Meituan releases 1.6T parameter LongCat-2.5 AI model
Meituan has launched LongCat-2.5, a multimodal AI model featuring 1.6 trillion parameters, optimized for long-sequence agent tasks and visual coding.
Meituan’s LongCat division has announced the release of LongCat-2.5, an advanced multimodal model designed to handle complex, long-range agentic tasks. The model, which contains 1.6 trillion parameters, activates approximately 48 billion parameters per token. This update marks a transition toward enhanced end-to-end engineering capabilities, specifically in complex codebase development and software building.
To improve task performance, the development team employed large-scale reinforcement learning, training the model on long-range trajectory data. By converting previous task failures into verifiable training sequences, the system addresses capacity gaps identified in earlier versions. The architecture represents visual signals as multi-layered discrete tokens, allowing the model to process visual and textual inputs through a unified framework.
The model is supported by an updated infrastructure designed to optimize long-context training and inference on domestic hardware. Meituan has also focused on numerical consistency to ensure efficient operation during large-scale reinforcement learning tasks.
LongCat-2.5 is currently accessible via the company’s web portal and API platform. Developers can expect further technical releases on GitHub and HuggingFace in the near future.