Seed1.5-VL 技术报告
计算机视觉与模式识别
2025-05-13 v1 人工智能
摘要
我们介绍 Seed1.5-VL,这是一个面向通用多模态理解与推理的视觉语言基础模型。Seed1.5-VL 由 532 百万参数的视觉编码器和 20B 活跃参数的混合专家 (Mixture-of-Experts, MoE) 大语言模型构成。尽管其相对紧凑的架构,Seed1.5-VL 在广泛的公开 VLM 基准测试和内部评估套件上均表现出强劲性能,在 38 个公开基准测试中实现最先进性能。此外,在 GUI 控制和游戏等智能体中心任务中,Seed1.5-VL 超越了包括 OpenAI CUA 和 Claude 3.7 在内的领先多模态系统。除了视觉和视频理解,Seed1.5-VL 还展现出强大的推理能力,在视觉拼图等多模态推理挑战中尤为有效。我们认为这些能力将使其在多样化任务中发挥更广泛的应用。本报告主要综述了我们在模型设计、数据构建和各阶段训练方面的经验,期望为后续研究提供启发。Seed1.5-VL 现已开放访问,地址为 https://www.volcengine.com/ (Volcano Engine 模型 ID: doubao-1-5-thinking-vision-pro-250428)
引用
@article{arxiv.2505.07062,
title = {Seed1.5-VL Technical Report},
author = {Dong Guo and Faming Wu and Feida Zhu and Fuxing Leng and Guang Shi and Haobin Chen and Haoqi Fan and Jian Wang and Jianyu Jiang and Jiawei Wang and Jingji Chen and Jingjia Huang and Kang Lei and Liping Yuan and Lishu Luo and Pengfei Liu and Qinghao Ye and Rui Qian and Shen Yan and Shixiong Zhao and Shuai Peng and Shuangye Li and Sihang Yuan and Sijin Wu and Tianheng Cheng and Weiwei Liu and Wenqian Wang and Xianhan Zeng and Xiao Liu and Xiaobo Qin and Xiaohan Ding and Xiaojun Xiao and Xiaoying Zhang and Xuanwei Zhang and Xuehan Xiong and Yanghua Peng and Yangrui Chen and Yanwei Li and Yanxu Hu and Yi Lin and Yiyuan Hu and Yiyuan Zhang and Youbin Wu and Yu Li and Yudong Liu and Yue Ling and Yujia Qin and Zanbo Wang and Zhiwu He and Aoxue Zhang and Bairen Yi and Bencheng Liao and Can Huang and Can Zhang and Chaorui Deng and Chaoyi Deng and Cheng Lin and Cheng Yuan and Chenggang Li and Chenhui Gou and Chenwei Lou and Chengzhi Wei and Chundian Liu and Chunyuan Li and Deyao Zhu and Donghong Zhong and Feng Li and Feng Zhang and Gang Wu and Guodong Li and Guohong Xiao and Haibin Lin and Haihua Yang and Haoming Wang and Heng Ji and Hongxiang Hao and Hui Shen and Huixia Li and Jiahao Li and Jialong Wu and Jianhua Zhu and Jianpeng Jiao and Jiashi Feng and Jiaze Chen and Jianhui Duan and Jihao Liu and Jin Zeng and Jingqun Tang and Jingyu Sun and Joya Chen and Jun Long and Junda Feng and Junfeng Zhan and Junjie Fang and Junting Lu and Kai Hua and Kai Liu and Kai Shen and Kaiyuan Zhang and Ke Shen and Ke Wang and Keyu Pan and Kun Zhang and Kunchang Li and Lanxin Li and Lei Li and Lei Shi and Li Han and Liang Xiang and Liangqiang Chen and Lin Chen and Lin Li and Lin Yan and Liying Chi and Longxiang Liu and Mengfei Du and Mingxuan Wang and Ningxin Pan and Peibin Chen and Pengfei Chen and Pengfei Wu and Qingqing Yuan and Qingyao Shuai and Qiuyan Tao and Renjie Zheng and Renrui Zhang and Ru Zhang and Rui Wang and Rui Yang and Rui Zhao and Shaoqiang Xu and Shihao Liang and Shipeng Yan and Shu Zhong and Shuaishuai Cao and Shuangzhi Wu and Shufan Liu and Shuhan Chang and Songhua Cai and Tenglong Ao and Tianhao Yang and Tingting Zhang and Wanjun Zhong and Wei Jia and Wei Weng and Weihao Yu and Wenhao Huang and Wenjia Zhu and Wenli Yang and Wenzhi Wang and Xiang Long and XiangRui Yin and Xiao Li and Xiaolei Zhu and Xiaoying Jia and Xijin Zhang and Xin Liu and Xinchen Zhang and Xinyu Yang and Xiongcai Luo and Xiuli Chen and Xuantong Zhong and Xuefeng Xiao and Xujing Li and Yan Wu and Yawei Wen and Yifan Du and Yihao Zhang and Yining Ye and Yonghui Wu and Yu Liu and Yu Yue and Yufeng Zhou and Yufeng Yuan and Yuhang Xu and Yuhong Yang and Yun Zhang and Yunhao Fang and Yuntao Li and Yurui Ren and Yuwen Xiong and Zehua Hong and Zehua Wang and Zewei Sun and Zeyu Wang and Zhao Cai and Zhaoyue Zha and Zhecheng An and Zhehui Zhao and Zhengzhuo Xu and Zhipeng Chen and Zhiyong Wu and Zhuofan Zheng and Zihao Wang and Zilong Huang and Ziyu Zhu and Zuquan Song},
journal= {arXiv preprint arXiv:2505.07062},
year = {2025}
}