DeepSWE: Training a Fully Open-sourced, State-of-the-Art Coding Agent by Scaling RL

What this post added

Introduces DeepSWE-Preview, a state-of-the-art open-source coding agent trained from Qwen3-32B using only reinforcement learning (RL). Details the training recipe, including scalable dataset curation with R2E-Gym, environment setup with tools like Bash execution and file editing, and a sparse outcome reward model. The training process utilizes an enhanced GRPO++ algorithm with features like Clip High, No KL Loss, and Compact Filtering for stable and performant training of multi-turn agents. Integrates Kubernetes for scalable agent rollout collection, managing thousands of containers for RL training.