Orchard: An open framework for scalable agentic AI
Microsoft is releasing Orchard as open infrastructure for training agents inside the same harnesses where they run.
Orchard Env uses Kubernetes to manage isolated rollout environments across coding, web, and assistant tasks. The note says its recipes include Orchard-SWE, which reaches 69.7% on SWE-bench Verified, or 73.0% with value-model reranking, using about 3 billion active parameters. Orchard-GUI and Orchard-Claw are presented as evidence that smaller open-weight models can perform strongly when the environment and training workflow are built for real agent use. Microsoft Research AI's note
Orchard Env uses Kubernetes to manage isolated rollout environments across coding, web, and assistant tasks. The note says its recipes include Orchard-SWE, which reaches 69.7% on SWE-bench Verified, or 73.0% with value-model reranking, using about 3 billion active parameters. Orchard-GUI and Orchard-Claw are presented as evidence that smaller open-weight models can perform strongly when the environment and training workflow are built for real agent use. Microsoft Research AI's note
score 6