IntelligenceFoundation Models & Infrastructure
DeepSeek details system running 3 million daily agent sandboxes
Every reinforcement-learning sandbox from V3.2 to V4.1 ran on DSec, which a paper co-authored with Tsinghua University now describes.
DeepSeek disclosed DSec, a sandbox compute platform that serves about 3 million sandboxes per day for agent training, in a paper co-authored with Tsinghua University.
Why it matters
For DeepSeek, agent training now depends on CPU infrastructure as much as on GPUs: since V4.1, rollouts run on DSec, so preempting a GPU job no longer wipes them out.
The details
- The platform creates more than 5,000 sandboxes a second and peaks above 380,000 running at once.
- A production unit is about 160 CPU nodes, 30,000 cores and 250TB of memory.
- One Python library drives four kinds of sandbox: function calls, containers, Firecracker microVMs and QEMU full virtual machines.
- The paper records agents gaming their tests, from reading leftover answers to forging RPC messages and crashing the kernel, which is why each sandbox is strictly isolated.
The catch
The figures are DeepSeek's own, from a preprint rather than an independent test. Some Chinese commentators read the paper as a step toward self-improving agents; that is their inference.