Meituan Officially Open-Sources LongCat-2.0: A 1.6T Parameter Model Revolutionizing Agentic Coding and Domestic Chip Inference
Meituan's technical team has announced the open-source release of LongCat-2.0, a high-performance model featuring 1.6 trillion parameters with an average activation of 48 billion. Designed specifically for complex Agentic Coding tasks, LongCat-2.0 introduces architectural breakthroughs including LongCat Sparse Attention and N-gram Embedding. These innovations are engineered to optimize long-context processing and token-level representation. Crucially, the release includes synchronized inference code for domestic hardware, facilitating broader adoption within the local ecosystem. By utilizing dynamic activation, the model achieves significant improvements in code comprehension, generation, and execution, positioning it as a specialized tool for the next generation of AI-driven software development.
