Meituan releases LongCat-2.5, a 1.6 trillion parameter native multimodal AI model for long-horizon agents

Meituan has unveiled LongCat-2.5, a new native multimodal AI model designed specifically for long-horizon agent tasks. This model boasts a total parameter count of 1.6 trillion, with approximately 48 billion parameters activated per token. It represents an advancement over LongCat-2, significantly enhancing native visual understanding and agent programming capabilities. The development involved incorporating extensive long-trajectory data into the long-sequence training process, coupled with large-scale long-horizon reinforcement learning. This has bolstered the model's end-to-end engineering proficiency in complex tasks such as full project construction and code repository development. Furthermore, the team is expanding the training pipeline through data synthesis, environment construction, and cross-Harness mixed training. LongCat-2.5 processes visual and text inputs using a unified token format by representing visual signals as multi-level discrete tokens. This integration of native visual understanding and agent programming allows the model to excel in Visual Coding and support creative design and 3D applications. The infrastructure supporting these upgrades includes optimizations for long-context training and W4A4 inference on domestic computing power, ensuring efficient large-scale long-horizon RL operation.