On July 27, Meituan officially launched the CatPaw, an AI Agent platform powered by LongCat 2.0, marking its transition from technical open-source to industrial application for its self-developed trillion-parameter large model. As the newly open-sourced 1.6T total parameter model with native support for 1M ultra-long context, LongCat 2.0 is the first trillion-parameter model to complete full-process training and inference on 50,000 domestic computing cards.

CatPaw further transforms this underlying computing power and model capability into an enterprise-level intelligent workbench and agent development and management platform that is ready to use out of the box.

QQ20260727-163531.jpg

CatPaw deeply integrates Meituan's full-chain cognition in the local lifestyle sector, packaging expert and skill modules that are ready to use for scenarios such as store operations, review diagnosis, marketing conversion, and fulfillment delivery. The platform supports seamless collaboration between mobile and PC devices, 7×24-hour cloud-based long-term operation, cross-session long-term memory, and autonomous planning and execution capabilities, enabling multi-Agent collaborative processing of complex tasks.

QQ20260727-163628.jpg

In the enterprise segment, CatPaw Managed Agents provides an engineering foundation with strictly isolated operating environments, supporting cold start at millisecond level and dynamic scaling, and pre-installing multiple IM digital employee role templates. Currently, CatPaw has covered 90,000 employees within Meituan and successfully built over 30,000 agents.

Through the comprehensive launch from LongCat 2.0 open source to CatPaw, Meituan has completed the evolution path from "self-developed large models to all-scenario AI applications." This not only reflects the shift of large models from "being able to run" to "being effectively used," but also provides a scalable and practical reference model for the digital and intelligent upgrade of the local lifestyle industry.