NVIDIA has released the Personal AI Router (PAIR), an open-source virtual inference router designed to distribute local AI model requests across multiple devices on a local network. Published under the Apache 2.0 license on GitHub, PAIR ships as a public beta with desktop installers for Windows, macOS, and Linux.
Rather than acting as a standalone inference engine or pooling VRAM across machines, PAIR acts as a proxy for existing interfaces like Ollama and LM Studio. It uses mDNS to discover local nodes—including GeForce RTX GPUs, DGX Spark systems, and Apple M-series Mac devices—and automatically routes incoming multi-agent subtasks to idle systems based on model availability, GPU load, and network capacity.
In NVIDIA benchmarks, a five-subagent task that took 18 minutes on a single laptop completed in under 9 minutes when distributed across a three-device PAIR cluster. The tool requires no modifications to existing agent harnesses or frameworks.
Why it matters
Enables local multi-agent systems to scale workloads across idle desktop, server, and Mac hardware.
Requires zero code changes for applications already compatible with Ollama, LM Studio, or OpenAI APIs.
Provides a cost-effective alternative to cloud inference by maximizing existing local hardware utilization.
Source: marktechpost.com
