Liang Wenfeng's byline, DeepSeek publishes a new paper, aiming at large-scale Agent training.

September 23 — A latest paper co-authored by DeepSeek founder Liang Wenfeng has been made public, systematically revealing for the first time the technical details of DeepSeek's Agent training sandbox platform DSec (DeepSeek Elastic Compute). The paper was submitted on September 19, with more than 130 authors, including Liang Wenfeng.

The DSec platform first appeared in the DeepSeek V4 technical report. Its main purpose is to provide a sandbox for Agent training to enable stable operation of large-scale Agent training. The paper explicitly states that from DeepSeek V3.2 to V4.1, all sandbox workloads for RL training and evaluation ran on DSec. Now that the technical details are public, it is effectively handing over the "secret manual."

The paper shows that the DSec platform is enormous in scale. One production unit consists of about 160 CPU nodes, 30,000 cores, and 250TB of memory, hosting PB-level images. In terms of capability, the DSec platform serves about 3 million sandboxes per day, with peak concurrency exceeding 380,000, a creation speed of more than 5,000 per second, and a single training task can spin up as many as 32,000 sandboxes at once.