Following the launch of the LongCat - Flash series models by Meituan and the open - sourcing of two major versions, the LongCat team has now unveiled a new addition to the family: LongCat - Flash - Omni. This model is constructed on a highly efficient architecture and leverages ScMoE technology. It seamlessly integrates multimodal perception and speech reconstruction modules. Although it boasts a colossal parameter scale of up to 560 billion (with 27 billion parameters activated), it still manages to deliver low - latency real - time audio - video interaction.

In full - modality benchmark tests, this model has attained the cutting - edge level in the open - source realm, achieving “full - modality intelligence without any performance degradation.” It adopts an integrated full - modality architecture and an end - to - end design, and also introduces a progressive early multimodal fusion training strategy. Performance tests clearly show that the model performs exceptionally well across multiple domains. It has significantly enhanced capabilities in text and image comprehension, along with state - of - the - art audio - video processing abilities.

Users can get a taste of some of its functions through the official website. The official App has already been rolled out, supporting internet search and voice calls. Video call functionality is set to be introduced in the near future. The relevant code can be accessed on Hugging Face and Github.