Meituan Releases LongCat-2.5: A 1.6T Parameter Native Multimodal AI Model
Meituan's LongCat team released LongCat-2.5, a 1.6-trillion parameter native multimodal AI model designed for long-horizon agent tasks and visual coding. Building on LongCat-2.0, it features expanded long-sequence training, large-scale reinforcement learning, and a unified token representation for both text and visual signals. This release represents a significant advancement in large-scale AI engineering, focusing heavily on end-to-end capabilities for complex code development and multi-harness collaboration. By tackling long-horizon agent tasks and visual coding, it pushes the boundaries of how AI handles extended, multi-step engineering workflows. LongCat-2.5 activates approximately 48 billion parameters per token and represents visual signals as multi-level discrete tokens. It is supported by infrastructure optimizations including domestic hardware support for long-context training, W4A4 inference, and numerical consistency in reinforcement learning rollouts.
## BACKGROUND
Meituan is a major Chinese technology company known for operating a comprehensive local services platform covering food delivery, travel bookings, and instant retail. In the context of large language models, long-horizon agent tasks refer to complex workflows where an AI agent must autonomously plan, use tools, and execute multi-step operations over hours or days without losing context or failing.