Agentic training of large language models requires developers to run models in isolated, stateful environments where they can inspect code repositories, invoke tools, execute commands, and interact with task‑specific services. These workloads generate sandboxes in large bursts, demand heterogeneous isolation levels, must preserve state across long interactions, and frequently load large image corpora that cannot be reused efficiently. Traditional single‑runtime sandbox approaches cannot elastically scale to meet such variable, high‑volume demands.
DeepSeek Elastic Compute (DSec) addresses this by providing a production sandbox platform that exposes FnCall, container, microVM, and full‑VM backends through a unified software development kit. DSec coordinates the placement and lifecycle of sandboxes across a cluster, builds environments from independently versioned software layers, combines memory sharing, reclamation, and CPU scheduling for high‑density execution, and loads image data on demand from the Fire‑Flyer File System (3FS), a distributed filesystem shared cluster‑wide.
A single scale unit of DSec spans roughly 160 CPU nodes with about 30K cores and 250TB of DRAM, managing petabytes of layers and images. In operation it serves around 3 million sandbox instances per day, maintains a peak concurrency of approximately 380000 sandboxes, and sustains a creation rate exceeding 5000 instances per second. These capabilities reduce environment‑setup and image‑distribution overhead, improve memory efficiency, and preserve latency‑sensitive performance even when the system is overcommitted beyond its nominal capacity.
Because DSec is co‑designed with the reinforcement‑learning framework that drives agentic training, it decouples stateful rollout execution from preemptible GPU training, allowing the system to preserve rollout state while reclaiming idle resources. The platform also incorporates mitigations for agent misbehavior such as reward hacking, where an agent might exploit loopholes in its reward function rather than solving the intended task.