NVIDIA Releases Open-Source PAIR Router to Distribute Local AI Requests Across Devices
NVIDIA has released PAIR, an open-source virtual inference router that routes local AI requests across RTX, DGX Spark and Mac nodes, cutting a five-subagent demo task from 18 minutes on one laptop to 8:48 on three devices.
Multi-agent workflows have changed local inference. A lead agent decomposes a task and spawns subagents, turning one user request into dozens of independent model calls. When pointed at a single local engine, those calls compete for execution slots while other capable machines on the same network sit idle, MarkTechPost explained.
PAIR is not a new inference engine. It discovers compatible machines on a home network and schedules independent inference requests across them. Ollama or LM Studio still executes the model on whichever node PAIR selects. It runs entirely on the local network, with internet needed only to download models.
PAIR ships as v0.1.1 with signed installers for Windows, macOS, and Linux, and its full source code is on GitHub under the Apache 2.0 license. Node discovery uses mDNS, and systems can be added by IP address when discovery fails. Trust is established by a six-digit PIN, and traffic between paired nodes is secured with mutual TLS.
The design decision that matters most, according to the report, is that PAIR introduces no cluster API. It proxies Ollama-compatible and LM Studio-compatible interfaces agents already speak, taking over default ports. The repository also exposes OpenAI-compatible proxy endpoints. As a result, existing agent harnesses require no changes; the agent decides what work to request and PAIR decides where it runs.
Each node runs Ollama or LM Studio, and PAIR can install an engine and start model downloads on paired systems. A node is eligible for a request only when the required engine is enabled and the exact requested model is present. For each request, the scheduler weighs five signals: online and ready status, enabled engine, exact model presence, current node and engine job load, and existing GPU utilization. PAIR assigns each request to one eligible node; it does not pool VRAM, merge GPUs, or shard a single request across machines.
NVIDIA's demonstration paired PAIR with Hermes Desktop, which creates a five-subagent workload over a synthetic household inbox. Ollama executed Qwen 3.6 35B A3B on each selected node. On one RTX Spark laptop, the workload took 18 minutes on average; on a three-device PAIR cluster consisting of an RTX Spark laptop, a DGX Spark, and an RTX 5090, it averaged 8 minutes and 48 seconds. MarkTechPost noted the demo numbers are unofficial.
PAIR supports GeForce RTX 20 Series and newer, RTX PRO workstation GPUs from Turing onward, DGX Spark, and Apple M4 or newer silicon. Windows, Linux, and macOS nodes can be paired with each other on x64 and arm64, though Windows on ARM is experimental. Validated configurations list 8 GB RAM or higher and a recommended 20 GB of disk.