DeepSeek Elastic Compute (DSec) sandbox platform
DeepSeek describes the sandbox platform behind its agentic RL: about 3 million sandboxes a day, 380,000+ concurrent, 5,000+ created per second.
A production platform exposing function-call, container, microVM and full-VM sandboxes behind one SDK, loading image layers on demand from 3FS and co-designed with the RL framework to preserve rollout state and curb reward hacking. One unit spans about 160 nodes.
- Date
- Saturday, 19 September 2026
- Lab
- DeepSeek
- Kind
- paper
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Sandboxes per day (one unit) | about 3 million About 160 nodes; 380,000+ concurrent sandboxes and 5,000+ creations per second in production | company |
| Author count | 131 Per the arXiv listing (v1 2026-09-19); affiliations not opened | independent |
arXiv 2609.22978 (2026-09-19). A rare look at agent-training infrastructure at a Chinese lab; all figures are DeepSeek's. The paper does not say which models were trained on it (inferred to be V3.2 and V4-era agent post-training).
Sources
This record was partly confirmed: some claims could not be checked on 6 October 2026. How we check