NVIDIA PAIR: Open Source Beta Routes Local AI Inference Across RTX and Mac Nodes

Multi-agent workflows turn one user request into dozens of independent model calls. When all those calls hit a single local engine, they queue up while other capable machines on the same network sit idle. NVIDIA Personal AI Router (PAIR) targets exactly that bottleneck.

PAIR is a virtual inference router that discovers compatible systems on your local network and routes each independent inference request to the best available node. It is not a new inference engine; Ollama or LM Studio still runs the model on whichever machine PAIR selects.

No cluster API, no agent changes

PAIR ships as a public beta (v0.1.1) with signed installers for Windows, macOS, and Linux, and the full source is on GitHub under Apache 2.0. It runs entirely on the local network, needing internet only to download models. The key design decision: PAIR introduces no cluster API. It proxies the Ollama-compatible and LM Studio-compatible interfaces that agents already speak. If an agent harness listens on a different port, the proxy port is configurable. The result is that existing agent harnesses require zero modifications. The agent decides what to request; PAIR decides where it runs.

How PAIR discovers, secures, and schedules

PAIR uses mDNS to find nearby systems automatically. If discovery fails, a node can be added by IP address. Trust is established with a six-digit PIN shown on the inviting machine and entered on the invited one. All node-to-node communication is blocked until pairing completes, and traffic is then secured with mTLS using generated certificates.

Each node runs Ollama or LM Studio. PAIR can install the engine and start model downloads on paired systems, reducing cross-machine setup work. A node is eligible for a request only when the required engine is enabled and the requested model is present. Models do not need to match across the cluster; PAIR routes according to model location.

For each request the scheduler weighs five signals:

• Is the node online and ready?
• Is a supported engine enabled?
• Is the exact model present?
• What is the current node and engine job load?
• What is existing GPU utilization?

This is workload-level concurrency. PAIR assigns each request to exactly one eligible node where it stays for its lifetime. It does not pool VRAM, merge GPUs, or shard a single request across machines.

Hardware support and real-world demo

Supported hardware includes GeForce RTX 20 Series and newer, RTX PRO workstation GPUs, DGX Spark, and Apple M4 or newer silicon. Nodes can run Windows, Linux, or macOS on x64 and arm64 (Windows on ARM is experimental).

NVIDIA demonstrated PAIR with Hermes Desktop creating a five-subagent workload over a synthetic household inbox, using Ollama to execute Qwen 3.6 35B A3B on each selected node. On a single RTX Spark laptop the workload averaged 18 minutes. On a three-device cluster (RTX Spark laptop, DGX Spark, and RTX 5090) it averaged 8 minutes 48 seconds, roughly halving the total completion time.

What this means for builders

If you run multi-agent local workflows and have multiple capable machines in your home or small office, PAIR directly addresses the queueing problem. It saves the effort of manually distributing models or building a custom scheduling layer. The no-cluster-API approach means you can plug it into existing agent harnesses and see throughput gains immediately.

However, PAIR cannot accelerate a single large model that exceeds one GPU's VRAM, and it does not parallelize a single request across GPUs. For workloads that are already parallel at the request level, such as multi-agent orchestration, it is a practical and lightweight solution. As a v0.1.1 beta, expect rough edges, especially on Windows on ARM.