NVIDIA PAIR

// Overview

PAIR — the Personal AI Router — is NVIDIA's Apache-2.0 tool for using more than one machine to serve local AI requests. It runs no models itself. Instead it discovers the compatible computers on the same network, keeps track of which inference engines and models each one has, and exposes Ollama-compatible and OpenAI-compatible endpoints that any existing client can point at. Each incoming request is routed to a single node that can serve it, based on which systems have capacity.

The important limit to understand up front is what PAIR does not do: it does not pool VRAM and it does not shard a model across machines. A model still has to fit on one node. What PAIR buys you is parallelism — when an agent fans a task out into several independent jobs, those jobs can land on different GPUs instead of queueing behind one. NVIDIA's own framing is an assistant that splits inbox triage across subagents while PAIR distributes the work over the PCs already sitting idle.

It ships as a public beta written in Go, with signed installers for Windows 11, Linux and macOS on both x64 and arm64, and mixed clusters are supported — a Windows desktop and a Mac can pair together. Supported hardware is GeForce RTX 20 Series and newer, RTX PRO workstation GPUs from Turing onward, DGX Spark, and Apple M4 or newer silicon. The supported inference engines at launch are Ollama and LM Studio.

// What it does

• Automatic discovery of compatible machines on the local network, with pairing confirmed by a six-digit PIN
• Ollama-compatible and OpenAI-compatible endpoints, so existing agent and client code does not have to change
• Per-request scheduling to a single node based on which systems have an enabled engine, the model in place, and free capacity
• Heterogeneous clusters — Windows, Linux and macOS nodes can be paired with each other
• Ollama and LM Studio installed and managed per node from PAIR's own Engine settings
• Graphical and terminal interfaces, with .exe, .deb and .dmg installers

// Getting started

PAIR is installed per machine, then the machines are paired with each other. Start with a single node and confirm a request works locally before adding a second one.

sudo apt install ./NVPAIR-Setup-*.deb

curl http://127.0.0.1:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen4:12b","messages":[{"role":"user","content":"In one sentence, what does a router do?"}]}'

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Run an agent that fans a task out to several subagents without every job queueing on one GPU
• Put an idle desktop or a DGX Spark to work serving requests from a laptop that is short on VRAM
• Keep an entire local AI setup offline while still using more than one machine's worth of compute
• Give a household or small team one endpoint in front of several Ollama and LM Studio installs

// How NVIDIA PAIR compares

NVIDIA PAIR alongside other open-source gateways & routing tools AI/TLDR tracks, ranked by GitHub stars.