Meituan releases LongCat-2.5 multimodal model

LongCat describes LongCat-2.5 as a native multimodal model for long-horizon agentic tasks and says it builds on LongCat-2.0 with stronger native visual understanding and agentic coding.[1] The company says the model has 1.6 trillion total parameters, with approximately 48 billion activated per token.[1] Its backbone processes visual and textual inputs in a unified token-based format, according to LongCat.[1] LongCat says combining visual understanding and agentic coding supports visual coding, creative design and 3D tasks.[1]

LongCat says it uses large-scale long-horizon reinforcement learning to strengthen end-to-end engineering across complex codebase development and full-project builds.[1] Its agent-training pipeline generates tasks, interactive environments and verifiable tests, using filtered GitHub pull requests and linked issues as training material.[1] LongCat also says it turns real-world agent failure trajectories into reproducible, verifiable training tasks.[1] These are descriptions of the training approach; the supplied information does not establish how the model performs on a particular team’s codebase or project.[1]

Meituan reports that its FSDP configuration achieved 86% higher per-device throughput than a DDP baseline in long-sequence supervised fine-tuning tasks.[1] For W4A4 inference, Meituan reports 43.3% higher prefill throughput and 42.77% higher decode throughput, alongside 15.91% lower TTFT and 19.72% lower TPOT versus its W8A8 INT8 baseline.[1] LongCat lists Terminal Bench v2.1 scores of 88.4 and 70.8, a Terminal Bench v4.0 score of 38.5, and a Vision2Web score of 78.6.[1]

Meituan released the model weights on Nov. 9; the weights are under the MIT License.[2][3] LongCat says its pay-as-you-go service currently supports LongCat-2.5, and lists uncached input at $0.30, cached input at $0.006 and output at $1.20.[4] The API guide shows that developers can submit screenshots, charts, images and videos.[2] VentureBeat reported that, as of Nov. 9, it had found no publicly disclosed U.S. federal procurement or use restriction specifically naming Meituan or LongCat-2.5.[2]