NVIDIA released PAIR on September 3, 2026. The free open-source software turns compatible computers on the same network into a shared pool for local AI inference.
PAIR does not split one model across several machines. It watches available devices and routes independent requests to the computer that is ready to handle them. That makes it especially useful for agent workflows that run several tasks at once.
What it supports
PAIR works with Ollama and LM Studio without requiring changes to an existing agent harness. NVIDIA lists support for Windows, macOS, and Linux systems using compatible hardware, including:
- GeForce RTX 20 Series and newer
- RTX PRO systems
- DGX Spark
- Apple M4 and newer
Before routing a request, the software checks device readiness, model availability, engine status, and GPU use.
What NVIDIA measured
In NVIDIA's five-subagent demonstration, three connected devices completed the workload in 8 minutes and 48 seconds. A single RTX Spark laptop took 18 minutes.
That comparison is a demonstration, not a promise for every network. Results will depend on model size, device mix, network speed, task structure, and how many requests can run independently.
Why this matters for smaller teams
Local AI can improve privacy and reduce dependence on a metered cloud API, but one computer often becomes the bottleneck. PAIR offers a way to use hardware that is already available before buying a larger server.
A good first test is a workflow with several independent jobs, such as document classification, image tagging, or research subagents. Measure throughput, energy use, failures, and setup time before moving an important process onto the cluster.
The SideQuest Stack take
PAIR makes local AI more practical for homes, studios, and small offices with several compatible machines. Its appeal is simple: add capacity without rebuilding the workflow. The best use cases will be parallel jobs, not one huge request that needs all devices at once.
Read the NVIDIA technical announcement.