DeepSeek V4 First Reveal of Production-Grade Agent Sandbox DSec: Single Cluster Dispatches Hundreds of Thousands of Concurrent Tasks, Unified Four Heterogeneous Foundations

According to Beating Monitoring, the DeepSeek V4 technical report publicly disclosed for the first time the core infrastructure supporting agent-based post-training and large-scale evaluation, the production-grade elastic computing sandbox DSec (DeepSeek Elastic Compute).

Currently, large model reinforcement learning requires an extremely large code trial-and-error environment. The report reveals that in actual production, a single DSec cluster can schedule hundreds of thousands of concurrent sandboxes simultaneously. The system is written in Rust, interfacing with a self-developed 3FS distributed file system, and breaks through the performance bottleneck of cold-starting massive sandboxes through hierarchical on-demand loading.

In terms of developer experience, DSec uses a unified Python SDK that consolidates four execution bases: function calls, containers, micro virtual machines, and full virtual machines. Switching between them only requires modifying one parameter. To address the common preemption issues in computing clusters, DSec introduces a global trace log: when a task resumes, the system directly “fast-forwards” and replays the cached command execution results, enabling ultra-fast checkpoint resume and avoiding non-idempotent errors caused by repeated execution.