Nvidia has released Personal AI Router (PAIR), a free, open-source software tool that connects multiple computers on your home network into a shared local inference pool — letting you run larger language and image models than any single machine could handle alone.
Key takeaways
PAIR is software, not hardware. Despite its name, it is an open-source application, not a physical network router.
It works with Ollama and LM Studio, two of the most common local model runners used by AI-art creators today.
Idle machines contribute compute. A laptop that is plugged in but not in use can feed GPU and CPU cycles to a heavier inference job running on a different machine.
No cloud fees, no data leaving your network. All inference stays local, which matters for creators working with proprietary styles or sensitive reference images.
RTX GPUs get priority scheduling, but the tool can also use CPU-only machines as secondary nodes.
How PAIR distributes inference across machines
PAIR acts as a local orchestration layer — software that sits between your model runner (Ollama, LM Studio, or a compatible API endpoint) and the models themselves. When you send a prompt, PAIR checks which connected machine has available compute and routes the job there, or splits it across several nodes if the model supports it. Think of it as a traffic director for your home network's unused GPU cycles.
The practical upshot for AI-art creators is meaningful. Running a 70-billion-parameter text model locally today typically requires either a single GPU with 40 GB or more of VRAM — expensive hardware — or accepting slow, CPU-only generation. PAIR lets a creator with, say, an RTX 4070 desktop and an older RTX 3060 laptop combine their VRAM budgets, making previously unrunnable model sizes accessible without a cloud API bill.
Nvidia PAIR's graphical interface displaying connected home machines and active inference routing.
What the Ollama and LM Studio integration means in practice
Both Ollama and LM Studio expose a local OpenAI-compatible API endpoint — a standardized interface that many image-generation pipelines and prompt tools can already talk to. PAIR slots in at that same layer, so any workflow already pointed at a local Ollama server should work with PAIR in front of it without rewriting prompts or changing model configs. The transition is closer to a settings change than a workflow overhaul.
A prompt routed through PAIR to the Hermes model, showing how the tool distributes the inference job.
For creators running image-generation pipelines that use a local LLM for prompt expansion or caption refinement — a common pattern in ComfyUI workflows — this could mean faster turnaround on that preprocessing step, since a second machine's GPU handles it while the primary GPU stays focused on diffusion.
According to The Verge, PAIR also supports MacBooks as nodes, which broadens the pool of machines a creator might already own and can contribute to the cluster.
The architecture of Nvidia PAIR, illustrating how desktop, laptop, and other devices form a unified local compute pool.
RTX hardware context and what non-Nvidia users should know
Nvidia naturally optimizes PAIR's scheduling for its own RTX GPUs — the same hardware family covered in the Nvidia DLSS 5 launch earlier this year. RTX cards benefit from CUDA-accelerated inference, so they will process tokens and latents faster per watt than CPU-only nodes. That said, CPU-only machines are not excluded; they join the pool as lower-priority contributors, which is useful for offloading smaller preprocessing tasks.
AMD and Intel GPU users are in a less clear position. PAIR's open-source nature means community drivers for ROCm or OpenCL may emerge, but Nvidia has not confirmed first-party support for competing GPU architectures at launch.
For creators already exploring local model running through the Charmloop guides, PAIR represents a concrete path to scaling up without new hardware purchases. The ceiling on what you can run locally just moved — how far depends on how many machines you have sitting idle on your network.
