Nvidia has released a beta tool that lets local AI applications send separate model requests to different computers on the same network. Personal AI Router, or PAIR, is designed for agent workflows where a coordinator launches several independent tasks and one GPU becomes a bottleneck.
PAIR sits between an application and local inference engines such as Ollama or LM Studio. It checks each request’s model and engine requirements, selects a compatible node, and returns the result through the same connection the application expects. Nodes can run Windows 11, Linux or macOS, including a mix of x64 and Arm64 systems.
The tool does not merge GPUs or pool their memory into one larger accelerator. Each request runs from start to finish on a single eligible machine, so PAIR helps parallel workloads rather than models too large for any one node.
In an Nvidia demo using an RTX Spark, DGX Spark and RTX 5090, five delegated analyses completed in roughly half the time required on one RTX Spark laptop. Nvidia cautions that this is not a performance guarantee: results depend on workload parallelism, models, settings, networking and node availability. PAIR is available from GitHub, giving developers a practical way to use otherwise idle local machines without rewriting an agent harness.