MiniMaxAI/MiniMax-M3
MiniMaxAI's MiniMax-M3 is a native multimodal model featuring approximately 428 billion parameters with 23 billion activated, and a 1 million token context length. It is distinguished by its mixed-modality training from inception, enabling deep semantic fusion across text, image, and video. M3 utilizes MiniMax Sparse Attention (MSA) for efficient long-context processing, achieving significant speedups and reduced compute, and excels in agentic benchmarks for coding and cowork capabilities.
Loading preview...
MiniMax-M3: A Native Multimodal Model with 1M Context
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 length. Its core innovation lies in its native multimodality, undergoing mixed-modality training from the very first step to achieve deeper semantic fusion across text, image, and video data.
Key Capabilities & Innovations
- Native Multimodality: Designed for deep semantic understanding across text, image, and video through integrated training.
- Context Scaling via MiniMax Sparse Attention (MSA): M3 introduces MiniMax Sparse Attention (MSA), a high-performance sparse attention operator, to significantly improve long context efficiency. This technology delivers 9x prefill and 15x decode speedups compared to its predecessor (M2) at 1M context, reducing per-token compute by 20 times. More details are available in the technical report on arXiv.
- Coding & Cowork Capability: The model demonstrates frontier-level performance in long-horizon agentic benchmarks, making it highly proficient in both coding tasks and collaborative work scenarios.
Reasoning Modes
M3 supports three distinct reasoning modes, controlled via the thinking parameter, allowing users to balance reasoning depth with latency and throughput:
enabled: Reasoning is always active.adaptive: M3 intelligently decides when to apply additional reasoning.disabled: Reasoning is turned off for minimal latency and maximum throughput.
Recommended Use Cases
MiniMax-M3 is particularly well-suited for applications requiring:
- Advanced multimodal understanding and generation.
- Processing and generating content within very long contexts (up to 1 million tokens).
- Agentic tasks, especially in coding and collaborative environments, where its performance on long-horizon benchmarks is a significant advantage.