LongCat-Flash-Omni Technical Report
Abstract
We introduce LongCat-Flash-Omni, a state-of-the-art open-source omni-modal model with 560 billion parameters, excelling at real-time audio-visual interaction. By adopting a curriculum-inspired progressive training strategy that transitions from simpler to increasingly complex modality sequence modeling tasks, LongCat-Flash-Omni attains comprehensive multimodal capabilities while maintaining strong unimodal capability. Building upon LongCat-Flash, which adopts a high-performance Shortcut-connected Mixture-of-Experts (MoE) architecture with zero-computation experts, LongCat-Flash-Omni integrates efficient multimodal perception and speech reconstruction modules. Despite its immense size of 560B parameters (with 27B activated), LongCat-Flash-Omni achieves low-latency real-time audio-visual interaction. For training infrastructure, we developed a modality-decoupled parallelism scheme specifically designed to manage the data and model heterogeneity inherent in large-scale multimodal training. This innovative approach demonstrates exceptional efficiency by sustaining over 90% of the throughput achieved by text-only training. Extensive evaluations show that LongCat-Flash-Omni achieves state-of-the-art performance on omni-modal benchmarks among open-source models. Furthermore, it delivers highly competitive results across a wide range of modality-specific tasks, including text, image, and video understanding, as well as audio understanding and generation. We provide a comprehensive overview of the model architecture design, training procedures, and data strategies, and open-source the model to foster future research and development in the community.
Cite
@article{arxiv.2511.00279,
title = {LongCat-Flash-Omni Technical Report},
author = {Meituan LongCat Team and Bairui Wang and Bayan and Bin Xiao and Bo Zhang and Bolin Rong and Borun Chen and Chang Wan and Chao Zhang and Chen Huang and Chen Chen and Chen Chen and Chengxu Yang and Chengzuo Yang and Cong Han and Dandan Peng and Delian Ruan and Detai Xin and Disong Wang and Dongchao Yang and Fanfan Liu and Fengjiao Chen and Fengyu Yang and Gan Dong and Gang Huang and Gang Xu and Guanglu Wan and Guoqiang Tan and Guoqiao Yu and Haibo Qiu and Hao Lu and Hongbo Liu and Hongyu Xiang and Jiaheng Wu and Jian Yang and Jiaxing Liu and Jing Huang and Jingang Wang and Jinrui Ding and Juchao Jiang and Jun Kuang and Jun Wang and Junhui Mei and Ke Ding and Kefeng Zhang and Lei Chen and Liang Shi and Limeng Qiao and Liming Zheng and Lin Ma and Liuyang Guo and Liya Ma and Luying Sun and Man Gao and Mengshen Zhu and Miao Cao and Minliang Lin and Nuo Xu and Peng Shi and Qi Zhang and Qian Fang and Qian Wang and Qian Yang and Quanxiu Wang and Rongxiang Weng and Rongxin Guo and Ruoxuan Liang and Senbin Yang and Shanbo Xu and Shanglin Lei and Shengze Ye and Shimin Chen and Shuaiqi Chen and Shujie Hu and Shuo Li and Siqi Yang and Siyu Xu and Siyu Ren and Song Li and Songxiang Liu and Tianhao Bai and Tianye Dai and Wei Hong and Wei Wang and Weixiao Zhao and Wengang Cao and Wenlong Zhu and Wenlong He and Xi Su and Xi Nan and Xiaohan Zhao and Xiaohao Wang and Xiaoyu Zhao and Xiaoyu Wang and Xiaoyu Li and Xin Pan and Xin Chen and Xiusong Sun and Xu Xiang and Xudong Xing and Xuezhi Cao and Xunliang Cai and Yang Yang and Yanli Tan and Yao Yao and Yerui Sun and Yi Chen and Yifan Lu and Yin Gong and Yining Zhang and Yitian Chen and Yiyang Gan and Yuchen Tang and Yuchen Xie and Yueqian Wang and Yuewen Zheng and Yufei Zhang and Yufeng Zhong and Yulei Qian and Yuqi Peng and Yuqian Li and Yuwei Jiang and Zeyang Hu and Zheng Zhang and Zhengkun Tian and Zhiqing Hong and Zhixiong Zeng and Zhuqi Mi and Ziran Li and Ziwen Wang and Ziyi Zhao and Ziyuan Zhuang and Zizhe Zhao},
journal= {arXiv preprint arXiv:2511.00279},
year = {2025}
}