Glonce

DeepSeek details DSec, its sandbox platform for agentic training at scale

infrastructureThe Glonce editorsHacker News🔥79Glonce rating79Impact70Novelty72Relevance96Surprise55Players72Credibility88

DeepSeek submitted a paper to arXiv on 19 September 2026 describing DeepSeek Elastic Compute (DSec), a production sandbox platform built to support large-scale agentic training and evaluation of large language models. The paper argues that agentic reinforcement-learning workloads, where a model inspects repositories, calls tools, runs commands and talks to task-specific services, create sandboxes in large bursts, need very different isolation and functionality, must hold state over long interactions, and pull from huge libraries of environment images with little reuse. DeepSeek says a single sandbox runtime cannot handle this and that an elastic platform is required instead. DSec exposes four sandbox backend types (FnCall, container, microVM and full VM) through one SDK, and handles placement and lifecycle management of sandboxes across the cluster. It builds environments from independently versioned layers, combines memory sharing, memory reclamation and CPU scheduling to run sandboxes at high density, and loads image data on demand from Fire-Flyer File System (3FS), DeepSeek's own cluster-wide distributed filesystem. The platform is co-designed with DeepSeek's reinforcement-learning framework: it separates stateful rollout execution (the agent acting in its sandbox) from preemptible GPU training, coordinates sandbox lifecycle with the training loop so rollout state is preserved while idle resources are reclaimed, and includes mechanisms meant to curb agent misbehavior such as reward hacking. On production scale, DeepSeek states that a single production-scale unit of DSec spans around 160 nodes and serves about 3 million sandboxes per day; in production it supports over 380,000 concurrent sandboxes and sustains over 5,000 sandbox creations per second. The authors say their evaluation and deployment experience show these mechanisms cut environment setup and image-distribution overhead, improve memory efficiency, and keep latency-sensitive performance stable even under heavy overcommit. The abstract does not name individual authors, does not give a release date or say whether DSec will be open-sourced, and includes no comparison against other sandbox platforms or baseline numbers.

Key facts

  • DeepSeek published a report on 19 September 2026 describing DSec, a sandbox infrastructure platform for large-scale agentic LLM training and evaluation.
  • DSec exposes FnCall, container, microVM and full-VM sandbox backends through one SDK, and loads images on demand from DeepSeek's 3FS distributed filesystem.
  • A single production-scale unit of DSec spans around 160 nodes and serves about 3 million sandboxes per day.
  • In production, DSec supports over 380,000 concurrent sandboxes and sustains over 5,000 sandbox creations per second.
  • DSec is co-designed with DeepSeek's reinforcement-learning framework, decoupling stateful rollout execution from preemptible GPU training and including mechanisms meant to mitigate reward hacking.

Why it matters

Training LLM agents with reinforcement learning means constantly spinning up isolated environments where the model can act, and doing that at DeepSeek's scale exposes a problem a single sandbox runtime cannot solve: sandboxes are created in huge bursts, need different levels of isolation, must hold state over long interactions, and draw on large, poorly-reused libraries of environment images. DSec is DeepSeek's answer, an elastic platform that ties sandbox lifecycle directly to its RL training loop rather than treating sandboxes as a generic side service.

Who it affects

This is infrastructure for DeepSeek's own agentic training pipeline, most directly relevant to engineers building or scaling reinforcement-learning systems for LLM agents elsewhere, since the report lays out a concrete architecture (unified SDK across four sandbox backends, on-demand image loading from a distributed filesystem, memory and CPU scheduling for high-density execution) that others working on agent RL infrastructure can study.

How to use it

The report describes a production system rather than a released product: it gives no release date, no availability details and no indication of whether DSec will be open-sourced, so there is nothing here to install or adopt directly.

How solid is it

The account comes from a paper submitted to arXiv, not from a peer-reviewed venue, and the abstract does not name any authors or institutions. The scale figures, around 160 nodes and 3 million sandboxes a day per unit, over 380,000 concurrent sandboxes, and over 5,000 sandbox creations per second, are DeepSeek's own reported production numbers rather than independently verified or externally benchmarked figures.

Risks and caveats

The abstract offers no comparison to other sandbox platforms and no baseline performance numbers, so there is no independent way to judge how DSec's figures stack up against alternative approaches. The claim that DSec mitigates agent misbehavior such as reward hacking is presented by the authors themselves, with no detail in the abstract on how that mitigation is measured or verified.

“these mechanisms reduce environment setup and image-distribution overhead, improve memory efficiency, and preserve latency-sensitive performance under high-density overcommit”

— DSec paper