NVIDIA and Microsoft have unveiled a broad push to move frontier artificial intelligence directly onto local desktop hardware. Announced at IFA 2026, the collaboration brings faster inference speeds and streamlined 1 click setup tools for open agent applications. The ecosystem expands further this October with compact RTX Spark PCs from partners like Acer and Lenovo.
The hardware centerpiece of the announcement is the NVIDIA RTX Spark platform scheduled to ship in October. Built around a 20 core Grace processor and an RTX Blackwell graphics chip capable of 1 Petaflop of compute, these machines support up to 128GB of unified memory. Acer revealed a compact desktop concept, while Lenovo introduced the Yoga Pro 9n and the Yoga 9n 2 in 1. Major game studios including Electronic Arts, Embark, and Ubisoft have also confirmed upcoming support for the architecture.
Running local models previously meant wrestling with quantization profiles and manual backend setup. NVIDIA is smoothing out that workflow through native optimizations in llama.cpp and vLLM. Benchmarks show inference throughput jumps up to 1.9x on the GeForce RTX 5090, with similar gains available through tools like Ollama and LM Studio. Popular open frameworks are taking advantage immediately. Hermes Agent now offers 1 click automated configuration on Windows, while the OpenClaw Windows app simplifies local deployments for systems packing at least 24GB of memory.
Perplexity is also expanding its footprint with the rollout of Portable Computer to Windows systems with 24GB of video memory. The software executes full research, coding, and document analysis tasks directly on local silicon without chewing through cloud subscriptions. When complex queries require external reasoning, the software prompts for explicit user permission before pinging remote frontier models.
To take advantage of unused household hardware, NVIDIA introduced a free open source tool called PAIR, short for Personal AI Router. The software scans your local network and splits heavy agentic tasks across idle machines, routing workloads to any system with spare graphics capacity. On the creative side, CyberLink announced PhotoDirector AI PC Mode, which uses local TensorRT acceleration to handle generative image editing directly on device without cloud lag.
