MiniMaxAI/MiniMax-M3

Hugging Face
VISIONConcurrent Unit Cost:4Model Size:427BQuant:FP8Context Size:195kPublished:Jun 2, 2026License:otherArchitecture:Transformer1.4K Warm

MiniMaxAI/MiniMax-M3 is a native multimodal model developed by MiniMaxAI, featuring approximately 428 billion parameters with 23 billion activated parameters and a 1 million token context window. It is distinguished by its mixed-modality training from inception, enabling deep semantic fusion across text, image, and video. The model utilizes MiniMax Sparse Attention (MSA) for efficient long-context processing, offering significant speedups and reduced compute, making it highly effective for long-horizon agentic benchmarks, coding, and cowork capabilities.

Loading preview...

MiniMax-M3: A Native Multimodal Model

MiniMax-M3, developed by MiniMaxAI, is a powerful native multimodal model with approximately 428 billion parameters (23 billion activated) and an extensive 1 million token context window. Its core innovation lies in its mixed-modality training approach, integrating text, image, and video data from the very first training step to achieve deeper semantic understanding across modalities.

Key Capabilities & Innovations

  • Native Multimodality: Achieves profound semantic fusion by being trained on mixed-modality data from the outset, unlike models that add modalities later.
  • Context Scaling with MiniMax Sparse Attention (MSA): M3 introduces MiniMax Sparse Attention (MSA), a high-performance sparse attention operator specifically designed for million-token contexts. This technology dramatically improves efficiency, delivering 9x prefill and 15x decode speedups compared to its predecessor (M2) at 1M context, while reducing per-token compute by 20x. More details are available in the technical report.
  • Frontier-Level Agentic Performance: Excels in long-horizon agentic benchmarks, demonstrating strong capabilities in both coding and cowork scenarios.

Recommended Use Cases

  • Complex Multimodal Tasks: Ideal for applications requiring deep understanding and generation across text, image, and video.
  • Long-Context Applications: Highly efficient for tasks demanding extensive context, such as detailed document analysis, long-form content generation, or complex coding projects.
  • Agentic Workflows: Suited for developing AI agents that perform multi-step reasoning and interaction, particularly in coding and collaborative environments.

MiniMax-M3 also offers flexible reasoning modes (enabled, adaptive, disabled) to optimize for specific latency and throughput requirements.