NVIDIA PAIR Routes Local Inference Across Home Devices to Ease Multi-Agent Bottlenecks

NVIDIA released the beta of its Personal AI Router (PAIR), a virtual inference router that distributes independent inference requests across compatible systems on a local network. PAIR is not a new inference engine; Ollama or LM Studio still runs each model on the selected machine. The router discovers participating systems, tracks readiness, schedules jobs, and returns responses to the originating application. It proxies the Ollama and LM Studio interfaces, so agent harnesses do not need changes.

How PAIR works

PAIR uses mDNS for local-network discovery, with the option to add a node by IP address. Users approve a secure pairing request, and all node-to-node communication is blocked until pairing is established. Once connected, traffic is secured with MTLS and generated certificates. Each participating node runs a supported local inference engine, and PAIR can help install an engine and initiate model downloads on paired systems. A node becomes eligible for a request only when the required engine is enabled and the exact requested model is present. Models do not have to be identical across nodes; different systems can host different models, and PAIR routes based on model location.

The router takes over the default port that Ollama and LM Studio use, and the proxy port can be configured in PAIR engine settings if the agent harness uses a different port. PAIR inspects the request's engine and model requirements, then passes those to the router. The agent decides what work to request, while PAIR decides where eligible work runs. The scheduler filters paired systems using readiness, engine state, model presence, and job load, then selects one eligible node. Independent calls from other subagents can be assigned to other ready nodes simultaneously. The selected engine executes the request, and PAIR streams the response back through the same local interface. Jobs and metrics views show which node handled each request.

Elastic home clusters

PAIR is designed for dynamic home hardware, not dedicated data centers. Client nodes can join the available pool when ready and drop away when needed, such as when a laptop sleeps or a GPU is reclaimed for a foreground application. For each request, PAIR considers whether a paired node is online, whether a supported engine is enabled, whether the exact model is present, the current node and engine workload, and existing GPU utilization. It does not depend on every system being identical or permanently available.

Demo results

In a demonstration with Hermes Desktop creating a five-subagent workload and Ollama executing models, a single NVIDIA RTX Spark laptop took 18 minutes on average to complete the task using Qwen 3.6 35B A3B. A three-device PAIR cluster containing an RTX Spark laptop, a DGX Spark, and an RTX 5090 took 8 minutes and 48 seconds on average. NVIDIA notes this is an unofficial, configuration-specific demo, not a general benchmark or a promise of linear scaling. Results depend on workload parallelism, model, engine settings, hardware, network, and node availability. The Jobs view in PAIR is the ground truth for where inference ran; multinode execution should be claimed only when PAIR telemetry shows jobs on more than one eligible node.

PAIR supports NVIDIA GeForce RTX 20 Series GPUs and newer, RTX PRO workstation GPUs (Turing architecture and newer), DGX Spark, and Apple M4+ silicon. The beta is available for Windows, macOS, and Linux through graphical and terminal interfaces. The project is open source, with a repository for code inspection, issue reporting, and contributions.