Meituan Open-Sources LongCat-2.5 Multimodal Model
Meituan’s LongCat team open-sourced LongCat-2.5, a native multimodal model with 1.6 trillion total parameters, 48 billion activated parameters per token, and a 1 million-token context window. The MIT-licensed release includes model weights, a W4A4 quantized version, and deployment guidance based on SGLang-FluentLLM for domestic AI accelerators. Meituan’s evaluation placed the model first in four of 31 comparisons, while OpenCode data cited by Zhidx showed about 109 billion tokens used across roughly 4,700 users and 50,000 sessions.
The MIT release provides developers with weights and deployment options for a 1.6-trillion-parameter multimodal model, although its leading benchmark results come from Meituan’s own evaluation.
智东西 作者 | 孙小雯 编辑 | 程茜
智东西11月10日报道，昨晚，美团LongCat团队开源原生多模态模型LongCat-2.5。据OpenCode公开数据，截至今日上午12时，LongCat-2.5在该平台的累计Token用量约1090亿，独立用户约4700名，已完成约5万次会话。
▲LongCat-2.5在OpenCode上的公开数据（图源：OpenCode）
LongCat-2.5采用MIT协议开放这款1.6万亿参数模型，支持商用推理部署与二次训练，无需额外授权。