DeepSeek's DSec system runs 3 million AI sandboxes a day
A paper from more than 130 DeepSeek engineers describes a production platform that isolates reinforcement-learning agents at a scale the company says tops 380,000 concurrent sandboxes.
DeepSeek said its new DSec sandbox platform creates 3 million isolated environments a day per production unit, for training AI agents with reinforcement learning.
The system exposes a single SDK across four sandbox backends: lightweight function calls, containers, microVMs and full virtual machines. It manages where each one runs, its memory use and how its image spreads across the cluster, the paper said. That design, DeepSeek said, lets reinforcement-learning frameworks train agents at scale without each backend's usual setup overhead.
In production, DeepSeek said one deployment of about 160 nodes creates 3 million sandboxes a day, sustains more than 380,000 running concurrently and can spin up over 5,000 new ones per second. The paper attributes this to overcommitting resources at high density while still protecting latency for time-sensitive workloads.
DeepSeek submitted the paper to arXiv on September 19, and it reached Hacker News's front page Saturday with 68 points and 14 comments. The paper lists more than 130 authors, led by Jialiang Huang with Wenfeng Liang as corresponding author. DeepSeek has not published independent benchmarks against sandbox systems from other labs, so the figures above are the company's own and may not generalize beyond its infrastructure.