SwarmLLM: Distributed Inference and Orchestration for Edge-Native Agent Swarms

ABSTRACT

The computational and bandwidth requirements for massive multi-agent swarms present a critical bottleneck for cloud-centric AI infrastructure. We introduce SwarmLLM, a framework for distributed inference that orchestrates agentic workflows across networks of heterogeneous edge devices. By leveraging dynamic model quantization, token routing, and decentralized state synchronization, SwarmLLM allows a cluster of smartphones and IoT devices to collaboratively execute complex reasoning tasks that typically require a datacenter GPU. Our results demonstrate a 15x reduction in cloud latency and a highly resilient, privacy-preserving architecture for local agentic ecosystems.