Meituan LongCat has released LongCat-2.5, a native multimodal model designed for long-horizon agent tasks. The model features 1.6 trillion total parameters, with approximately 48 billion activated per token. Building on LongCat-2.0, it delivers enhanced native visual understanding and agentic coding capabilities. To strengthen end-to-end engineering performance, the model incorporates more long-horizon trajectory data and applies large-scale long-horizon reinforcement learning. It also converts real-world agent failure trajectories into training tasks to address key shortcomings. In addition, the model represents visual signals as multi-level discrete tokens, enabling tasks such as visual coding, creative design, and 3D. These upgrades are supported by continuously evolving infrastructure, including optimized long-context training on domestic compute and W4A4 inference.