Multi-agent AI workflows have created a mess for local inference. A single user request gets decomposed into dozens of independent model calls, all competing for slots on one machine while other capable hardware sits idle on the same network. NVIDIA’s answer is PAIR — the Personal AI Router — now shipping as a public beta that spreads those requests across whatever compatible machines happen to be nearby.
PAIR isn’t an inference engine. It’s a scheduler that proxies the Ollama and LM Studio interfaces agents already use, meaning existing harnesses don’t need a single line of code changed. The agent decides what work to request; PAIR decides where it runs. Discovery happens over mDNS, trust gets established with a six-digit PIN, and node-to-node traffic is then locked down with mTLS. The whole thing runs locally, with internet only needed for model downloads.
Eligibility for routing is refreshingly strict. A node only gets a request when the exact model is present, the engine is enabled, and the scheduler has weighed current job load and GPU utilization. What PAIR deliberately doesn’t do is pool VRAM or shard a single request across machines. That’s the right call for agent workloads, which are embarrassingly parallel by nature, but it also means anyone expecting a poor man’s NVLink cluster will be disappointed.
The demo numbers come with caveats. NVIDIA paired PAIR with Hermes Desktop running a five-subagent workload over a synthetic inbox, executing Qwen 3.6 35B A3B. One RTX Spark laptop averaged 18 minutes. A three-device cluster — laptop, DGX Spark, RTX 5090 — cut that to 8:48. Unofficial, sure, but the direction is what matters. The scheduler currently uses one policy and ignores VRAM capacity, GPU class, and model warmness, which leaves performance on the table for a v0.1.1. Still, the design decisions — no new API, Apache 2.0, signed installers for Windows, macOS, and Linux — suggest NVIDIA understands that adoption lives or dies on friction. If PAIR can evolve its scheduler without breaking the proxy model, it could quietly become the default way home labs handle multi-agent inference.
