One-sentence summary
Meituan on November 2 launched LongCat-2.5, a native multimodal model for long-horizon agent tasks: 1.6T total parameters, ~48B activated per token, 1M-token context, fully trained on domestic Chinese compute.
Event Overview
On November 2, 2026, Meituan launched LongCat-2.5, its next-generation LLM positioned as a native multimodal model for long-horizon agent tasks, billed as a new milestone in domestic chips running domestic models.
Key Information
- 1.6T total parameters in a MoE architecture, with about 48B activated per token.
- Native multimodal, supporting 1M-token context and image understanding, strengthening native vision and agentic coding.
- Builds on LongCat-2.0, with training completed entirely on domestic Chinese compute.
- LongCat-2.5-Preview is live, targeting long-horizon tasks across browsers, terminals, desktop software, spreadsheets and design tools.