Nvidia launches PAIR, a local inference router that spreads workloads across workstations

The problem Nvidia is tackling with PAIR — Personal AI Router — is familiar to anyone running multi-task agents on local hardware: a lead agent breaks a job into dozens of independent model calls, and all of them pile up in a single inference engine's queue. While your GPU runs hot, the DGX Spark in the corner or your colleague's MacBook sits idle. PAIR, announced this week as public beta 0.1.1, is a virtual router that discovers compatible machines on the home network and routes individual inference requests to them, without replacing the inference engine itself. Ollama or LM Studio still run the model on the chosen node; PAIR only decides where each request goes.

No new API, just a proxy for existing interfaces

The most significant architectural decision is that PAIR does not expose a new cluster API. Instead it acts as a proxy to the Ollama-compatible and LM Studio-compatible interfaces agents already speak, capturing each engine's default port. If the harness listens on a different port, an alternative proxy port can be configured in PAIR's engine settings. The repository also exposes OpenAI-compatible endpoints. The result: existing agents need zero code changes. The agent decides *what* to ask for; PAIR decides *where* it runs.

Discovery, pairing, and secured traffic

Discovery happens automatically over mDNS, with a manual IP-address fallback when discovery fails. Trust is established through a six-digit PIN displayed on the inviting machine and entered on the invited one; all inter-node traffic is blocked until pairing completes. After that, traffic is secured with mTLS using locally generated certificates. Every node runs Ollama or LM Studio, and PAIR can install the engine and pull models on paired systems, eliminating most of the manual cross-machine setup.

The scheduler: five signals, no VRAM pooling

A node becomes eligible for a request only when the required engine is running and the exact requested model is present. Models do not need to be identical across the cluster — different systems can hold different models — and PAIR routes based on model location. Loading the same tag on multiple nodes simply expands the eligible pool. For each request the scheduler weighs five signals: node availability, supported engine running, exact model presence, current task load on the node and engine, and existing GPU utilization. This is workload-level concurrency, and the boundary is sharp: PAIR assigns each request to a single eligible node for its entire lifetime. It does not pool VRAM, merge GPUs into one large accelerator, or shard a single request across machines.

Demo numbers, with caveats

In Nvidia's demo, PAIR was paired with Hermes Desktop generating a load of five sub-agents against a synthetic inbox. Ollama ran Qwen 3.6 35B A3B on every selected node. On a single RTX Spark laptop the workload took 18 minutes on average. On a three-device cluster — RTX Spark laptop, DGX Spark, and RTX 5090 — the average dropped to 8 minutes 48 seconds. These figures reflect one specific synthetic workload; real-world performance depends on network topology, hardware differences, and the nature of the models.

Supported hardware and minimum requirements

PAIR supports GeForce RTX 20-series and newer, RTX PRO workstation cards from Turing onward, DGX Spark, and Apple M4 silicon and newer. Windows, Linux, and macOS nodes can pair with each other on both x64 and arm64 (Windows on ARM is experimental). Verified configurations require at least 8 GB RAM and 20 GB disk is recommended. Other Linux distributions build from source. The full code is on GitHub under Apache 2.0, and installers are signed for Windows, macOS, and Linux.