DeepSeek V4 Unveils Production-Grade Agent Sandbox DSec: Single-Cluster Scheduling Hundreds of Thousands of Concurrent Jobs, Unifying Four Heterogeneous Platforms

According to Dynasty Beating Monitoring, the DeepSeek V4 Technology Report has unveiled for the first time the core infrastructure supporting post-training of the Agent and large-scale evaluation, the production-level elastic compute sandbox DSec (DeepSeek Elastic Compute).

Current large-scale model reinforcement learning requires an extremely large-scale trial-and-error environment. The report reveals that in actual production, a single DSec cluster can concurrently schedule tens of thousands of sandbox instances. The system is implemented in Rust and interfaces with the in-house 3FS distributed file system. By using on-demand loading, it breaks through the performance bottleneck of cold-starting a large number of sandboxes.

In terms of developer experience, DSec has unified function invocation, containers, micro VMs, and full VMs into a set of Python SDK, allowing for a parameter change to switch between the four execution bases. To address the common problem of task preemption in computing clusters, DSec has introduced a global trajectory log: when a task resumes, the system will directly "fast-forward" to replay the cached command execution results, achieving both rapid checkpoint continuation training and avoiding non-idempotent errors caused by repeated execution.