NVIDIA introduces PAIR for local agent inference across PCs

NVIDIA has introduced PAIR, an open-source router that distributes local AI inference across compatible PCs. The IFA 2026 update also covers faster llama.cpp and vLLM inference, streamlined local-agent setup, and RTX Spark Windows PCs planned for October. Teams should test routing, device membership and data boundaries before relying on distributed local inference.

Local agents get a larger toolkit

NVIDIA has announced a set of local AI updates at IFA 2026 centred on making agents easier to install, faster to run and able to use more than one computer on a local network. The September 3 announcement spans software and hardware: simplified agent setup, updated inference backends, NVIDIA PAIR and RTX Spark Windows PCs expected in October.

The common thread is reducing the friction of running capable models near the user. Local systems can offer lower latency, more direct control of data and the ability to keep some work off a cloud service. They still need good hardware and careful setup, but the updates aim to make that setup more accessible to developers, creators and people experimenting with autonomous workflows.

PAIR distributes work across available machines

The headline addition is NVIDIA Personal AI Router, or PAIR. NVIDIA describes it as free, open-source software that discovers compatible PCs on a local network and directs independent inference requests to a system with available capacity. It works with Ollama and LM Studio, and can adapt as devices enter or leave the network.

This is especially relevant for agentic work, where a complex request can be divided into smaller jobs. Rather than waiting for every subtask to queue behind one GPU, PAIR can distribute independent work across idle capacity. NVIDIA gives the example of a planning workflow split among subagents, but the same pattern could apply to local research, document processing or code-analysis tasks where inputs are appropriate for local handling.

Compatibility and availability

The PAIR beta is available for Windows, macOS and Linux through graphical and terminal interfaces. NVIDIA says it supports GeForce RTX 20 Series GPUs and newer, RTX PRO workstation GPUs based on Turing or newer, DGX Spark and Apple M4 or newer silicon. That broad platform coverage makes it a potentially useful layer for households and small teams with mixed hardware.

Compatibility does not mean every machine will deliver the same experience. Network speed, GPU memory, model choice and the shape of the workload will still influence results. Users should start with a small task, observe how it is divided and verify that local routing does not create unexpected data movement or performance bottlenecks.

Faster inference underneath

NVIDIA also reports new inference improvements in llama.cpp and vLLM. The company says llama.cpp can provide up to 1.9 times higher throughput on a GeForce RTX 5090 through kernel improvements, speculative decoding and faster prefill. For vLLM, it cites gains on an RTX PRO 6000 Blackwell Workstation Edition and on two DGX Spark clusters.

Those figures are hardware- and workload-specific, so they should be treated as directional rather than a promise for every configuration. The practical message is that the local-agent experience depends as much on the runtime as the model. Faster prefill and decoding can make multi-step interactions feel more responsive, particularly when an agent needs to process long context before acting.

Simpler local setup

The announcement also points to easier model configuration in Hermes Agent, OpenClaw and Perplexity Portable Computer. In the OpenClaw case, Windows RTX systems with at least 24GB of VRAM can use a more guided route to an optimised local model. This is distinct from PAIR: onboarding gets an agent running on one machine, while PAIR helps use a group of machines once local inference is in place.

For organisations evaluating local agents, that division suggests a phased approach. First establish a safe, supportable installation and a known model configuration. Then test whether distribution adds enough capacity or responsiveness to justify the extra operational complexity. The right answer will depend on the tasks, data sensitivity and the systems already available.

RTX Spark arrives in October

NVIDIA says new RTX Spark Windows PCs from Lenovo and Acer are scheduled to arrive in October. The platform combines a Blackwell GPU, Grace CPU and up to 128GB of unified memory in devices intended for creators, gamers and always-on agents. NVIDIA also connects the hardware to the Windows Agent framework and operating-system-level controls.

Together, PAIR, faster runtimes and RTX Spark present a more complete local-agent stack: a way to install, run, accelerate and scale work across devices. The useful next step is practical testing, with explicit permissions and workload boundaries. Local AI can broaden control, but it works best when the user understands where models run, which machines participate and what an agent is allowed to do.

How to evaluate a local cluster

Before routing an important workflow across machines, test the system with representative but non-sensitive inputs. Confirm which computers PAIR discovers, which runtime receives each request and what happens when one device becomes unavailable. Monitor latency, memory use and network traffic as well as model output quality. A small local cluster can improve responsiveness, but it also introduces new questions about updates, device ownership and data boundaries. Establish a simple policy for enrolling and removing machines, and ensure every participant runs a supported software version. Where tasks are divided among subagents, give each one a narrow scope and retain a review step before any external action. These controls make local distribution a deliberate engineering choice rather than an invisible convenience feature. They also help teams learn whether multiple PCs deliver a material benefit for the workloads they actually run.