SearchSwarm-30B-A3B

SearchSwarm-30B-A3B

SearchSwarm: Towards Delegation Intelligence in Agentic LLMs for Long-Horizon Deep Research

Project Page | GitHub Repository

SearchSwarm is a foundation action model designed for long-horizon deep research tasks. It introduces delegation intelligence, enabling a main research agent to use subagents as an active context-management mechanism. The main agent decomposes complex tasks, dispatches bounded subtasks to subagents, and synthesizes citation-grounded reports while staying within a finite context budget.

Overview

SearchSwarm focuses on several key aspects of agentic intelligence:

• Subagents as Context Management: Subagents work in independent contexts and return compact, evidence-grounded reports, preventing the main agent's context window from being overwhelmed.
• Delegation Intelligence: The model is trained to determine when and what to delegate, how to brief subagents effectively, and how to verify their findings.
• Harness-guided Synthesis: The authors used a specialized harness to synthesize high-quality SFT data to internalize these delegation capabilities into the model weights.

Performance

SearchSwarm-30B-A3B achieves state-of-the-art results among comparable 30B-A3B open-source lightweight research agents on deep research benchmarks:

Citation

@article{ning2026searchswarm,
  title={SearchSwarm: Towards Delegation Intelligence in Agentic LLMs for Long-Horizon Deep Research},
  author={Ning, Pu and Chen, Quan and Tao, Kun and Tang, Xinyu and Wang, Tianshu and Cao, Qianggang and Kong, Xinyu and Wen, Zujie and Zhang, Zhiqiang and Zhou, Jun},
  journal={arXiv preprint arXiv:2606.09730},
  year={2026}
}

Acknowledgements

This project builds on open-source infrastructure including vLLM, ms-swift, Megatron-LM, Qwen-Agent, Serper, and Jina.