MiniMax M3.1 Flash: An Open Multimodal Agent Model
MiniMax published M3.1 Flash weights on Hugging Face. It is a native multimodal MoE built for coding, agentic workflows, computer use and cybersecurity, accepting text, images and video with up to a 1M-token context. Versus M3 it adds sparse attention throughout the model, quantization-aware training and DSpark speculative decoding, and the model card ships a comparison table spanning automation, terminal work, code repair and visual reasoning
Tool Interface
Interactive tool will be available soon
Features
- ✓ Native multimodal MoE with text, image and video input and a context window of up to 1M tokens, with usable length depending on GPU memory and serving configuration
- ✓ Three main changes over M3: sparse attention throughout the model, quantization-aware training and DSpark speculative decoding, aimed at both task performance and inference efficiency
- ✓ On Automation-Bench v1.0.6 it scores 62.8, where the same table lists DeepSeek V4.1 Flash-1017 at 60.5 and Kimi K3 at 46.7
- ✓ Terminal and desktop work: Terminal Bench 4.0 at 38.1 and OSWorld 2.0 at 61.7, plus 60.3 on the long-context LOCA-Bench 256K
- ✓ Coding and repo-level tasks: DeepSWE v1.1 at 73.7, NL2Repo-Bench at 61.2 and SWE Refactor Bench at 23.0, with CyberGym at 82.2 on the security side
- ✓ Vision and knowledge: CharXiv RQ 82.5, BabyVision with tools 91.0 and HLE without tools 43.4; recommended sampling is temperature=1.0 and top_p=0.95
How to Use
- Start with the model card at MiniMaxAI/MiniMax-M3.1-Flash on Hugging Face and confirm your VRAM can hold the context length you actually need
- Pull the weights with hf download MiniMaxAI/MiniMax-M3.1-Flash --local-dir MiniMax-M3.1-Flash, then serve them through one of the inference frameworks the card recommends
- Sample with temperature=1.0 and top_p=0.95 as suggested, and set thinking depth through reasoning_effort in the chat template, which offers low, medium, high, xhigh and max
- Define permission boundaries before wiring it into an agent: give it as little access as possible for computer-use and terminal tasks, because benchmark scores are not a substitute for runtime safety
FAQ
What is M3.1 Flash?
Per the model card, it is a native multimodal MoE model for coding, agentic workflows, computer use and cybersecurity, supporting text, image and video inputs with a context window of up to 1M tokens. The weights are at huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash.
What changed from M3?
MiniMax says it keeps M3's core MoE architecture with three main changes: sparse attention throughout the model, quantization-aware training, and DSpark speculative decoding, targeting both task performance and inference efficiency.
How do the published scores look?
The model card includes a multi-model comparison. M3.1 Flash scores 62.8 on Automation-Bench v1.0.6 (DeepSeek V4.1 Flash-1017 at 60.5, Kimi K3 at 46.7), 73.7 on DeepSWE v1.1 (DeepSeek is higher there at 77.0), 38.1 on Terminal Bench 4.0, 61.7 on OSWorld 2.0, 82.2 on CyberGym and 43.4 on HLE without tools. The card also warns that third-party quantizations may differ from the official weights.
How do I deploy it locally?
The card's command is hf download MiniMaxAI/MiniMax-M3.1-Flash --local-dir MiniMax-M3.1-Flash, followed by one of the recommended inference frameworks. Integrations with Transformers, KTransformers, unsloth and ATOM are listed as in progress. Context goes up to 1M tokens, but usable length and throughput depend on GPU memory and serving configuration.
Any parameter gotchas?
Recommended sampling is temperature=1.0 and top_p=0.95, and reasoning depth is set through reasoning_effort in the chat template with five levels: low, medium, high, xhigh and max. For computer-use or terminal tasks, tighten tool permissions and sandbox boundaries rather than treating benchmark scores as a proxy for production safety.