NVIDIA Releases Personal AI Router, an Open-Source Tool That Pools Local AI Inference Across a Home Network
NVIDIA's open-source PAIR routes multi-agent inference requests across paired local machines running Ollama or LM Studio, without pooling GPUs or changing agent code.
Overview
NVIDIA has released Personal AI Router (PAIR), an open-source tool that routes independent AI inference requests across multiple machines on the same local network. According to NVIDIA’s developer blog, PAIR “is … a virtual inference router that maximizes the AI compute in your home,” and explicitly “is not a new inference engine” — Ollama or LM Studio still executes the model, just on whichever paired machine PAIR selects.
What We Know
The problem PAIR addresses is described in similar terms by both NVIDIA and outside coverage. As MarkTechPost put it, “Multi-agent workflows have changed the shape of local inference. A lead agent decomposes a task and spawns subagents. What looked like one user request becomes dozens of independent model calls. Pointed at a single local engine, those calls compete for the same execution slots. The queue grows while a workstation, laptop, or DGX Spark on the same network sits idle.” NVIDIA frames the same shift on its blog: “AI agents are learning to do more by working together. A lead agent can break a complex task into smaller jobs and assign those jobs to specialized subagents.”
According to the project’s README on GitHub, PAIR “is a local inference router for a group of compatible computers on the same network. It discovers participating nodes, manages supported inference engines, and presents Ollama-compatible and OpenAI-compatible proxy endpoints to applications and agents.” Because the endpoints match what agent tooling already expects, NVIDIA notes there is “No new API: PAIR proxies compatible Ollama and LM Studio interfaces,” meaning existing agent harnesses require no code changes to take advantage of another machine’s GPU.
PAIR explicitly does not turn a cluster into one giant accelerator. The README states in bold terms that PAIR “routes each independent request to one node. It does not pool GPU memory, combine GPUs into a larger logical GPU, shard one model across machines, or split an in-flight inference request between nodes.” MarkTechPost describes this as “workload-level concurrency,” adding that a scheduler “weighs five signals” per request: whether a node is online and ready, whether a supported engine is enabled, whether the exact requested model is present, current job load on the node and engine, and existing GPU utilization.
Setting up a cluster relies on a manual trust step. According to PAIR’s getting-started documentation, “On the inviting system, PAIR displays a six-digit PIN and sends an invitation,” and the invited machine’s user must accept the invitation and “enter the PIN shown on the inviting system.” The documentation warns that this “PIN is a temporary bootstrap code, not a durable high-entropy credential,” so pairing should only happen “while both devices and the local network are trusted.” Once paired, NVIDIA says “communications are secured with MTLS and generated certificates so that the communications between the nodes stay private on the network,” and discovery of nearby machines happens automatically “via local-network discovery (mDNS),” with manual IP entry as a fallback.
NVIDIA demonstrated PAIR’s effect on a multi-agent workload using a companion application called Hermes Desktop, where, per NVIDIA’s blog, “Hermes owns decomposition, delegation and synthesis” while “PAIR owns inference routing.” MarkTechPost describes the demo as “a five-subagent workload over a synthetic household inbox,” run with Ollama executing the Qwen 3.6 35B A3B model on each selected node. Both NVIDIA and MarkTechPost report the same result: on a single RTX Spark laptop, the workload took 18 minutes on average, while a three-device PAIR cluster made up of an RTX Spark laptop, a DGX Spark, and an RTX 5090 completed it in an average of 8 minutes and 48 seconds. NVIDIA labels this an “unofficial, configuration-specific demonstration,” cautioning that “results depend on workload parallelism, model, engine settings, hardware, network, and node availability.”
PAIR supports Windows 11, Linux, and macOS, on both x64 and arm64 architectures, though the README notes that Windows on ARM is experimental. Installers are provided as a Windows .exe, a Linux .deb package, and a macOS .dmg, and “Windows, Linux, and macOS nodes can all be paired with each other.” NVIDIA’s blog lists the supported inference hardware as GeForce RTX 20 Series GPUs and newer, RTX PRO workstation GPUs from the Turing architecture onward, DGX Spark, and Apple M4 or newer silicon. The project is released under the Apache 2.0 license, and according to GitHub’s release history, the current version, v0.1.1, was published on August 28, 2026, following the initial v0.1.0 release on August 26, 2026.
What We Don’t Know
NVIDIA has not published independent, third-party benchmark results beyond its own labeled “unofficial” demonstration, so how PAIR performs across a wider range of hardware mixes, model sizes, and network conditions remains untested outside that single scenario. Long-term plans for the project, including a timeline toward a stable 1.0 release, have not been disclosed in the sources reviewed.
Analysis
PAIR arrives as multi-agent tooling increasingly assumes that a single machine can field many concurrent model calls at once, an assumption that breaks down on consumer and prosumer hardware where one GPU quickly becomes a bottleneck. By deliberately avoiding request-splitting or GPU-pooling — routing whole requests to whichever paired node is ready, rather than trying to merge multiple GPUs into a larger virtual one — NVIDIA is targeting a narrower and arguably more tractable problem than distributed model serving at data-center scale: making idle compute already sitting on a home or small-office network usable without asking developers to touch their agent code.