Ling与Ring 2.6技术报告：万亿参数规模下的高效即时Agentic智能

# Ling和Ring 2.6技术报告：高效即时的万亿参数规模智能体智能
来源：https://arxiv.org/abs/2606.15079
作者：Ang Li (https://arxiv.org/search/cs?searchtype=author&query=Li,+A),Ben Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+B),Bin Han (https://arxiv.org/search/cs?searchtype=author&query=Han,+B),Bin Hu (https://arxiv.org/search/cs?searchtype=author&query=Hu,+B),Bin Jing (https://arxiv.org/search/cs?searchtype=author&query=Jing,+B),Binbin Hu (https://arxiv.org/search/cs?searchtype=author&query=Hu,+B),Bing Li (https://arxiv.org/search/cs?searchtype=author&query=Li,+B),Cai Chen (https://arxiv.org/search/cs?searchtype=author&query=Chen,+C),Caizhi Tang (https://arxiv.org/search/cs?searchtype=author&query=Tang,+C),Changxin Tian (https://arxiv.org/search/cs?searchtype=author&query=Tian,+C),Chao Huang (https://arxiv.org/search/cs?searchtype=author&query=Huang,+C),Chao Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+C),Chen Liang (https://arxiv.org/search/cs?searchtype=author&query=Liang,+C),Chen Qian (https://arxiv.org/search/cs?searchtype=author&query=Qian,+C),Chengfu Tang (https://arxiv.org/search/cs?searchtype=author&query=Tang,+C),Chengyao Wen (https://arxiv.org/search/cs?searchtype=author&query=Wen,+C),Chilin Fu (https://arxiv.org/search/cs?searchtype=author&query=Fu,+C),Chunwei Wu (https://arxiv.org/search/cs?searchtype=author&query=Wu,+C),Cong Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+C),Cunyin Peng (https://arxiv.org/search/cs?searchtype=author&query=Peng,+C),Daixin Wang (https://arxiv.org/search/cs?searchtype=author&query=Wang,+D),Dalong Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+D),Deng Zhao (https://arxiv.org/search/cs?searchtype=author&query=Zhao,+D),Dingnan Jin (https://arxiv.org/search/cs?searchtype=author&query=Jin,+D),Dingyuan Zhu (https://arxiv.org/search/cs?searchtype=author&query=Zhu,+D),Donghao Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+D),Fan Yuan (https://arxiv.org/search/cs?searchtype=author&query=Yuan,+F),Fangzheng Zhao (https://arxiv.org/search/cs?searchtype=author&query=Zhao,+F),Fanzhuang Meng (https://arxiv.org/search/cs?searchtype=author&query=Meng,+F),Feifan Wu (https://arxiv.org/search/cs?searchtype=author&query=Wu,+F),Feng Xu (https://arxiv.org/search/cs?searchtype=author&query=Xu,+F),Fengbin Fang (https://arxiv.org/search/cs?searchtype=author&query=Fang,+F),Gangshan Wang (https://arxiv.org/search/cs?searchtype=author&query=Wang,+G),Guodong Yang (https://arxiv.org/search/cs?searchtype=author&query=Yang,+G),Hailin Zhao (https://arxiv.org/search/cs?searchtype=author&query=Zhao,+H),Haitao Wang (https://arxiv.org/search/cs?searchtype=author&query=Wang,+H),Haitao Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+H),Hanxiao Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+H),Hanzi Wang (https://arxiv.org/search/cs?searchtype=author&query=Wang,+H),Hao Dai (https://arxiv.org/search/cs?searchtype=author&query=Dai,+H),Hao Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+H),Hao Qian (https://arxiv.org/search/cs?searchtype=author&query=Qian,+H),Hao Wu (https://arxiv.org/search/cs?searchtype=author&query=Wu,+H),Haoxiong Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+H),Haoyu Xu (https://arxiv.org/search/cs?searchtype=author&query=Xu,+H),Heng Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+H),Hong Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+H),Hongliang Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+H),Hongrui Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+H),Hongxun Li (https://arxiv.org/search/cs?searchtype=author&query=Li,+H),Hongzhi Ruan (https://arxiv.org/search/cs?searchtype=author&query=Ruan,+H),Huaidong Xiong (https://arxiv.org/search/cs?searchtype=author&query=Xiong,+H),Huihuang Zheng (https://arxiv.org/search/cs?searchtype=author&query=Zheng,+H),Huikang Tang (https://arxiv.org/search/cs?searchtype=author&query=Tang,+H),Jia Guo (https://arxiv.org/search/cs?searchtype=author&query=Guo,+J),Jia Li (https://arxiv.org/search/cs?searchtype=author&query=Li,+J),Jia Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+J),Jiameng Wang (https://arxiv.org/search/cs?searchtype=author&query=Wang,+J),Jiaming Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+J),Jiannan Shi (https://arxiv.org/search/cs?searchtype=author&query=Shi,+J),Jianping Wei (https://arxiv.org/search/cs?searchtype=author&query=Wei,+J),Jiaolong Yang (https://arxiv.org/search/cs?searchtype=author&query=Yang,+J),Jiapeng Wang (https://arxiv.org/search/cs?searchtype=author&query=Wang,+J),Jie Gao (https://arxiv.org/search/cs?searchtype=author&query=Gao,+J),Jie Wang (https://arxiv.org/search/cs?searchtype=author&query=Wang,+J),Jiewei Wu (https://arxiv.org/search/cs?searchtype=author&query=Wu,+J),Jin Yang (https://arxiv.org/search/cs?searchtype=author&query=Yang,+J),Jinjin Li (https://arxiv.org/search/cs?searchtype=author&query=Li,+J),Jinjing Huang (https://arxiv.org/search/cs?searchtype=author&query=Huang,+J),Jinquan Sun (https://arxiv.org/search/cs?searchtype=author&query=Sun,+J),Jinyao Chen (https://arxiv.org/search/cs?searchtype=author&query=Chen,+J),Juanhui Tu (https://arxiv.org/search/cs?searchtype=author&query=Tu,+J),Jun Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+J),Jun Mei (https://arxiv.org/search/cs?searchtype=author&query=Mei,+J),Jun Xu (https://arxiv.org/search/cs?searchtype=author&query=Xu,+J),Jun Zhou (https://arxiv.org/search/cs?searchtype=author&query=Zhou,+J),Junjie Ou (https://arxiv.org/search/cs?searchtype=author&query=Ou,+J),Junnan Sipan (https://arxiv.org/search/cs?searchtype=author&query=Sipan,+J),Junpeng Fang (https://arxiv.org/search/cs?searchtype=author&query=Fang,+J),Kaihong Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+K),Kaiqin Hu (https://arxiv.org/search/cs?searchtype=author&query=Hu,+K),Ke Shi (https://arxiv.org/search/cs?searchtype=author&query=Shi,+K),Kuan Xu (https://arxiv.org/search/cs?searchtype=author&query=Xu,+K),Kun Tang (https://arxiv.org/search/cs?searchtype=author&query=Tang,+K),Kunlong Chen (https://arxiv.org/search/cs?searchtype=author&query=Chen,+K),Lanyin Mei (https://arxiv.org/search/cs?searchtype=author&query=Mei,+L),Lei Chen (https://arxiv.org/search/cs?searchtype=author&query=Chen,+L),Lei Liang (https://arxiv.org/search/cs?searchtype=author&query=Liang,+L),Lei Xu (https://arxiv.org/search/cs?searchtype=author&query=Xu,+L),Li Tang (https://arxiv.org/search/cs?searchtype=author&query=Tang,+L),Liang Jiang (https://arxiv.org/search/cs?searchtype=author&query=Jiang,+L),Liangcheng Fu (https://arxiv.org/search/cs?searchtype=author&query=Fu,+L),Lihui Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+L),Linfeng Shi (https://arxiv.org/search/cs?searchtype=author&query=Shi,+L),Lintao Ma (https://arxiv.org/search/cs?searchtype=author&query=Ma,+L),Liyuan Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+L),Longfei Li (https://arxiv.org/search/cs?searchtype=author&query=Li,+L),Longfei Zheng (https://arxiv.org/search/cs?searchtype=author&query=Zheng,+L),Lu Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+L),Lu Yu (https://arxiv.org/search/cs?searchtype=author&query=Yu,+L)
等（另有118位作者未列出）

查看 PDF (https://arxiv.org/pdf/2606.15079)

> 摘要：高效且可扩展的智能体智能要求模型既能提供低延迟响应，又能具备强大的推理能力，同时在训练、服务和部署方面保持实用性。在本报告中，我们介绍了Ling-2.6和Ring-2.6，这是一系列旨在大规模解决这一挑战的模型。Ling-2.6针对即时响应生成和每个输出token的高能力进行了优化，而Ring-2.6则专为深度推理和更高级的智能体工作流而设计。我们没有从零开始训练，而是通过架构迁移预训练和大规模后训练对Ling-2.0基础模型进行升级。这一升级基于模型架构、优化目标、服务系统和智能体训练环境的统一协同设计，从而提升模型能力和部署效率。在架构层面，我们引入了混合线性注意力设计，将Lightning Attention与MLA相结合，提高了长上下文训练和解码的效率。为了进一步提升token效率，我们通过进化思维链、语言单元策略优化、双向偏好对齐和最短正确响应蒸馏来优化每个输出token的能力。在智能体能力方面，我们提出了KPop，这是一个强化学习框架，旨在支持在大型环境接地数据上稳定训练Ring-2.6-1T。KPop通过在编码、搜索、工具使用和工作流执行之间进行异步调度来提高训练效率，从而能够从复杂的智能体-环境交互中进行可扩展学习。Ling-2.6和Ring-2.6共同为构建高效、可扩展且开放的智能体系统提供了实用路径。我们开源了2.6系列的所有检查点，以支持实用智能体智能领域的进一步研究和发展。

## 提交历史

来自：Yuxin Tian [查看邮件 (https://arxiv.org/show-email/bc333059/2606.15079)] **\[v1\]** 2026年6月13日星期六 03:21:49 UTC（4,866 KB）