Microsoft Research releases Orchard for training agentic models

Orchard is an open-source framework released with three specialist models. Qwen3-30B-based Orchard-SWE scored 67.5% on SWE-bench Verified. The 4B Orchard-GUI reached 74.1% success on WebVoyager and 67.0% on Online-Mind2Web. Orchard Env provides reusable sandbox lifecycle controls for training coding, GUI, and personal assistant agents.

Visit original source