On November 3, Meituan made an official announcement regarding the launch and open-sourcing of its LongCat-Flash-Omni model. Constructed on the foundation of the efficient architectural design inherent in the LongCat-Flash series—which features Shortcut-Connected Mixture of Experts (MoE) and incorporates zero-computation experts—this model ingeniously combines efficient multimodal perception modules with a speech reconstruction module.

Despite boasting an impressive total of 560 billion parameters (with 27 billion active parameters), LongCat-Flash-Omni still manages to deliver low-latency, real-time audio-video interaction capabilities. This provides developers with a more streamlined and efficient solution for tackling multimodal application scenarios. The model has attained the state-of-the-art (SOTA) status among open-source models in full-modality benchmark tests. Moreover, it exhibits robust competitiveness in pivotal single-modality tasks, including text, image, and video comprehension, as well as speech perception and generation.

In conjunction with this release, Meituan has also rolled out the official LongCat App. This app comes equipped with internet search capabilities and voice call functionalities, further enhancing its utility and appeal.