The news of September 24thDeepSeek Internal Agent training and assessment sandbox infrastructure DSec was recently released。paperIt was submitted to arXiv on 19 September, and Liang Wenbing is listed as the author. It describes the operational platform that has been put into production and not the release of new models。
In the course of training, Agent needs to open a code warehouse, execute commands and call tools, and each mission must have an independent environment of sustainable preservation. DSec provides functional calls, containers, mini-virtual machines and four sandbox backends of complete virtual machines through a unified SDK, and selects the segregation by task. The platform also contains on-demand mirrors from distributed file systems to reduce the centralized download pressure when the environment is activated。
According to the paper, a production unit with about 160 nodes handles approximately 3 million sandbox cases per day. The production environment has peaks of over 380,000, with more than 5,000 created per second; the DeepSeek V3.2 to V4.1 intensive learning training and sandbox assessment missions are on DSec. These size and performance figures are derived from the team ' s systematic reports and are not yet independently measured by third parties。