Meituan has proudly announced the official release of its latest innovation—the next-generation, trillion-parameter large model, LongCat-2.0. In a bold move, the company has also decided to open-source this cutting-edge technology. LongCat-2.0 has undergone comprehensive end-to-end training and inference processes on a robust, domestically-sourced computing power cluster comprising 50,000 cards. Boasting an impressive total of 1.6 trillion parameters, the model features an average activation of around 48 billion, with a dynamic parameter range fluctuating between 33 billion and 56 billion.

Built from the ground up, LongCat-2.0 is pre-trained to seamlessly support ultra-long contexts, accommodating up to a staggering 1 million tokens. Its architectural design is meticulously crafted to empower the model to comprehend, generate, and execute code with enhanced efficiency and stability in real-world Agentic Coding scenarios.