- Published on
NVIDIA Releases Personal AI Router (PAIR): An Open Source Virtual Inference Router that Distributes Local AI Requests...
- Authors
- Name
- DeepHorus
NVIDIA has launched a public beta of its Personal AI Router (PAIR), an open-source virtual inference router designed to solve a growing bottleneck in local AI workflows. The tool automatically distributes multiple independent AI requests across compatible devices on a home network, such as RTX-powered PCs, DGX Spark systems, and Macs, preventing requests from queuing on a single machine.
The release addresses a key challenge in multi-agent AI systems, where a single user prompt can spawn dozens of model calls that compete for resources on one engine. PAIR aims to utilize idle hardware on the same network, significantly accelerating processing times for complex tasks. The public beta (v0.1.1) is available now with installers for Windows, macOS, and Linux, with the full source code released under the Apache 2.0 license.
Crucially, PAIR introduces no new cluster API. It works by proxying the existing Ollama and LM Studio interfaces that many tools already use. This design means existing agent harnesses require no modification; they continue sending requests to a standard port, while PAIR handles routing the work to an optimal available node.
Security and setup are managed through a local pairing process. PAIR uses mDNS for automatic discovery on a local network, and devices are connected via a six-digit PIN for trust. All communication between paired nodes is secured with mTLS certificates. The router can even initiate model downloads on new nodes to simplify cluster expansion.
The scheduling algorithm intelligently selects a node based on several factors: whether the node is online, if the required engine is running, if the exact model file is present, current job load, and GPU utilization. Importantly, PAIR assigns each entire request to one node for its lifetime, avoiding complex techniques like VRAM pooling or request sharding across machines.
In a demonstration, NVIDIA showed a five-subagent workload drop from an average of 18 minutes on a single laptop to 8 minutes and 48 seconds on a three-device cluster. This cluster combined an RTX Spark laptop, a DGX Spark, and an RTX 5090 desktop. The performance gain highlights the potential for distributed local inference.
PAIR supports a wide range of modern hardware, including NVIDIA GeForce RTX 20 Series and newer GPUs, RTX PRO workstations, DGX Spark, and Apple M4 or newer silicon. The tool enables users to build a personal, scalable AI compute cluster from existing devices, marking a significant step toward efficient, decentralized AI processing.
