DeepSeek Releases DSec: Elastic Sandbox Infrastructure for Large-Scale Agentic Training

DeepSeek has introduced DeepSeek Elastic Compute (DSec), a production-grade sandbox infrastructure designed to support large-scale agentic training and evaluation for Large Language Models (LLMs). The system addresses the challenges of isolated, stateful execution environments required for agents to inspect repositories, invoke tools, and interact with services during reinforcement learning (RL) workflows.

Architecture and Capabilities

DSec exposes multiple sandbox backends—FnCall, container, microVM, and full-VM—through a unified SDK. It coordinates placement and lifecycle management across the cluster, composing environments from independently versioned layers. Key technical features include:

• High-Density Execution: Combines memory sharing, reclamation, and CPU scheduling to maximize resource utilization.
• On-Demand Image Loading: Loads image data from Fire-Flyer File System (3FS), a cluster-wide distributed filesystem, to handle large image corpora with limited reuse.
• RL Co-Design: Decouples stateful rollout execution from preemptible GPU training, preserving rollout state while reclaiming idle resources.
• Safety Mechanisms: Mitigates agent misbehaviors such as reward hacking.

Production Scale and Performance

A single production-scale unit of DSec spans approximately 160 nodes and serves about 3 million sandboxes per day. In live production environments, the platform supports over 380,000 concurrent sandboxes and sustains creation rates exceeding 5,000 sandboxes per second. Evaluation results indicate that DSec reduces environment setup and image-distribution overhead, improves memory efficiency, and maintains latency-sensitive performance under high-density overcommit conditions.