DeepSeek presented DeepSeek Elastic Compute (DSec), a production sandbox platform for large-scale agent training and evaluation with large language models. It provides four execution backends—FnCall, containers, microVMs and full virtual machines—through one software development kit (SDK), while coordinating sandbox placement and lifecycles across a cluster. Its environments are built from independently versioned layers, and image data is loaded on demand from the Fire-Flyer File System (3FS), a distributed filesystem.

DSec combines memory sharing, resource reclamation and CPU scheduling to run sandboxes at high density. Co-designed with a reinforcement learning (RL) framework, it separates stateful rollout execution from GPU training that can be interrupted, while coordinating sandbox lifecycles to preserve rollout state and reclaim idle resources. The platform also includes measures against agent misbehavior such as reward hacking.

A single production-scale DSec unit spans around 160 nodes and serves about 3 million sandboxes per day. In production, the system supports over 380,000 concurrent sandboxes and sustains over 5,000 sandbox creations per second. The report says its mechanisms reduce environment setup and image distribution overhead, improve memory efficiency, and maintain performance for latency-sensitive tasks under high-density overcommit. The paper was submitted to arXiv on September 19, 2026.

DeepSeek presented DeepSeek Elastic Compute (DSec), a production sandbox platform for large-scale agent training and evaluation with large language models. It provides four execution backends—FnCall, containers, microVMs and full virtual machines—through one software development kit (SDK), while coordinating sandbox placement and lifecycles across a cluster. Its environments are built from independently versioned layers, and image data is loaded on demand from the Fire-Flyer File System (3FS), a distributed filesystem.

DSec combines memory sharing, resource reclamation and CPU scheduling to run sandboxes at high density. Co-designed with a reinforcement learning (RL) framework, it separates stateful rollout execution from GPU training that can be interrupted, while coordinating sandbox lifecycles to preserve rollout state and reclaim idle resources. The platform also includes measures against agent misbehavior such as reward hacking.

A single production-scale DSec unit spans around 160 nodes and serves about 3 million sandboxes per day. In production, the system supports over 380,000 concurrent sandboxes and sustains over 5,000 sandbox creations per second. The report says its mechanisms reduce environment setup and image distribution overhead, improve memory efficiency, and maintain performance for latency-sensitive tasks under high-density overcommit. The paper was submitted to arXiv on September 19, 2026.