In a recent development, the company Meituan has released an enormous multimodal AI model that can help accomplish complex coding and other tasks with greater efficiency. This multimodal AI model, known as LongCat-2.5, has been designed specifically for handling longer tasks.
Breakthrough Capabilities of Meituan’s New AI Agent LongCat-2.5
On November 2, 2026, Meituan unveiled its AI agent, LongCat-2.5, which contains 1.6 trillion total parameters and about 48 billion parameters per token. It consists of one architecture that processes both visual and textual inputs and can perform various tasks such as visual coding, design, and 3D work.
Unlike other AI models, LongCat-2.5 is capable of accomplishing longer tasks such as working in large software projects and managing coding workflows. Meituan has also trained its model using large-scale reinforcement learning and converted actual failures of AI agents into training examples to make it able to accomplish multi-step tasks.
One of the most important characteristics of the model LongCat-2.5 is its ability to function as an autonomous AI agent. While maintaining consistency throughout the whole workflow process, the enormous number of parameters, along with its sparse architecture, enables the model to activate just a part of its parameters in each of its computational steps. LongCat-2.5 is currently accessible on Meituan’s online chat interface and API platform, but its upcoming releases on GitHub and Hugging Face platforms are expected soon.
However, some of the details, such as benchmark scores, API prices, context window sizes and the license type of the open weights, remain unknown. Thus, developers are eagerly waiting for its public release to evaluate the possibilities of using the model.
For more information, Click Here.
