DeepSeek details DSec, the sandbox platform behind its agent training: about 3 million sandboxes a day per unit
A DeepSeek technical report on arXiv describes DeepSeek Elastic Compute (DSec), its platform for large-scale agent training and evaluation. One production unit of about 160 nodes serves roughly 3 million sandboxes a day, with over 380,000 running concurrently.
DeepSeek has published a 31-page technical report on arXiv describing DeepSeek Elastic Compute (DSec), the internal sandbox platform behind its large-scale agent reinforcement learning. The report was submitted on September 19 and drew a Hacker News discussion this week.
The facts
- What it does: agent training needs huge numbers of isolated, stateful environments where models inspect repos, call tools and run commands. DSec exposes four sandbox backends through one SDK: function calls, containers, microVMs and full VMs.
- Scale: the report says a production unit of about 160 nodes serves roughly 3 million sandboxes a day, sustains over 380,000 concurrently, and creates more than 5,000 per second.
- Key design choices: environments are composed from independently versioned layers, image data loads on demand from DeepSeek’s 3FS distributed filesystem, and sandbox execution is decoupled from preemptible GPU training so rollout state survives while idle resources are reclaimed.
- Safety: DSec is co-designed with the RL framework, partly to curb agent misbehavior such as reward hacking.
Our take
The report’s value is that it spells out the environment layer of agent training. Beyond model quality, whoever can spin up millions of trustworthy sandboxes cheaply and fast can run more RL. Set against OpenAI’s sandbox-escape incidents this week, isolation and scheduling are also safety problems.
DSec itself isn’t open source. What you can borrow is the architecture: layered images, on-demand loading, and execution decoupled from training. If you’re building agent evals or an RL pipeline, it’s a useful design reference.