Orchard: Open-Source Framework for Scalable Agentic AI Modeling
A new open-source framework named Orchard has been launched to enable scalable agentic modeling, with the goal of evolving LLMs into independent agents capable of handling intricate tasks. This framework tackles the limitations in infrastructure and training that hinder open research, as many top-performing systems depend on proprietary code. At its core is Orchard Env, a streamlined environment service that offers reusable components for managing sandbox lifecycles across various task domains, agent harnesses, and pipeline stages. Additionally, three agentic modeling recipes have been developed: Orchard-SWE focuses on coding agents, utilizing 107K trajectories from MiniMax-M2.5 and Qwen3.5-397B, and incorporates credit-assignment SFT to learn from effective segments. The framework is built to facilitate planning, reasoning, tool usage, and multi-turn interactions with environments.
Key facts
- Orchard is an open-source framework for scalable agentic modeling.
- Orchard Env is a lightweight environment service for sandbox lifecycle management.
- Orchard-SWE targets coding agents with 107K distilled trajectories.
- Trajectories distilled from MiniMax-M2.5 and Qwen3.5-397B.
- Credit-assignment SFT introduced for learning from productive segments.
- Addresses infrastructure and training gaps in open research.
- Many high-performing systems rely on proprietary codebases.
- Supports planning, reasoning, tool use, and multi-turn interaction.
Entities
Institutions
- arXiv
- MiniMax
- Qwen