[Submitted on 19 Sep 2026]
Authors:Jialiang Huang, Hongxuan Tang, Jingchang Chen, Yuxuan Liu, Yixiao Chen, Yuan Cheng, Yi Tao, Jingli Zhou, Yupeng Chen, Haoyu Chen, Jiarui Wang, Shengkai Lin, Chuqi Zhang, Bryan Lee Teng, Lian Guo, Zhe Fu, Wenjun Gao, Yisong Wang, Liang Zhao, Zehao Wang, Ziwei Xie, Yongqiang Guo, Peixin Cong, Ziyi Gao, Shuiping Yu, Hanwei Xu, Zuofan Wu, Zhizhou Ren, Yuyang Zhou, Bowei Zhang, Zhihuan Huang, Qihao Zhu, Lei Wang, Tianle Lin, Han Yu, Jiewen Hu, Dejian Yang, Shuo Yang, Shanghao Lu, Shaoyuan Chen, Junjie Qiu, Zhangli Sha, Yinmin Zhong, Yongtong Wu, Shiyu Wang, Wei Liu, Bingzheng Xu, Longhao Chen, Qiushi Du, Yuzhen Huang, Shirong Ma, Yaohui Wang, Mingshu Chen, Tongrui Xiong, Y.C. Yan, Haowen Luo, Haofen Liang, Xiaokang Zhang, Weihao Zeng, Runxin Xu, Peiyi Wang, Jinhua Zhu, Ruoyu Zhang, Wenkai Yang, Qi Tang, Jiping Yu, Tian Ye, Ruizhe Pan, Honghui Ding, Xiaodong Liu, Lingxiao Luo, Zhihong Shao, Yuhan Wu, Jibai Lu, Wen Liu, Haoling Zhang, Jingcheng Hu, Yaoyang Ye, Chaofan Lin, Zhaochen Zhang, Jianan Tong, Hengxu Wu, Zhihao Li, Yicheng Wang, Luyao Wang, Yuzhuo Bai, Lingyue Fu, Ruifan Xu, Y.Z. Wang, Zonglin Li, Mingqi Wei, Haiyang Shen, Chengyuan Zhang, Chao Jin, Zili Zhang, R.H. Yang, Xinbo Xu, Jian Zhou, Ruidong Zhu, Yuzhe Guo
et al. (31 additional authors not shown)
Abstract:Large-scale agentic training and evaluation alongside ample tongue models (LLMs) depend on isolated, stateful implementation environments in which models inspect repositories, invoke tools, execute commands, and engage alongside task-specific services. These workloads create sandboxes in ample bursts, extend heterogeneous functionality and isolation requirements, keep province throughout lengthy interactions, and diagram from ample depiction corpora alongside constricted reuse. Supporting them hence requires an elastic implementation phase fairly than a sole sandbox runtime.
This study presents DeepSeek Elastic Compute (DSec), a manufacturing sandbox phase that exposes FnCall, container, microVM, and full-VM sandbox backends through a unified SDK. DSec coordinates placement and lifecycle administration throughout the cluster, composes environments from independently versioned layers, combines recollection sharing, reclamation, and CPU scheduling for high-density execution, and loads depiction data on petition from Fire-Flyer File System (3FS), a cluster-wide distributed filesystem. DSec is co-designed alongside the reinforcement learning (RL) framework, decouples stateful rollout implementation from preemptible GPU training, coordinates sandbox lifecycle alongside training to maintain rollout province during reclaiming idle resources, and mitigates delegate misbehavior specified as reward hacking.
A sole production-scale component of DSec spans about 160 nodes, serving concerning 3 myriad sandboxes per day; in production, it supports complete 380,000 concurrent sandboxes and sustains complete 5,000 sandbox creations per second. Our evaluation and deployment cognition display that these mechanisms decrease surroundings setup and image-distribution overhead, enhance recollection efficiency, and maintain latency-sensitive achievement under high-density overcommit.
Submission history
From: Wenfeng Liang [view email]
[v1] Sat, 19 Sep 2026 12:20:26 UTC (504 KB)
