The PAIR beta connects local models on PCs and Macs. It can distribute concurrent requests, but it cannot pool GPU memory for one large model.
If you have several computers running AI models locally, NVIDIA wants to route requests among them through one application. Its PAIR beta arrived on September 3 and works with Ollama and LM Studio, two tools for running models on your own hardware. NVIDIA’s announcement describes the setup.
An application sends a request to one endpoint. PAIR then selects a computer to handle it. That machine must already have the required model and a running instance of Ollama or LM Studio.
Where it helps
A team might have several AI agents sending requests to a local model at the same time. PAIR can send those independent requests to different computers, allowing each to process part of the queue.
PAIR does not pool GPU memory. If a large model cannot fit on any one machine, adding two more PCs will not make it fit. One computer handles each request, and a single sequential task is not guaranteed to run faster. NVIDIA’s FAQ explains the limitation.
Hardware and operating systems
PAIR supports Windows, macOS and Linux. NVIDIA lists GeForce RTX 20-series and newer GPUs, DGX Spark and Macs with M4 chips or newer among the hardware it has validated. Check a specific configuration against the beta requirements before installing it.
Local inference does not require an internet connection, although downloading models does. We have not benchmarked PAIR, so its effect on a particular team’s throughput remains to be measured.