Meituan releases multimodal LongCat-2.5 with 1.6 trillion parameters
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Another major Chinese model combines vision, long context, and agent work in familiar APIs, giving developers an alternative to US services.
According to Meituan's announcement, the model is designed for long-horizon agent tasks and combines text and visual understanding with coding, including working from mockups and creating 3D interfaces. About 48 billion of its 1.6 trillion total parameters are activated per token. The context window is 1 million tokens, with a maximum response of 128,000. Access is available through APIs compatible with OpenAI and Anthropic formats, allowing the model to be connected to several existing coding agents by changing the endpoint and key. The company links the transition from the preview to reinforcement learning on long action sequences; no independent comparison results were available when the ExplainX review was published.
Sources 1