Nvidia releases open-source PAIR tool for local compute pooling

Nvidia's newly announced Personal AI Router (PAIR) is free, open-source software designed to link idle home computers for local AI inference using tools like LM Studio and Ollama. [1] The system connects compatible PCs on a network to process agentic workflows in parallel, which breaks complex tasks into smaller jobs and prevents single-GPU bottlenecks. [1] PAIR functions with Apple M4 processors and newer, as well as Nvidia RTX 20-series, RTX Pro, and DGX Spark systems. [1] The software uses in-home hardware only when it is idle and can adapt if devices leave the network, such as when a user begins playing a PC game. [1]

Recent tests on the DGX Spark evaluated the performance of the Qwen3.8-Flash-Next model utilizing Nvidia's official NVFP4 quantization on upstream vLLM. [2] A single 128 GB DGX Spark achieved a median speed of 32.5 tokens per second and a single-stream peak of 43.8 tokens per second. [2] Linking two Sparks yielded a median speed of 53.7 tokens per second and a peak of 63.7 tokens per second when utilizing a specific speed profile with a 4096-token prefill chunk. [2] Developers noted that linking four Sparks required enabling expert parallel settings because the intermediate size could not be padded four ways. [2]

Nvidia is expanding its software ecosystem to maximize the utility of distributed local hardware for AI workloads. By allowing Apple silicon to participate in PAIR clusters, Nvidia acknowledges the mixed-hardware reality of consumer networks while ensuring its own GPUs, like the RTX series and DGX Sparks, remain central to inference scaling. The forum benchmarks indicate that linking high-end local nodes yields tangible token generation improvements for heavily quantized models running on standard frameworks like vLLM. What the sources don't address: Whether PAIR introduces measurable latency overhead when routing small tasks across disparate hardware architectures and network configurations.