NVIDIA Open-Sources PAIR for Local Multi-Device AI Inference Scheduling
NVIDIA has released Personal AI Router (PAIR), an open-source tool that schedules and routes local AI inference requests across compatible local network devices like Apple Silicon Macs, RTX PCs, and DGX Spark systems. It acts as a unified traffic controller, presenting OpenAI- and Ollama-compatible endpoints so local AI applications can offload tasks to idle computers on the LAN. PAIR enables users to easily transform heterogeneous hardware into a private, multi-node personal AI cluster without sending sensitive data to cloud servers. This significantly improves execution speed for multi-agent workflows and concurrent LLM tasks by taking full advantage of all compute resources available on a local network. PAIR distributes independent concurrent requests across nodes rather than parallelizing a single response, meaning individual response speeds remain bound to the processing device. In benchmark tests running five AI sub-agent tasks, distributing the workload across an RTX laptop, a DGX Spark, and an RTX 5090 reduced total task completion time from 18 minutes down to 8 minutes and 48 seconds.
## BACKGROUND
Running Large Language Models (LLMs) locally using tools like Ollama or LM Studio allows developers to keep code, documents, and prompts entirely private. However, as developers adopt multi-agent frameworks where several AI assistants process tasks simultaneously, a single local computer can quickly run out of compute capacity. Local inference engines traditionally operate on individual machines, making it difficult to utilize idle processing power from other computers on the same home or office network.