DeepSeek Elastic Compute (DSec): Agentic Training Sandbox
At a Glance
| Item | Value |
|---|---|
| Published | 2026-09-27 |
| Source type | Unverified (not confirmed by a primary source) |
Values determined by this site’s code at collection time. Dates are JST.
Overview
Please note that the information covered in this article is based on a paper submitted to the arXiv preprint server and is unverified information that has not been officially confirmed.
According to a paper published on arXiv, the architecture and operational results of “DeepSeek Elastic Compute (DSec)," a sandbox execution infrastructure designed for agentic training and evaluation of Large Language Models (LLMs), have been reported. The subjects of this report are a unified SDK and cluster execution infrastructure that bundle FnCall, containers, microVMs, and full VMs, as well as the cluster-wide distributed file system “Fire-Flyer File System (3FS)." The report presents a mechanism that decouples stateful rollout execution from GPU reinforcement learning (RL) processing, enabling the rapid startup and maintenance of large-scale sandbox groups under high-density resource overcommit environments.
Claims and Rationale
According to the presenters’ report, large-scale agent training and evaluation require sudden, massive demands for isolated, stateful execution environments to perform tasks such as repository inspection, tool calling, command execution, and interaction with individual services. Such environments require diverse functionalities and isolation levels, need to maintain state through long-running interactions, and face the challenge of limited reusability as they are invoked from various images. DSec is claimed to solve these issues not by relying on a single runtime, but as a highly elastic platform spanning the entire cluster.
The main system architecture and rationale cited by the presenters are as follows:
Integration of Multiple Sandbox Backends via a Unified SDK
DSec is reported to provide sandbox backends with different isolation levels—FnCall, containers, microVMs, and full VMs—through a single unified SDK. This is said to centralize placement and lifecycle management across the entire cluster, making it possible to configure environments by combining individually version-controlled layers.
On-Demand Loading in Coordination with the Distributed File System “3FS"
To suppress image distribution overhead, a design is adopted in which image data is loaded on-demand from the cluster-wide distributed file system, the “Fire-Flyer File System (3FS)." Furthermore, it is claimed that high-density execution environments are constructed by combining memory sharing, memory reclamation, and CPU scheduling.
Co-Design with Reinforcement Learning (RL) Frameworks
According to the presenters, DSec is designed to work in coordination with RL frameworks. By decoupling state-maintaining rollout execution from preemptible GPU training and linking the sandbox lifecycle with training processing, the system is equipped with a mechanism to reclaim idle resources while preserving rollout states. It is also reported that mechanisms to mitigate improper agent behaviors such as reward hacking are incorporated.
Measured Values of Operational Results in Production Environments
As a result of measurements from their own operation, the presenters report that the following scale and processing performance were achieved in a single production-scale unit (around 160 nodes):
- Processing scale: Approximately 3 million sandboxes provided per day per 1-node cluster unit (around 160 nodes)
- Concurrency: Supports over 380,000 concurrent execution sandboxes in a production environment
- Throughput: Sustainable creation of over 5,000 sandboxes per second
According to the presenters’ evaluations and production operation experience, these mechanisms reduce the overhead associated with environment setup and image distribution, successfully balancing improved memory efficiency with the maintenance of low-latency performance under high-density overcommit environments.
Prerequisites
The achievements and verification results of DSec presented in this paper are reported on the premise of a large-scale distributed computing environment and a specific system configuration. The specific prerequisites and architectural components identifiable from the documentation are as follows:
-
Target workload: Agentic training and evaluation processing in Large Language Models (LLMs). The targets are processes that require long-term state retention (stateful) and sudden creation of execution environments, such as repository inspection, tool calls, command execution, and interaction with various services.
-
Sandbox backends: Four types—FnCall (function calls), containers, microVMs, and full VMs (Full-VM). These backends with different isolation levels are integrated and controlled via a unified SDK.
-
Cluster scale and hardware configuration: A cluster environment consisting of approximately 160 nodes per production-scale unit. High-density resource overcommit operation is reported.
-
Storage infrastructure: The “Fire-Flyer File System (3FS)," a distributed file system targeting the entire cluster. It is said to be used as an infrastructure for loading image data on-demand.
-
Integrated framework: Co-design with reinforcement learning (RL) frameworks. The premise is an environment that separates rollout processing and training processing on GPUs, combining preemptible operation, memory sharing, memory reclamation, and CPU scheduling.
Reproducible Scope Locally
Regarding the extent to which readers can reproduce and evaluate DSec mechanisms on their own PCs or local environments, it is reported to be limited based on the currently available documentation.
-
Published materials: The PDF and HTML versions of the paper (arXiv:2609.22978) are published on arXiv, making it possible to review technical details regarding the system architecture and design philosophy.
-
Source code and reproduction steps: This material does not contain descriptions regarding the provision of source code repositories or reproduction commands for DSec’s unified SDK, code for integration with 3FS, or for directly verifying its operation. Therefore, it is judged that reproduction steps to run it as-is in a local environment have not been published.
-
Scale of the target environment: This system is designed and evaluated targeting a data center-scale production infrastructure premised on a cluster scale of approximately 160 nodes and over 380,000 concurrent sandboxes. It is stated that it is difficult to directly reproduce equivalent high-density operations and performance results in limited equipment environments such as a single PC or a small local server.
What the Material Does Not Cover
In this material (such as the abstract information of the preprint paper), the following information necessary for evaluating the technical assessment and practicality of DSec locally is not described:
-
Open-source availability: There is no description regarding whether DSec’s unified SDK, management infrastructure, and peripheral tools are scheduled to be released generally as open-source software, nor about their license format.
-
Minimum hardware requirements and detailed specs: There are no descriptions of detailed hardware specs for the nodes constituting the production unit, such as the number of CPU cores per node, installed memory capacity, network bandwidth, or the model of GPUs used.
-
Detailed specifications of backend technology: Detailed information on the specific virtualization technologies used to realize microVMs and full VMs (such as hypervisor names) or the internal structure of FnCall’s lightweight execution mechanism is not mentioned.
-
Costs and overhead in deployment and operation: Specific figures regarding infrastructure costs required for system construction and maintenance, as well as effective storage bandwidth or initial construction costs when introducing 3FS, are not described.
-
Evaluation data in small-scale or other environments: Operation verification data and comparison results under small-scale configurations other than the 160-node scale, different distributed file systems, or heterogeneous GPU cluster environments are not shown.
Sources
This article contains unverified information. We will append an update note once it is confirmed by a primary source.