Meituan has released LongCat-2.5, a native multimodal model built for coding work that stretches across many steps. The release, reported on 2 November 2026, follows a September API preview and is aimed at AI agents that have to keep going inside large codebases — editing complex projects and building complete software, rather than answering one-off prompts.

The model processes text and images in a unified token framework, which Meituan says supports visual coding, creative design and 3D tasks. Training leaned on long sequences of agent actions, including failed attempts turned into reproducible training tasks, plus large-scale reinforcement learning for work inside complex codebases. Reported figures put the model at 1.6 trillion total parameters with about 48 billion activated per token — though that parameter count has been disputed, with one report noting the figures were carried over from the earlier LongCat-2.0 model card, so treat the headline size with caution until Meituan publishes a full model card.

On Meituan's own reported benchmarks, LongCat-2.5 scores 88.4 on Terminal Bench v2.1, 51.8 on AutomationBench and 91.9 on BrowseComp — numbers that have not been independently verified, and the model does not lead every test in its own charts. It is available now through Meituan's online chat service and the LongCat API, with support for agent tools including Claude Code and Codex, while GitHub and Hugging Face releases are planned to follow.

Sources