DeepSeek recently published a systems paper, revealing the core technical details of its large-scale agent training dedicated sandbox infrastructure for the first time. The paper is titled "DeepSeek Elastic Compute (DSec): An Efficient Sandbox Infrastructure for Large-Scale Agent Training," which is 31 pages long and was submitted on September 19th. Notably, the author team consists of more than 130 people, with Liang Wenfeng, the founder of DeepSeek, listed as the last author.

Self-developed sandbox platform DSec, serving 3 million instances per day per unit

According to the paper, DSec (DeepSeek Elastic Compute) is DeepSeek's self-developed production-grade elastic computing sandbox platform, used as its internal infrastructure. Its core purpose is to provide a secure and stable sandbox environment for large-scale training and evaluation tasks of large model agents, ensuring efficient and stable deployment of massive agent training tasks.

The production environment data disclosed in the paper shows that DSec has strong capabilities for large-scale concurrent processing: a standard DSec production unit consists of 160 CPU nodes, with a total of 30,000 CPU cores and approximately 250TB of memory, capable of managing PB-level image files and environmental layer resources. In daily production scenarios, this unit can serve about 3 million sandbox instances per day, stably supporting more than 380,000 concurrent sandboxes running at the same time, with a sandbox creation rate peak exceeding 5,000 per second.

Reducing resource costs and improving memory efficiency

Through actual deployment and model evaluation experience, several key mechanisms of DSec have reduced the resource costs of setting up training environments and image distribution, and significantly improved overall memory utilization efficiency. This paper's release of this infrastructure allows the outside world to glimpse the underlying engineering capabilities supporting DeepSeek's massive agent training.