Orchard: Open-Source Agentic Framework Sets SWE-Bench SOTA at 67.5% with Qwen3-30B

arXiv·high signal

Orchard introduces credit-assignment supervised fine-tuning to learn from productive segments of unresolved trajectories, plus Balanced Adaptive Rollout for sparse-reward RL. Using Qwen3-30B-A3B-Thinking, Orchard-SWE achieves 64.3% on SWE-bench Verified after SFT and 67.5% after RL — new SOTA among open-source models of comparable size. Orchard-GUI reaches 68.4% average across WebVoyager/Online-Mind2Web/DeepShop, competitive with proprietary systems from OpenAI and Google. Distills 107K trajectories from MiniMax-M2.5 and Qwen3.5-397B.