NVIDIA Announces PAIR, Faster Local AI and RTX Spark PCs

NVIDIA announced a collection of local AI updates at IFA 2026, including a tool for sharing inference workloads across nearby computers, performance optimizations for two inference backends and simplified model setup for agent applications. The company also said NVIDIA RTX Spark Windows PCs will arrive in October 2026.

Routing work across local computers

NVIDIA introduced Personal AI Router, or PAIR, as a free open-source software tool that distributes independent AI inference requests among compatible computers on a local network. According to NVIDIA, PAIR automatically discovers supported devices, selects systems with available capacity and adjusts when devices join or leave the network.

The tool works with Ollama and LM Studio and is available in beta through graphical and terminal interfaces. NVIDIA lists support for Windows, macOS and Linux, as well as GeForce RTX 20 Series and newer GPUs, RTX PRO workstation GPUs using Turing or newer architectures, DGX Spark and Apple M4 or newer silicon.

NVIDIA said PAIR can distribute jobs created by multiple subagents rather than leaving them queued on one GPU. It can also move AI workloads to another computer while a user employs the primary system for gaming, creative tasks or other work.

Simplified agent setup

NVIDIA said Hermes Agent, OpenClaw and Perplexity Portable Computer are receiving simplified support for local models on its hardware. The setup experiences use llama.cpp and include NVIDIA’s inference optimizations.

For Hermes Agent, one-click local model setup is available on Windows across RTX and DGX systems, with Linux support planned. NVIDIA said the software detects the installed GPU, chooses a model and configuration, and runs the model through an integrated llama.cpp implementation. Hermes can use tools, preserve context across tasks, retain information between sessions and create reusable skills.

Perplexity Portable Computer is available on Linux systems equipped with NVIDIA RTX GPUs carrying at least 24GB of VRAM, according to NVIDIA, and Windows support is planned. NVIDIA said the application can run workflows locally without consuming credits and can ask permission before sending content to one of more than 15 cloud models.

NVIDIA, Microsoft and OpenClaw are also working on an OpenClaw Windows application. NVIDIA said it simplifies setup of an optimized local model on RTX GPUs with at least 24GB of VRAM.

Inference performance updates

NVIDIA reported that llama.cpp can provide up to 1.9 times higher throughput on a GeForce RTX 5090 through kernel changes, speculative decoding enhancements and faster prefill. The company also reported a 1.2-times gain for vLLM on an RTX PRO 6000 Blackwell Workstation Edition and up to a 1.4-times gain on two DGX Spark clusters.

NVIDIA said the improvements are available through the llama.cpp and vLLM inference backends and can also be used through LM Studio and Ollama.

RTX Spark systems

NVIDIA said RTX Spark Windows PCs will ship in October, with Acer showing a compact desktop concept and Lenovo announcing the Yoga Pro 9n and Yoga 9n 2-in-1. The company described RTX Spark configurations as including a 1-petaflop RTX Blackwell GPU, as much as 128GB of unified memory and a 20-core Grace CPU.

CyberLink’s PhotoDirector AI PC Mode is also scheduled to launch with RTX Spark in October. NVIDIA said it will offer generative editing, image enhancement, object removal, background tools, portrait refinement and image creation, with users able to select local or cloud processing.

Source: NVIDIA’s “Sparks Fly: NVIDIA Accelerates Local AI at IFA 2026,” published September 3, 2026.