XiaomiMiMo/MiMo-v2.5 (huggingface.co)

Xiaomi has announced its latest AI model, MiMo-V2.5, a sophisticated omnimodal system capable of processing text, images, video, and audio within a single architecture. Built on the MiMo-V2-Flash backbone, this model features advanced agentic capabilities that enhance multimodal perception and long-context reasoning. A standout feature is its Hybrid Attention Architecture, which combines Sliding Window Attention and Global Attention in a scalable fashion that significantly reduces storage needs while efficiently handling context lengths of up to 1 million tokens. The significance of MiMo-V2.5 lies in its ability to streamline multimodal AI tasks, making it particularly relevant for applications requiring nuanced understanding across different data types. The model's native encoders—comprising a Vision Transformer with 729 million parameters and an audio transformer initialized from MiMo-Audio—enable high-quality outputs. Additional enhancements such as Multi-Token Prediction modules accelerate inference and improve reinforcement learning training efficiency. By leveraging efficient pre-training methods with around 48 trillion tokens, MiMo-V2.5 positions itself as a powerful tool for developers and researchers aiming to advance AI-driven solutions in complex environments.