❯ Paper signed by Liang Wenfeng details DeepSeek’s DSec agent-training system, which creates 5,000 sandboxes a second and runs 380,000 at once

the scaleA paper bearing the name of DeepSeek founder Liang Wenfeng was published on arXiv, describing DSec (DeepSeek Elastic Compute), the company’s elastic compute platform for agent training. According to the paper, a production unit of about 160 nodes can create more than 5,000 sandboxes a second, up to about 3 million a day, with more than 380,000 running at peak. A single cluster has about 30,000 CPU cores and 250TB of memory.

how it worksDSec exposes four sandbox backends (function calls, containers, microVMs and full VMs) through one Python SDK and schedules resources through a six-layer pipeline. Environments are assembled on demand from three independently versioned EROFS read-only image layers, with image data loaded on demand from 3FS, DeepSeek’s in-house distributed filesystem. Memory sharing, CPU priority scheduling and cloud bursting push up sandbox density. The system was designed together with the reinforcement learning framework, so when GPU training is preempted, sandbox state is kept and rollouts resume later.

anti-cheatingThe paper addresses agents that exploit reward loopholes or damage systems during training, which DSec guards against with AppArmor and eBPF. It also acknowledges that these defenses will need constant upgrading as models get stronger.

▪ SIGNALThe bottleneck in agent training is shifting from GPUs to CPUs and environment scheduling, and by publishing DSec, DeepSeek signals it no longer sees this layer as a moat.