LongCat-Flash-Omni: A SOTA Open-Source Omni-Modal Model with 560B Parameters with 27B activated, Excelling at Real-Time Audio-Visual Interaction

Meituans LongCat team has developed LongCat Flash Omni, an open-source, omni-modal model featuring 560 billion parameters and 27 billion active per token, designed for real-time multi-modal interaction across text, images, video, and audio. Built on a shortcut-connected Mixture of Experts architecture, the model extends traditional language models by integrating perception modules such as a LongCat ViT encoder for vision and video, and an audio encoder with a dedicated Audio Codec, enabling seamless audio-visual communication and long-context understanding with a 128K token window. The architecture maintains the core language model