Meituan releases LongCat-2.5 AI model for coding and visual tasks
Meituan has launched the LongCat-2.5 AI model, featuring 1.6 trillion parameters, specifically optimized for complex coding and multi-step software development tasks.
On November 2, Meituan introduced LongCat-2.5, an AI model designed to enhance visual reasoning and programming capabilities. The architecture utilizes 1.6 trillion total parameters, with 48 billion activated per token, focusing on the requirements of long-running AI agent tasks.
To improve performance on complex codebases, Meituan expanded training on lengthy task sequences and employed large-scale reinforcement learning. A key feature of the model's development is the use of failure-to-training conversion, where errors encountered by autonomous agents are transformed into reproducible scenarios to address performance gaps.
LongCat-2.5 processes visual and text inputs within a unified framework to support visual coding, design, and 3D tasks. The model is currently accessible via Meituan’s chat service and API, with planned releases on GitHub and Hugging Face in the near future.