Orchestra-o1
A multi-agent orchestration framework that splits omnimodal tasks into subtasks handled by specialized sub-agents equipped with perception and action tools, coordinated by a planning main agent.
Orchestra-o1 uses a MainAgent/SubAgent architecture: the MainAgent decomposes an omnimodal task into independent or dependent subtasks across multiple turns, then creates specialized SubAgents equipped with perception tools such as image, audio and video analysis, and action tools such as web search, page visits and code execution, running them in parallel. An online specialization module handles sub-task preparation, model selection, tool integration and memory allocation for each SubAgent.
The released model, Orchestra-o1-8B, was trained with Group Relative Policy Optimization on Qwen3-8B using an LLM-as-judge reward, and is evaluated on the OmniGAIA benchmark for omnimodal question answering.
The repository includes training and inference code, configuration for connecting arbitrary LLM API backends, and instructions for preparing the OmniGAIA dataset. It is written in Python.