MiniMax-M3 is a natively multimodal large language model published by MiniMax and released with open weights, an accompanying GitHub repository, and a research paper on arXiv (2606.13392). The model is part of MiniMax's text LLM family on minimax.io, sitting alongside related MiniMax M2.7 and M2.5 offerings, and is positioned as a next-generation step that integrates text, image, and video understanding from the very first training step rather than bolting modalities on after a text-only base. That mixed-modality pretraining approach is intended to give M3 deeper cross-modal semantic fusion, which is useful for tasks that require grounding language in visual or video context rather than treating images and video as an afterthought.

The architecture combines a very large total parameter count of roughly 428B with about 23B activated parameters per token, suggesting a Mixture-of-Experts style design aimed at keeping inference cost tractable while preserving broad knowledge. To make the model's 1M-token context practical, M3 introduces MiniMax Sparse Attention (MSA), a long-context efficiency mechanism that reduces the cost of attending over very long inputs and supports sustained generation across extended documents and agent-style workloads. Together, the sparse attention design and large active memory make M3 a natural fit for long-horizon agentic applications, multi-document reasoning, and rich multimodal analysis where both deep context and cross-modal understanding matter.