LongCat 2.0 is a sparse mixture-of-experts language model from Meituan that activates 48 billion parameters out of a 1.6 trillion total, giving it very large world knowledge and tool-use capacity while keeping per-request compute closer to a mid-sized model. A published weights repository is hosted under the meituan-longcat organization on Hugging Face, making it one of the larger openly distributed MoE models of its generation. The design is aimed at coding, repository-level changes, long-horizon problem solving, and agentic workflows, where the model can hold large amounts of code, documentation, and tool state in context at once.

In deployment, LongCat 2.0 is served through an OpenRouter-hosted route with a roughly one-million-token context window, and it can be self-hosted via the SGLang serving stack. At the time of the SGLang documentation, support lived on the main branch and required a nightly wheel or rolling nightly Docker image until a tagged release picked it up, with CUDA 12.9 and CUDA 13 wheel indices available. The combination of a very large MoE with active-parameter efficiency, open weights, and a deep context window makes it a practical fit for code-assistant backends, multi-step agent pipelines, and any application that needs sustained reasoning over long, structured inputs.