On September 4, 2026, NVIDIA officially released the v0.1.1 public beta version of the open-source virtual inference router Personal AI Router (PAIR). This tool aims to address the queueing bottleneck when a single device handles AI requests. It automatically discovers and schedules compatible nodes in the network via mDNS, distributing local AI requests to hardware nodes such as RTX, DGX Spark, and Mac. PAIR is compatible with GeForce RTX 20 series and higher graphics cards, RTX Pro GPUs, DGX Spark, and Apple M4 chips, and supports Windows, Linux, and macOS systems. Its core mechanism involves dynamically allocating computing power based on the status of user devices: when it detects that a user is using a computer (such as playing games), it automatically exits the network to ensure performance; when idle, it joins the cluster. PAIR uses Ollama and LM Studio interfaces, without the need for modifications to existing…
Nvidia has released the open-source tool Personal AI Router (PAIR), which synchronizes compatible computers in a home network and combines their idle computing power for local AI inference and agent-based workloads. Devices compatible with PAIR include Nvidia GeForce RTX 20 and newer graphics cards, RTX Pro GPUs, and DGX Spark, as well as Apple M4 and higher chips. PAIR will decide whether to utilize the computing power of devices based on whether they are idle. If a user is playing games on their computer, the device will exit PAIR’s network. Combining the computing power of multiple devices prevents GPU bottlenecks from affecting inference performance on a single device. PAIR assigns all devices with a six-digit code and uses mTLS (Mutual Transfer Layer Security) to protect the communication channel, establishing a secure encrypted connection between computers. PAIR supports Windows, Linux, and macOS systems.
NVIDIA has released the open-source virtual inference router Personal AI Router (PAIR), designed to distribute local AI requests to RTX, DGX Spark, and Mac nodes to address single-engine queueing bottlenecks. As a public test version v0.1.1, PAIR automatically discovers network machines via mDNS and schedules requests based on model location, load, and GPU utilization, without requiring modifications to existing toolchains. It supports hardware such as GeForce RTX 20 series and above, DGX Spark, and Apple M4, and uses mTLS to ensure communication security, without merging VRAM or splitting models. Demos show that on a three-node cluster including RTX Spark, DGX Spark, and RTX 5090, the average time required to handle a five-proxy workload was reduced from 18 minutes per machine to 8 minutes and 48 seconds.