NVIDIA PAIR, RTX Spark PCs, and Faster Local Inference Land at IFA 2026

NVIDIA used IFA 2026 to push local AI forward with faster inference, simpler agent setup, and a new tool that spreads workloads across idle PCs. The company also confirmed that RTX Spark Windows PCs from Lenovo and Acer will ship in October, joining six previously announced OEM designs.

Faster inference and simpler agent setup

New optimizations in llama.cpp deliver up to 1.9x higher throughput on a GeForce RTX 5090, thanks to kernel improvements, enhanced speculative decoding, and faster prefill. vLLM gains 1.2x on RTX PRO 6000 Blackwell Workstation Edition and up to 1.4x on two DGX Spark clusters, with new XQA attention kernels in FlashInfer and backend optimizations. These gains are available directly in the llama.cpp and vLLM backends, and through LM Studio and Ollama.

Three agent apps are simplifying local model configuration on Windows. Hermes Agent from Nous Research now offers one-click setup that detects the NVIDIA GPU, picks a model and configuration, and runs it through integrated llama.cpp with NVIDIA optimizations. Perplexity Portable Computer, previously available on Linux with RTX GPUs that have at least 24GB of VRAM, will get Windows support soon. OpenClaw, the largest AI project on GitHub with more than 380K stars, is working with NVIDIA and Microsoft on an OpenClaw Windows App that streamlines local model setup on RTX GPUs with at least 24GB of VRAM.

NVIDIA PAIR routes inference across local PCs

NVIDIA Personal AI Router (PAIR) is a free, open source tool that discovers compatible PCs on a local network and routes independent inference requests to whichever system has capacity. It works with Ollama and LM Studio and adapts as devices join or leave the network. The beta is available for Windows, macOS, and Linux, supporting GeForce RTX 20 Series and newer, RTX PRO workstation GPUs with Turing architecture or newer, DGX Spark, and Apple M4 or newer silicon.

Agentic workflows often split tasks into parallel jobs, but performance drops when every request hits the same GPU. PAIR lets a user run Hermes to sort an inbox, for example, with subagents distributed across multiple PCs instead of waiting on one GPU. That frees the main system for gaming or other work while AI tasks run elsewhere.

RTX Spark Windows PCs and more local models

RTX Spark combines a 1 Petaflop RTX Blackwell GPU, up to 128GB of unified memory, and a 20-core Grace CPU. At IFA, Acer showed a compact desktop concept, and Lenovo announced the Yoga Pro 9n and Yoga 9n 2-in-1. Electronic Arts, Embark, and Ubisoft are among the latest publishers bringing titles to RTX Spark, joining KRAFTON, NetEase, Riot Games, and XBOX.

NVIDIA also recapped August releases. Nemotron 3.5 Lightning is a 30-billion-parameter model that runs on RTX PCs, RTX PRO workstations, DGX Spark, and Jetson. Z.ai's GLM-5.3-Flash is a multimodal mixture-of-experts model for DGX Station. Qwen released Qwen3.8-Flash-Next, an open-weight multimodal MoE model for DGX Spark and DGX Station, plus Qwen3.8-27B, a 27-billion-parameter model for local agentic and coding workloads. LTX 2.5 is an open-world video generation model with NVFP4, FastVideo, and ComfyUI enhancements. MiniMax-H3 is an open-weight video generation model with synchronized audio, and FastH3, a four-step distilled version, improves performance by 7x. Meta's Muse Glimmer is a 30-billion-parameter coding and agentic model for GeForce RTX PCs, DGX Spark, DGX Station, and Jetson, with NVFP4 quantization for DGX Spark. DeepSeek v4 Flash is a 284-billion-parameter MoE model with 13 billion active parameters that runs on a 2x DGX Spark cluster and DGX Station.

CyberLink's PhotoDirector AI PC Mode, optimized for RTX Spark, integrates diffusion models into the photo editor with TensorRT-RTX and FP8 acceleration, offering local or cloud processing. It launches with RTX Spark in October.