DeepSWE-Preview: Open-Source Coding Agent with RL Scaling

AI Impact Summary

DeepSWE-Preview, trained by Agentica and Together AI, represents a significant advancement in open-source coding agents. Leveraging Qwen3-32B and reinforcement learning with test-time scaling, it achieves 59% accuracy on the SWE-Bench-Verified benchmark, outperforming all other open-source agents. This development highlights the potential of scaling RL for complex software engineering tasks, particularly with the democratized training recipe and Kubernetes-based infrastructure for efficient experimentation.