DeepSeek DSec large-scale Agent training sandbox platform disclosed in paper listing Liang Wenfeng as author
BIBIBI
AT A GLANCE
The DeepSeek paper discloses the technical details of DSec, an Agent training sandbox platform.
Article
Liang Wenfeng listed as author on DeepSeek’s new paper targeting large-scale Agent training
PANews, September 23: According to Sina Finance, the latest paper listing DeepSeek founder Liang Wenfeng as an author has been made public. It systematically presents, for the first time, the technical details of DeepSeek’s Agent training sandbox platform, DSec (DeepSeek Elastic Compute). The author list includes more than 130 people, including Liang Wenfeng. The paper shows that the DSec platform is very large in scale: one production unit consists of approximately 160 CPU nodes, 30,000 cores, and 250TB of memory, and hosts petabyte-scale images. In terms of capabilities, the DSec platform serves approximately 3 million sandboxes per day, supports peak concurrency of more than 380,000, creates more than 5,000 sandboxes per second, and can launch up to 32,000 sandboxes at once for a single training task.
The paper lists Liang Wenfeng as an author, and the author list includes more than 130 people.
02
One DSec production unit consists of approximately 160 CPU nodes, 30,000 cores, and 250TB of memory, and hosts petabyte-scale images.
03
DSec serves approximately 3 million sandboxes per day, supports peak concurrency of more than 380,000, and creates more than 5,000 sandboxes per second.
04
A single training task can launch up to 32,000 sandboxes at once.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
DSec is a sandbox platform used by DeepSeek for Agent training. A sandbox can be understood as an isolated runtime environment; the figures in the article indicate that the platform can run and rapidly create large numbers of sandboxes simultaneously.
Why it matters to readers
These figures indicate that DSec is designed for large-scale Agent training scenarios, with a focus on creating and concurrently serving large numbers of isolated runtime environments.
If you are new to Agents, you can think of DSec as a platform that provides independent experimental environments in batches, allowing many training tasks to run in parallel.
Risks and unknowns
The source says that the paper has been “made public,” but does not provide the paper itself or independent verification information.
The source does not specify where DSec is deployed, its operating costs, or its specific training results.
Related Developments
Loading event timeline…
Related concepts
Agent
This term is not in the glossary yet. Browse related concepts in the glossary.
