Back to news
launchMeituan2026-11-02

Meituan LongCat releases LongCat-2.5 native multimodal model with 1.6 trillion parameters and 48 billion active per token

Meituan's LongCat team released LongCat-2.5 on November 2, a native multimodal model aimed at long-horizon agentic tasks with 1.6 trillion total parameters and roughly 48 billion active parameters per token, building on LongCat-2.0 with stronger native visual understanding and agentic coding.

Meituan's LongCat team released LongCat-2.5 on November 2, a native multimodal model aimed at long-horizon agentic tasks with 1.6 trillion total parameters and roughly 48 billion active parameters per token, building on LongCat-2.0 with stronger native visual understanding and agentic coding. The team said it added more long-horizon trajectory data to long-sequence training and applied large-scale long-horizon reinforcement learning to improve end-to-end engineering capability on complex codebases and full project builds, extending a training pipeline that spans data synthesis, environment construction and cross-harness mixed training. Real agent failure trajectories were converted into reproducible and verifiable training tasks to cover capability gaps. LongCat-2.5 represents visual signals as multi-level discrete tokens so the backbone handles visual and textual input in a unified token format, which combined with native visual understanding gives strong Visual Coding performance and supports creative design and 3D tasks. The company credits infrastructure work including optimized long-context training and W4A4 inference on domestic compute, plus train-inference numerical consistency and long-context rollout optimization for efficient large-scale long-horizon RL. Online chat and API access are live, with Github and HuggingFace weights to follow.

美团LongCat-2.5多模态模型智能体长程任务开源