AnswersPublished17 sources
DeepSeek DSec: The Sandbox Infrastructure Behind Large-Scale Agent Training
DeepSeek DSec runs isolated agent actions for reinforcement learning rollouts and evaluation. Four backends let tasks use lightweight function execution, containers, microVMs, or full VMs; layered images and on demand loading reduce the need to copy entire environments.
Published byEdited with GPT-6 SolImages generated with GPT Image 2
Research answer
Key takeaways
- DeepSeek DSec runs isolated agent actions for reinforcement learning rollouts and evaluation.
- Four backends let tasks use lightweight function execution, containers, microVMs, or full VMs; layered images and on demand loading reduce the need to copy entire environments.
- Secondary reports describe a single training task requesting 32,000 sandboxes at once.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: How does DeepSeek Elastic Compute (DSec), described in DeepSeek’s September 2026 arXiv report co-authored by Wenfeng Liang, support large-sc. Article summary: DeepSeek Elastic Compute (DSec) is the sandbox service DeepSeek uses to run agent actions during reinforcement-learning rollouts and evaluation, separately from the GPU training pods. Its purpose is to make large numbers. Topic tags: general, academic, general web. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fake num
DeepSeek Elastic Compute (DSec) provides the execution environments in which AI agents use tools, run code, and complete tasks during reinforcement-learning training and evaluation. Its September 2026 report describes a platform built to create and manage large numbers of sandboxes without treating every task as a full virtual machine or distributing a complete new image for every environment change. 1
Why agent training needs a sandbox platform
An agent rollout may involve a brief function call, a shell command, code running in a workspace, or a task that needs an entire operating-system environment. Running many such rollouts in parallel requires different execution options and rapid sandbox creation. The code and actions involved also cannot simply be assumed safe: reporting on the DSec paper describes risks including agent misbehavior and attempts to manipulate an environment or its results. Isolation is therefore part of the design, though no single sandbox mechanism should be treated as a complete security guarantee. 1
9
10
Four backends behind one interface
DSec exposes FnCall, containers, microVMs, and full VMs through a unified SDK. FnCall is suited to lightweight function-style execution; containers provide more general-purpose workspaces; microVMs offer a VM isolation boundary; and full VMs accommodate tasks that need a complete virtual machine. Training and evaluation clients can use the common interface while DSec coordinates sandbox placement and lifecycle across its hosts. 1
3
The practical distinction is that workloads need not all pay for the same environment. A short tool call and a task requiring a full VM can use different backends within the same platform. The backends offer different isolation properties, but the available evidence does not establish that any one of them eliminates the risks of untrusted agent execution. 1
10
How DSec makes environments reusable
DSec separates an environment into independently versioned, read-only base image, workspace, and toolkit layers using EROFS. It assembles the needed layers at startup with a writable overlay. If a toolkit changes, the system need not rebuild every complete combination of toolkit, base image, and workspace. 1
11
Those images are stored on 3FS, DeepSeek’s distributed filesystem. Rather than copying every image byte before execution, DSec can prepare metadata locally and fetch file data as a sandbox reads it. This on-demand approach targets image-transfer time and local storage pressure when many sandboxes start together; the report also describes memory management and CPU scheduling for dense execution. 1
11
What DeepSeek reports at production scale
DeepSeek describes one production-scale unit with approximately 160 CPU nodes, 30,000 cores, and 250 TB of DRAM, backed by petabyte-scale image storage. For that unit, it reports roughly 3 million sandbox instances per day, more than 380,000 concurrent sandboxes at peak, and a creation rate exceeding 5,000 sandboxes per second. The daily count, concurrency peak, and creation rate measure different things: the peak creation rate should not be read as a rate maintained throughout the day. Nor should a unit’s figures be presented as totals for every DSec deployment. 1
3
12
Several reports citing the paper say a single production training task requested or launched 32,000 sandboxes at once. That is useful evidence of bursty single-job demand, but the accessible reporting does not establish it as a hard per-job limit. 8
11
Use from V3.2 through V4.1—and GPU preemption
DeepSeek says DSec has handled sandbox workloads for its reinforcement-learning training and evaluation from V3.2 through V4.1. Its V4.1 technical report further describes long-lived rollouts running on DSec outside the preemptible GPU training pool. When a trainer is preempted, rollout execution can be suspended and offloaded with its state preserved for later resumption, rather than discarding the agent’s progress with the GPU training pod. 1
3
Together, these choices make DSec more than a way to launch isolated processes: it is infrastructure for provisioning varied agent environments, managing their execution at scale, and keeping long-running rollout work separate from the scheduling of GPU trainers. The scale and operational benefits above are DeepSeek’s reported results, not independently established capacity guarantees for other deployments. 1
3
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "DeepSeek DSec: The Sandbox Infrastructure Behind Large-Scale Agent Training"?
DeepSeek DSec runs isolated agent actions for reinforcement learning rollouts and evaluation.
What are the key points to validate first?
DeepSeek DSec runs isolated agent actions for reinforcement learning rollouts and evaluation. Four backends let tasks use lightweight function execution, containers, microVMs, or full VMs; layered images and on demand loading reduce the need to copy entire environments.
What should I do next in practice?
Secondary reports describe a single training task requesting 32,000 sandboxes at once.