LongCat-2.5

Meituan · proprietary · longcat-2.5

LongCat-2.5 is Meituan's native multimodal Mixture-of-Experts model for long-horizon agentic tasks, with about 1.6T total parameters and roughly 48B activated per token. It builds on LongCat-2.0 with stronger native visual understanding and agentic coding, representing images as multi-level discrete tokens (RVQ over native-resolution features) so the backbone processes visual and text inputs in one token format. Training adds more long-horizon trajectories, large-scale asynchronous agentic RL across mixed harnesses, and multi-teacher on-policy distillation. Meituan also reports optimizing long-context training and W4A4 inference for domestic AI accelerators. The LongCat API (OpenAI, OpenAI Responses and Anthropic compatible) exposes it as LongCat-2.5 with a 1M-token context window; Meituan's Quick Start and FAQ list a 128K maximum output, while the Model Parameters page allows max_tokens up to 262,144. Launch pricing is a limited-time discount of $0.30 / $1.20 per MTok (cached input $0.006). Weights were not public at announcement.