Meituan's LongCat Launches LongCat-2.5, a 1.6-Trillion-Parameter AI Model

Meituan's LongCat team has launched LongCat-2.5, a native multimodal model built for long-horizon agent tasks. The model has 1.6 trillion total parameters, with about 48 billion activated per token, and improves visual understanding and agentic coding over LongCat-2.0.

During training, the team added more long-horizon trajectory data and used large-scale reinforcement learning to boost end-to-end engineering skills, such as working with complex codebases and building full projects. Real agent failures were also turned into reproducible, verifiable training tasks.

LongCat-2.5 turns visual signals into multi-level discrete tokens, so the model handles images and text in a unified way. It performs well on Visual Coding and supports tasks like creative design and 3D work.

The team also optimized long-context training and W4A4 inference on domestic computing power. The model is now available online, with GitHub and HuggingFace versions coming soon.