Case Study: Hugging Face accelerates secure code execution for Open R1 with E2B Sandboxes

Hugging Face launches Open R1 in hours with E2B

Hugging Face, a leading AI research company, faced a significant challenge in its Open R1 project to replicate the DeepSeek-R1 model. A core part of this effort involved reinforcement learning that required executing potentially risky, AI-generated code to verify its correctness. The challenge was to find a secure, fast, and cost-effective method to run this code at scale without risking damage to their local systems.

The solution was implemented using E2B Sandboxes, which provided secure, isolated cloud environments for code execution. E2B enabled Hugging Face to securely run hundreds of code evaluation tasks in parallel per training step, with each sandbox starting in about 150-170 milliseconds. This integration was critical for minimizing GPU idle time and was both simple to set up and cost-effective, with a single training run costing only a few dollars. As a result, Hugging Face successfully scaled its reinforcement learning pipeline to use code execution feedback, a key factor in improving the reasoning capabilities of open-source models.

Hugging Face