Meituan Officially Releases LongCat-2.5: A Native Multimodal Model for Agentic Tasks

Meituan has officially released LongCat-2.5, a native multimodal model designed for long-horizon agentic tasks. The model has 1.6 trillion total parameters, with approximately 48 billion parameters activated per token, and further strengthens native visual understanding and agentic programming capabilities building on LongCat-2.0. The new model shows clear gains on long-horizon agentic tasks. This stems from adding more long-horizon trajectory data to long-sequence training and using large-scale long-horizon reinforcement learning to improve the model's end-to-end engineering capabilities in tasks such as developing complex codebases and building complete projects. We continue to expand this training pipeline around data synthesis, environment construction, and mixed training across harnesses. In addition, we convert failure trajectories from real agents into reproducible, verifiable training tasks, and keep adding training data targeting the model's weak points. At the same time, LongCat-2.5 has native visual understanding. It represents visual signals as multi-level discrete tokens, allowing the backbone model to process visual and text inputs in a unified token format. It performs strongly on Visual Coding and supports tasks such as creative design and 3D. Behind these upgrades is continuously evolving infrastructure. We optimized long-context training and W4A4 inference on domestic computing hardware, and supported efficient large-scale long-horizon RL through training-inference numerical consistency and long-context rollout optimization. Benchmark Dashboard To show LongCat-2.5's current capability level intuitively, the chart below compares it with GPT-5.5, GPT-5.6-Sol, Claude-Opus-4.8, and Claude-Opus-5. Detailed evaluation results covering more models are in the tables below. 01 Visual Representation Modeling LongCat-2.5 represents visual signals as multi-level discrete tokens, allowing the backbone model to process visual and text inputs in a unified token format. This design combines native-resolution visual discretization, offline visual preprocessing, and visual embeddings learned jointly with the backbone model. It supports native multimodal fusion while simplifying visual computation and load management during training. Native-resolution discrete visual Transformer: for