DeepSeek V4 has introduced its production-grade elastic computing sandbox, DSec, designed to enhance agent post-training and evaluation processes. The DSec infrastructure supports large model reinforcement learning by providing a vast code experimentation environment. A single DSec cluster can manage hundreds of thousands of concurrent sandboxes, overcoming cold-start performance issues with hierarchical on-demand loading. Built in Rust and integrated with the 3FS distributed file system, DSec offers a unified Python SDK for seamless switching between execution backends. To improve developer experience, DSec integrates function calls, containers, micro virtual machines, and full virtual machines under one system, requiring only a parameter change for backend transitions. Additionally, DSec addresses task preemption challenges in compute clusters with a global trajectory log, allowing for rapid checkpoint recovery and avoiding errors from redundant execution by replaying cached command results.