MiniMax M3

MiniMax M3 is a next-generation multimodal model designed for high-performance coding and autonomous agent execution. It delivers strong results across software development, workflow automation, and complex multi-step reasoning tasks, making it a solid foundation for AI assistants and production agents. Powered by MiniMax Sparse Attention (MSA), the model supports up to a 1M-token context window, enabling long-horizon reasoning across large codebases, documents, and extended agent sessions. It natively handles text, image, and video inputs, with multimodal capabilities deeply integrated into its architecture rather than added post-training. MiniMax M3 excels in areas like bug fixing, full-stack development, and performance optimization, while also demonstrating strong capabilities in tool use, task planning, and autonomous execution. It performs well in real-world workflows such as browsing, office tasks, and structured information retrieval, making it suitable for enterprise-grade agentic applications.