LongCat2.0 was officially open-sourced this month. The 1.6T-parameter model has an average active size of about 48B (dynamic activation range 33B–56B) and natively supports a 1M-token context. It is the first trillion-parameter model to complete end-to-end training and inference on 50,000 domestically made compute cards. Meituan today launched CatPaw, a full-scenario AI Agent platform powered by LongCat2.0, aiming to move the model from "can run" to "use well." CatPaw provides an out-of-the-box

LongCat2.0 was officially open-sourced this month. The 1.6T-parameter model has an average active size of about 48B (dynamic activation range 33B–56B) and natively supports a 1M-token context. It is the first trillion-parameter model to complete end-to-end training and inference on 50,000 domestically made compute cards. Meituan today launched CatPaw, a full-scenario AI Agent platform powered by LongCat2.0, aiming to move the model from "can run" to "use well." CatPaw provides an out-of-the-box AI workstation and hosted enterprise Agent development. The platform is already deployed internally at scale, covering 90k employees and 30k Agents, and has been validated in multiple live business scenarios.