Meituan Quietly Launches LongCat-2.0-Preview, Sets China Record for Domestic Compute Training Scale

Meituan has invested in large language model research and development for years. LongCat-2.0-Preview builds on its LongCat technology roadmap, focusing on planning, tool use and long-running task execution for AI agents. The release is significant because it advances model capabilities while validating China's domestic computing supply chain at scale.

Meituan recently and quietly launched LongCat-2.0-Preview, which uses a trillion-parameter MoE architecture and supports million-token context processing. Training used as many as approximately 60,000 domestically produced accelerator cards throughout the process, setting a record for the largest training run of a Chinese large model using domestic computing hardware.