Nvidia PAIR: How to Route Local AI Across Multiple Home Machines
Nvidia PAIR is a free open-source router that spreads Ollama and LM Studio requests across GPUs on your home network. Here's how it works.
What is Nvidia PAIR?
Nvidia PAIR (Personal AI Router) is a free, open-source tool that sits between your AI agents and the inference engines running on your home network, deciding which machine handles which request. It works with Ollama and LM Studio out of the box, runs on Windows, Linux, and Mac, and is released under an Apache 2.0 license, meaning anyone can fork it, modify it, or extend it. It’s currently at a 0.1 release, so treat it as an early, evolving project rather than a finished product.
TL;DR
- Nvidia PAIR is an open-source router (Apache 2.0 licensed) that distributes local inference requests across multiple machines on your home network instead of queuing everything on one GPU.
- It supports Ollama and LM Studio natively, works on Windows, Linux, and Mac, and community pull requests are already adding support for Anthropic-style APIs and llama.cpp.
- PAIR does not pool VRAM. Each request still runs entirely on one machine; adding more nodes gives you more parallel capacity, not a bigger combined memory pool for a single huge model.
- It works by taking over the ports normally used by Ollama and LM Studio, so existing apps and agents keep calling the same endpoints while PAIR quietly redirects traffic to whichever machine is free.
- The tool targets a specific bottleneck in agentic workflows: when a lead agent spins up parallel sub-agents for coding or research tasks, those sub-agents often just queue up waiting for a single local GPU instead of actually running in parallel.
- Nvidia is positioning this as part of a bigger push into local AI infrastructure, coming alongside its roughly $13 billion acquisition of Hugging Face and ahead of next month’s RTX Spark laptop launches.
- Because it’s Apache 2.0 licensed, if Nvidia is slow to accept community pull requests, the project can simply be forked by anyone who wants a different feature set.
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
How does Nvidia PAIR actually work?
PAIR functions as a proxy. When you install it, it takes over the network ports that Ollama and LM Studio normally listen on. Any app, agent, or script that used to talk directly to those engines, whether that’s a coding agent like Open Claw or an orchestration framework like Hermes, keeps calling the exact same ports as before. Nothing changes on the client side.
Behind the scenes, PAIR tracks which machines on your network are running Ollama or LM Studio, what models are loaded on each, and how much VRAM and system memory each machine has available. When a request comes in, PAIR checks which machine is free or best suited to handle it and routes the call there. If one machine is busy, PAIR sends the next request to another machine instead of letting it sit in a queue.
The setup process is straightforward: install PAIR on a machine, and it automatically detects locally installed inference engines like Ollama, showing available VRAM and currently loaded models. You then connect additional machines on the network running Ollama or LM Studio, and PAIR treats them as additional routing targets.
Why does this matter for local AI setups?
The problem PAIR is trying to solve comes from how modern agentic AI systems actually work. Tools built around a lead agent that delegates tasks to parallel sub-agents (research subtasks, coding subtasks, and so on) depend on those sub-agents running at the same time. On a single machine with a single GPU, that parallelism breaks down. Sub-agents end up in a line, waiting their turn for the one available GPU, which defeats much of the point of running an agentic system in the first place.
PAIR addresses this by treating every GPU-capable machine on your home network as a worker that can pick up part of the load. A gaming PC with an Nvidia GPU, a laptop with an RTX chip, and a workstation in another room can all be pulled into the same routing pool. Sub-agents that would have queued on one box can instead run at the same time across several.
This matters more as home hardware diversifies. A significant share of households already have more than one computer, and with RTX Spark laptops launching soon alongside other GPU-accelerated devices entering the market over the next few years, many homes will end up with multiple machines capable of running local inference, even if none of them individually match a high-end workstation.
Is Nvidia PAIR a VRAM pooling tool?
No, and this is worth being clear about. PAIR does not combine the VRAM of multiple machines into one larger virtual GPU. It does not let you load a model that’s too big for any single machine’s memory by spreading it across several devices. Each individual request still runs entirely on one machine, using only that machine’s resources.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
What PAIR adds is parallelism, not raw horsepower. Adding more machines to your PAIR network means you can handle more simultaneous requests, which is exactly what agentic workflows with multiple sub-agents need. But a single prompt sent to a single agent will not run any faster just because you’ve added more nodes. If you’re hoping to run a massive model that doesn’t fit on your best GPU, PAIR won’t solve that problem.
What can you connect to PAIR right now?
At its 0.1 release, PAIR officially supports Ollama and LM Studio as inference engines, and works across Windows, Linux, and Mac. On the Mac side, the tool reportedly favors Apple M4 chips and above, though it has been shown running on older hardware like an M2 Pro as well, just with more limited support. Demonstrations of the tool have shown it directing traffic to hardware ranging from Nvidia RTX 5090 GPUs to AMD Radeon graphics cards, suggesting the routing layer isn’t locked to Nvidia-only hardware despite Nvidia building it.
Community activity around the GitHub repository has moved fast for a 0.1 release. Pull requests are already in progress to add support for Anthropic-style API formatting rather than just OpenAI-compatible endpoints, and separate efforts are underway to bring llama.cpp into the supported engine list. Because the project is Apache 2.0 licensed, none of this depends on Nvidia’s approval alone. If Nvidia declines certain pull requests or moves slowly, developers are free to fork the project and maintain their own version with the features they want.
Why is Nvidia building this now?
PAIR didn’t launch in isolation. It arrived the same week Nvidia closed an acquisition of Hugging Face valued at just under $13 billion, and just ahead of the broader rollout of RTX Spark laptops from multiple hardware makers. Taken together, these moves point to Nvidia making a deliberate bet on local AI as a real market, not just a niche for hobbyists.
The reasoning tracks with broader trends in open-weight AI. Open models on Hugging Face have grown to more than four times last year’s volume so far this year, and generative AI models broadly have more than doubled. Open-weight models are also closing the gap with proprietary ones faster than expected: models like Qwen3 have already surpassed intelligence benchmark scores that flagship proprietary models were hitting just months earlier, according to Artificial Analysis intelligence index scores referenced in recent coverage. None of these open models run on phones, but many are now practical to run on capable local hardware, which is exactly the audience PAIR is built for.
Nvidia’s interest is straightforward: it sells the GPUs that make local inference possible, and a tool that makes multi-machine local setups more usable increases the value of owning more Nvidia-accelerated hardware, not less.
Is Nvidia PAIR worth setting up?
For anyone already running Ollama or LM Studio on more than one machine at home, PAIR is worth trying, with realistic expectations. It won’t make a single model run faster or let you load something too large for your best GPU. What it does offer is a low-effort way to stop parallel agent workflows from bottlenecking on one machine, which is a real and growing problem as agentic coding and research tools become more common in local setups.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Given the 0.1 version number, expect rough edges, inconsistent detection of installed engines, and gaps in hardware support. But because the project is open source and permissively licensed, the pace of improvement is likely to depend as much on community contributions as on Nvidia’s own roadmap.
Frequently Asked Questions
What does Nvidia PAIR stand for?
PAIR stands for Personal AI Router. It’s Nvidia’s open-source tool for routing local AI inference requests across multiple machines on a home network.
Does Nvidia PAIR require Nvidia GPUs?
No. While built by Nvidia, PAIR has been shown directing requests to AMD Radeon graphics as well as Nvidia GPUs, and it runs on Windows, Linux, and Mac.
Can PAIR combine VRAM from multiple machines?
No. PAIR routes individual requests to individual machines; it does not pool memory across devices. Each request runs entirely on whichever single machine handles it.
Which inference engines work with Nvidia PAIR?
Ollama and LM Studio are supported at launch. Community pull requests are in progress to add support for Anthropic-style APIs and llama.cpp.
How is PAIR different from a tool like SwitchYard?
SwitchYard routes requests across models based on cost and quality, often including cloud-hosted options. PAIR is strictly for hardware you own and control, splitting local inference traffic across your own machines rather than choosing between paid API models.
