DeepSeek Details DSec Sandbox Infrastructure for Agent Training

DeepSeek has detailed DSec, or DeepSeek Elastic Compute, a sandbox platform engineered for large-scale artificial intelligence agent training. The platform unifies various sandbox environments, including function-call, container, microVM, and full-VM sandboxes, coordinating their lifecycle management with reinforcement learning workloads.

According to the published paper, a production-scale DSec unit spans approximately 160 nodes. It supports around 3 million sandboxes daily and can handle over 380,000 concurrent sandboxes. The system is capable of provisioning more than 5,000 sandboxes per second.

DeepSeek founder Liang Wenfeng is listed among the paper's more than 130 authors, indicating significant internal development. The infrastructure aims to provide a robust and scalable environment for training complex AI agents.

The technical specifications and architecture of DSec were made public via a posting on the arXiv research repository. This release provides insight into the underlying technology enabling advanced AI agent development and deployment.