According to the report submitted to arXiv on 19 September 2026, DeepSeek Elastic Compute (DSec) is a production‑grade sandbox infrastructure designed to support large‑scale agentic training and evaluation of large language models. The platform exposes four sandbox backends—FnCall, container, microVM, and full‑VM—through a single software development kit, allowing developers to select the isolation level that best matches their workload. DSec coordinates the placement and lifecycle of these sandboxes across a cluster, composes environments from independently versioned software layers, and combines memory sharing, reclamation, and CPU scheduling to enable high‑density execution. Image data required by the sandboxes is loaded on demand from the Fire‑Flyer File System (3FS), a cluster‑wide distributed filesystem that minimizes redundant storage and network traffic. The system is co‑designed with the reinforcement‑learning framework used for training, which decouples stateful rollout execution from preemptible GPU training. By synchronizing sandbox lifecycle with the training process, DSec can preserve the state of active rollouts while reclaiming resources from idle sandboxes, thereby improving overall utilization. The paper notes that DSec also incorporates mechanisms to mitigate agent misbehavior, such as reward hacking, by monitoring and restricting sandbox actions during rollout evaluation. In terms of scale, a single production unit of DSec spans roughly 160 nodes. It serves about three million sandboxes per day, maintains over 380,000 concurrent sandboxes, and sustains a creation rate exceeding 5,000 sandboxes per second. Evaluation and deployment experience indicate that these architectural choices reduce the overhead associated with environment setup and image distribution, improve memory efficiency through sharing and reclamation, and preserve latency‑sensitive performance even when the system is overcommitted. The report concludes that DSec provides a flexible, elastic execution platform that addresses the unique demands of agentic LLM training at scale.
