Meituan’s LongCat update combines visual input with coding agents
Meituan has released LongCat-2.5 for software agents working across extended tasks and visual inputs. The developer describes training with longer task histories, including reproducible examples drawn from failed attempts. Chat and programming-interface links accompany the release, while downloadable distribution through GitHub and Hugging Face remains a later step in the announcement.
Artificial Intelligence··Midday
Images and text in the same model
Chinese technology company Meituan has introduced LongCat-2.5, a model intended for agents that use tools and carry out extended software tasks. IT Home’s account of the release describes images being converted into several levels of discrete tokens, the units processed by the model. This allows its main network to handle visual information and text in a common format. The developer identifies visual coding, creative design and three-dimensional work as target uses. The new model builds on LongCat-2.0.[1]
Failed attempts become training tasks
The team says its training includes longer sequences of agent actions, with work directed at complex repositories and building whole projects. Failed attempts by working agents are turned into tasks that can be rerun and checked. Generated training material, constructed task environments and practice across different agent frameworks supplement that process. The release also describes reinforcement learning over lengthy tasks, alongside changes intended to keep training and inference numerically consistent.[1]
Chat access before downloadable distribution
Access links point users to a chat service and a programming interface through which applications can call the model. GitHub and Hugging Face, services commonly used to distribute code and model files, are named as forthcoming distribution channels. The release account therefore describes access through hosted services while presenting those downloads as a future step. The model’s engineering focus includes multi-step work rather than only isolated code completions.[1]
