Case Study: Martian achieves 95% faster SWE-Bench Verified execution with Daytona

Martian speeds SWE-Bench Verified execution by 95% with Daytona

Martian, an AI research lab, needed to scale its agentic training and evaluation framework, ARES, which required the ability to provision hundreds of isolated, parallel sandboxes for reinforcement learning tasks. The challenge was to find a managed runtime solution that could handle this elastic, large-scale provisioning with sophisticated orchestration to avoid pulling engineers from core research. They partnered with Daytona for its sandbox infrastructure.

Daytona provided a solution through its runtime platform, embedding directly into ARES via an SDK and API to dynamically provision and tear down concurrent sandboxes on demand. This resulted in a 95% faster execution of the SWE-Bench Verified benchmark, allowed Martian to provision over 20,000 sandboxes in a single week, and saved an estimated three months of engineering time, enabling faster research cycles and more efficient compute usage.

Martian