DeepSeek published a paper on DSec, its infrastructure for making agent training elastic
This story was corrected after publication (Oct. 9, 2026). Details
The paper, with around 130 authors including founder Liang Wenfeng, describes DeepSeek Elastic Compute, or DSec. The system mass-produces the isolated environments in which AI agents are trained with reinforcement learning, managing four types, from simple function calls and containers to lightweight VMs and full VMs running complete operating systems such as Android, through a single Python interface. A unit of about 160 nodes can create more than 5,000 environments per second and about 3 million a day, and run more than 380,000 at once. The paper is about the environments agents are trained in, not about model serving.
Sources
- arXiv, “DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Effective Agentic Training at Scale”, (arxiv.org)
Corrections and updates
- The version first posted on Instagram listed full Windows and macOS systems among the environments DSec manages. The paper does not mention Windows or macOS; its example for full VMs is Android, and the text has been corrected.
About this story
This story was posted on Instagram by @jarrus.tech on Sept. 30, 2026.
Spotted an error in this story? [email protected] · Instagram
Short link: thejarrus.com/en/deepseek-dsec
This story in Turkish: DeepSeek, ajan eğitimini esnek hale getiren altyapısı DSec’i anlatan bir makale yayımladı
On the same topic
Robot tests: even the best model completed only 19% of 84 robot tasks, and 63 tasks were solved by no model
In RobotWorld, a benchmark released October 7, the top model, GPT-6 Astra, completed only 16 of 84 simulated robot tasks; no model solved 63 of the tasks.
OpenAI’s image model posts a near-perfect score on text-dense images
On UltraText Bench, a dense-text test led by Westlake University, OpenAI’s GPT Image 2 ranked first of 24 configurations with 99.35 out of 100.
HackerRank’s AI interviewer opens to all customers after more than 500,000 job interviews
HackerRank made its AI interviewer Chakra generally available to customers on October 5, after a roughly six-month beta with more than 500,000 interviews.