**[Meituan LongCat-2.5 Preview Version Launched: Native Multimodal, Focused on Long-Range Agent Tasks]** Beating AI News Flash: Meituan has launched LongCat-2.5-Preview. The model continues the specifications of LongCat-2.0, featuring 1.6T total parameters, approximately 48B active parameters, and a 1M token context window. However, it introduces native multimodal capabilities, enabling execution of long-process tasks across scenarios such as terminals, browsers, desktop software, spreadsheets, and design tools. LongCat-2.5-Preview is now integrated with the official API and is compatible with OpenAI and Anthropic interfaces, supporting a maximum output length of 128K tokens. The official documentation also provides integration methods for Codex, OpenCode, OpenClaw, and CatPaw, allowing existing Agent tools to directly invoke the model. Meituan is offering 5 million free tokens to all existing users for experiencing the new model, and previously purchased token packages remain usable. However, the official benchmark for LongCat-2.5 has not yet been released, and there is currently no publicly available data to compare improvements in GUI operations and long-process tasks. [Original Link]