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

Together AI · 1 year ago · 25

Agentica and Together AI released DeepSWE-Preview, a 32B parameter coding agent trained entirely with reinforcement learning that achieves 59% accuracy on SWE-Bench-Verified with test-time scaling. The model was trained on 4,500 real-world software engineering tasks over six days using 64 H100 GPUs, reaching 42.2% Pass@1 performance as the best open-weight coding agent. The team open-sourced the dataset, training code, and evaluation logs, along with improvements to the GRPO algorithm including techniques like compact filtering and length normalization for stable multi-turn agent training.