AINEWS
PAIR routes independent AI jobs across the PCs on your home network. It does not pool GPU memory, shard a model, or speed up a single machine, so it is worth installing only if you already own more than one.
PAIR, short for Personal AI Router, is a free, open-source tool from Nvidia that splits AI work across the PCs already sitting on your home network. Nvidia announced it at IFA 2026 in Berlin, and the code is on GitHub under an open-source license.
PAIR's job is request routing. When a local AI app makes a request, PAIR looks at every machine on the home network and routes the request to the one that has the right model loaded and the spare compute to handle it. Each request goes to a single node, runs there, and comes back. PAIR does not pool GPU memory across machines, does not combine GPUs into a larger logical GPU, does not shard a single model across machines, and does not split an in-flight request between nodes. The release notes for v0.1.0 and the v0.1.1 build follow-up confirm the shipping scope.
Two consequences follow from that limit. If one PC can already run a model, PAIR will not make that model faster. If one PC cannot run a model because it does not fit in that machine's VRAM, PAIR will not make it fit across several PCs. PAIR helps with throughput on independent jobs, for example a multi-agent workflow where several sub-tasks run at once, or a household where a few machines are idle while one is busy.
PAIR exposes Ollama-compatible and OpenAI-compatible proxy endpoints to apps and agents, so existing local-AI toolchains can point at it without code changes. The privacy benefit is concrete: when every client, model source, engine, and node lives on the local network, prompts and responses never leave the house. That is a property of the deployment, not a guarantee the router enforces on its own.
Alongside PAIR, Nvidia said local inference on RTX hardware is up to 1.9x faster through new llama.cpp and vLLM optimizations, with the gains shipping now in LM Studio and Ollama. Compact RTX Spark desktops from Lenovo and Acer land in October. Three agent products (Hermes Agent, OpenClaw, and Perplexity Portable Computer) will get simplified local-AI support on Nvidia GPUs.
Who should install PAIR today: anyone who already runs a local model on at least two PCs and wants those machines to share a queue. Who should skip it: anyone expecting it to behave like Nvidia's DGX Spark, a separate clustering product designed to pool memory across workstations. The GitHub repo walks through setup, and a v0.1.1 build is already in the release log for users on Intel and AMD x64 systems.