NVIDIA RTX Spark PCs ship October 2026; PAIR and Project Zenith expand local AI
RTX Spark gives the local AI stack a firm shipping date and concrete hardware specs, while PAIR and Project Zenith extend local inference to multi-PC home networks and high-memory Windows configurations respectively. These moves expand the surface area for running large models without cloud dependency.
NVIDIA confirmed at IFA 2026 that RTX Spark compact Windows PCs will ship in October 2026, featuring a 1 Petaflop RTX Blackwell GPU, up to 128GB of unified memory, and a 20-core Grace CPU. Lenovo announced the Yoga Pro 9n and Yoga 9n 2-in-1, and Acer showed a compact desktop RTX Spark concept at the event. Electronic Arts, Embark, and Ubisoft are among game publishers committing titles to the platform. RTX Spark integrates with the Windows Agent framework to run AI agents in the background under OS-level control, and NVIDIA and Microsoft are partnering to deliver faster inference and easier local agent setup on NVIDIA hardware.
NVIDIA launched PAIR (Personal AI Router), a free open-source tool that distributes AI inference requests across available GPUs on a home network, enabling parallel agentic workloads across multiple home systems. PAIR supports Windows, macOS, and Linux, covering GeForce RTX 20 Series and newer, RTX PRO workstations, DGX Spark, and Apple M4 or newer, and works with Ollama and LM Studio. NVIDIA noted that more than half of U.S. households have two or more PCs, representing untapped compute PAIR aims to leverage.
Microsoft announced Project Zenith, a Windows configuration for PCs with 64GB or more RAM, designed to run AI models with 30 billion or more parameters locally on-device.
NVIDIA contributed optimizations to llama.cpp delivering up to 1.9x higher throughput on a GeForce RTX 5090 through kernel optimizations and enhanced speculative decoding. vLLM achieves a 1.2x speedup on RTX PRO 6000 Blackwell Workstation Edition and up to 1.4x on two DGX Spark clusters, with new XQA attention kernels in FlashInfer contributing to the gains. These improvements are available through LM Studio and Ollama.
Three AI agent applications added simplified local model setup on NVIDIA hardware: Hermes Agent automatically detects the GPU and runs inference through integrated llama.cpp with NVIDIA optimizations on Windows; OpenClaw simplifies local model setup on Windows for RTX GPUs with at least 24GB of VRAM; and Perplexity Portable Computer runs the orchestrator LLM, subagent LLM, and agent harness entirely on local hardware with no cloud dependency, available on Linux for NVIDIA RTX GPUs with 24GB or more of VRAM.
At Gamescom, NVIDIA unveiled DLSS 4.5 Ray Reconstruction featuring a second-generation transformer model, with game titles including 007 First Light, CONTROL Resonant, and Gears of War: E-Day announced. MLCommons released MLPerf Client v2.0, developed with NVIDIA and other industry partners, adding benchmarks for agentic AI and image generation beyond prior LLM-only tests.
Microsoft announced Project Zenith on September 6, a Windows configuration for PCs with 64GB or more RAM, built to run AI models with 30 billion or more parameters locally on-device.
The announcement adds a software-side complement to the RTX Spark compact PCs and PAIR network inference tool NVIDIA described at IFA 2026.
At IFA 2026, NVIDIA confirmed RTX Spark ships in October and launched PAIR, a free open-source tool routing inference requests to whichever home PC has free GPU capacity.
NVIDIA also contributed llama.cpp optimizations reaching up to 1.9x throughput on an RTX 5090, announced a Microsoft partnership for faster local agent setup, and integrated the hardware with Windows Agent framework for background AI tasks. At Gamescom, NVIDIA unveiled DLSS 4.5, and MLCommons released MLPerf Client v2.0 adding agentic AI benchmarks.
I send one email each morning with the stories that moved. If you would rather just read here, that works too.
Subscribe free