DeepSeek details DSec sandbox infrastructure for agent training

DeepSeek detailed DeepSeek Elastic Compute (DSec), a sandbox platform for large-scale agent training, in a paper posted to arXiv with founder Liang Wenfeng among its 130-plus authors. The paper states a production-scale DSec unit spans around 160 nodes, supports about 3 million sandboxes per day and more than 380,000 concurrent sandboxes, and can create more than 5,000 sandboxes per second, unifying function-call, container, microVM and full-VM sandboxes while coordinating their lifecycle with reinforcement learning workloads.

A paper posted to arXiv details DeepSeek Elastic Compute, or DSec, a sandbox platform designed for large-scale agent training. DeepSeek founder Liang Wenfeng is listed among the paper’s 130-plus authors. DSec unifies function-call, container, microVM and full-VM sandboxes, while coordinating their lifecycle with reinforcement learning workloads. The paper says a production-scale unit spans around 160 nodes, supports about 3 million sandboxes per day and more than 380,000 concurrent sandboxes, and can create more than 5,000 sandboxes per second. arXiv https://arxiv.org/abs/2609.22978