Overview
Danau5tin/Orca-Agent-v0.1 is a 14 billion parameter orchestration agent, fine-tuned from Qwen3-14B, specifically designed to act as the 'brain' in multi-agent systems. Unlike traditional LLMs that directly generate code, Orca-Agent focuses on task analysis, decomposition, and delegation. It orchestrates explorer agents to understand systems, dispatches coder agents with precise instructions for implementation, and uses additional explorer agents for verification, all while maintaining a comprehensive context store of discovered knowledge.
Key Capabilities
- Task Orchestration: Analyzes user tasks, breaks them into focused subtasks, and manages the workflow.
- Agent Delegation: Dispatches specialized explorer and coder agents, providing clear instructions and managing their interactions.
- Context Management: Maintains a persistent context store, accumulating and leveraging discovered knowledge throughout a task's lifecycle.
- Verification: Utilizes explorer agents to verify changes and ensure task completion accuracy.
Performance
Orca-Agent-v0.1 (14B), when paired with a Qwen3-Coder-30B subagent, achieves an 18.25% success rate on TerminalBench. This performance is competitive, placing it within striking distance of much larger models like Qwen3-Coder-480B, which scored 19.7% on the same benchmark. The model was trained using GRPO with curriculum learning on 32 H100 GPUs, with 16 dedicated to training and 16 for inference across the orchestrator and subagents. The full training details and code are open-sourced here.
Good For
- Developing sophisticated multi-agent coding systems.
- Automating complex software development or system administration tasks.
- Researching agentic AI architectures and orchestration strategies.