Nvidia is testing free software that lets the computers in a home share the work of running AI models. The beta release, called PAIR, short for Personal AI Router, links compatible machines on one network into a single pool for local inference.
PAIR is code, not a box. Nvidia says it is neither a physical router nor a stand-in for engines like Ollama and LM Studio, which still execute each model on a single chosen machine. The program handles coordination instead: it finds participating systems, checks which are free, schedules independent jobs, and delivers each response back to the app that made the request.
The beta supports Windows 11, DGX OS, Ubuntu, and macOS Tahoe on x64 and arm64, with Windows on ARM listed as experimental. Nvidia names GeForce RTX GPUs from the 20 Series onward, RTX PRO workstations, DGX Spark and GB10 systems, and Apple M4 Macs or newer as compatible hardware.
No special cables or racks are needed, and once models are downloaded the cluster runs without an internet connection. The source code is published under Apache 2.0, and Nvidia is inviting developers to inspect it and file issues.
PAIR turns idle machines into an on-prem inference pool, letting homes and small teams take on bigger agent workloads than a single PC could carry. It arrives alongside Nvidia’s IFA push for local AI agents, offering a living-room answer to cloud dependence.
