Case Study: Hugging Face achieves secure, fast code execution for Open R1 with E2B Sandboxes

How Hugging Face Is Using E2B to Replicate DeepSeek-R1

The customer, Hugging Face, faced the challenge of needing to safely execute potentially risky, AI-generated code as part of a reinforcement learning pipeline for their Open R1 project. The execution was required for a verifiable reward function, but running unverified code locally posed significant security dangers, such as system corruption or data loss.

The vendor, E2B, provided a solution through their secure E2B Sandboxes, which are isolated cloud environments. This allowed Hugging Face to execute code safely and at scale, integrating it into their training process in just a few hours. The solution was praised for its security, speed—with sandboxes starting in about 150ms—and cost-effectiveness, costing only a few dollars per training run. This enabled the research team to launch hundreds of secure sandboxes in parallel per training step.

Hugging Face

Lewis Tunstall

Research Engineer