Meituan Launches LongCat-2.5-Preview Model: 1.6T Parameters, Focused on Long-Horizon AI Tasks and Multimodal Capabilities
Summary
Meituan's LongCat API platform launched its next-generation large model LongCat-2.5-Preview on September 25, emphasizing long-horizon tasks and multimodal capabilities, with API and web access opened simultaneously. The model follows a MoE architecture with roughly 1.6 trillion total parameters and about 48 billion activated per inference, natively supports a 1-million-token context window, and is deeply adapted to mainstream development environments such as Claude Code.
Bubble analysis
This is a model release — a product and technology milestone that carries no funding, capex or revenue figures, so it says nothing directly about whether AI investment is a bubble. At most it indirectly shows that Chinese model developers are still investing in R&D and competition has not cooled.
Related signals
linked by shared tags
- Behind the 4 Trillion Compute Network: What Does the Tech Foundation of 'New Quality Productive Forces' Look Like?#china
- Qineng Technology's Datong Super Energy Complex Tianzhen Computing Park Phase I Tops Out#china
- An education company selling compute? Doushen Education lands an RMB 800 million cloud computing services order#china
- Leading Optical Interconnect Player for Compute: Xihe Technology Discloses Second Round of Regulatory Inquiry#china
- Focus on "1144" | Guyang County: Building a New Industrial Track of "Green Power—Computing Power—Tokens"#china