LongCat 2.0

A sparse mixture-of-experts model from Meituan with 48B active parameters designed to handle very long contexts up to 1M tokens.

Strengths

Efficiently processes long documents and extended code repositories through sparse routing, enabling coherent reasoning across large codebases and multi-step problem-solving tasks.

Best for

Coding tasks, repository-wide refactoring, long-horizon reasoning, and agentic workflows that require understanding across thousands of lines of context.

Limitations

Sparse mixtures may produce less consistent results on tasks where dense model computation is critical; performance on very short contexts may not match dense baselines.