3.1 Flash free for commercial use?
+
MiniMax M3.1 Flash is released under the MiniMax Community License, a custom license. Commercial use may come with conditions, so read the license on Hugging Face before you deploy or fine-tune it.
What can MiniMax M3.1 Flash take as input?
+
MiniMax M3.1 Flash accepts text, image, video input and generates text.
Is MiniMax M3.1 Flash good for partner-bank RFIs?
+
Yes. MiniMax M3.1 Flash is our #3 pick for partner-bank RFIs. Multimodal with sparse attention built for million-token contexts, useful for long correspondence threads.
Is MiniMax M3.1 Flash good for chargebacks and disputes?
+
Yes. MiniMax M3.1 Flash is our #1 pick for chargebacks and disputes. Reads text, images, and video with a 1M-token context, so a whole dispute file fits in one pass, and it posts the top open-weight BankerToolBench score in MiniMax's report.
Can I run MiniMax M3.1 Flash in my own cloud with Runtime?
+
Yes. Download the weights from Hugging Face (MiniMaxAI/MiniMax-M3.1-Flash) and serve them with vLLM or SGLang in your cloud, or use a managed provider that hosts the model. Runtime agents can then use MiniMax M3.1 Flash through a harness like OpenCode, with your data staying in your environment.