MiniMax Debuts M3.1 Flash, a Multimodal MoE Model

MiniMax has released MiniMax-M3.1-Flash, a multimodal model that combines a mixture-of-experts (MoE) architecture with support for vision, code, and general any-to-any tasks. The model is available on Hugging Face under a custom license.

According to the release, M3.1 Flash is positioned as a versatile workhorse rather than a narrow specialist. It targets three overlapping use cases that have become the center of gravity for frontier open models:

• Agentic workflows, where a model plans and executes multi-step tasks
• Coding, a benchmark-heavy area where open models have made rapid gains
• Video understanding, extending the model beyond text and still images

Why it matters

The "Flash" naming signals a lean, speed-oriented tier, and the MoE design fits that goal: by activating only a subset of parameters per request, mixture-of-experts models can offer large-model capability at lower inference cost. That tradeoff has made MoE the default choice for labs shipping practical, deployable systems.

MiniMax has been steadily expanding its open-weight lineup, and M3.1 Flash reflects a broader industry shift toward consolidating multiple modalities into one model instead of maintaining separate systems for text, vision, and video. Full specifications — including parameter counts, active-expert configuration, and context length — were not detailed in the release record, so developers will want to consult the model card directly before building on it.