NVIDIA’s RTX Spark compact desktops went on sale on Wednesday, with ASUS, Dell and MSI shipping mini PCs built to run local AI agents, image and video generation and model fine-tuning on up to 128GB of unified memory.
A 20-core Grace CPU and a Blackwell RTX GPU with 6,144 CUDA cores push up to 1 petaflop of AI performance.
This is the desktop half of a family NVIDIA already seeded with RTX Spark laptops — and it changes the operating model. A dedicated box keeps agents and scheduled tasks running while a laptop travels, and running supported models locally keeps processed code and files on the PC. It also sidesteps per-token cloud API charges, as NVIDIA’s blog reported.
The 128GB local-model ceiling
Unified memory is the number that matters here. With up to 128GB shared between CPU and GPU, NVIDIA claims the desktops run 120-billion-parameter LLMs locally, render 90GB-plus 3D scenes in real time and edit 4:2:2 video without stutter — three workloads that typically force creators back to the cloud or a workstation.
Ten-gigabit Ethernet handles the unglamorous part: moving datasets, model checkpoints and creative assets to networked storage. The ASUS ProArt GR1X mini PC is positioned as the always-on agent machine; the MSI EdgeMesa N AI+ Series rounds out the launch lineup.

| Component | RTX Spark desktop spec |
|---|---|
| CPU | 20-core NVIDIA Grace |
| GPU | Blackwell RTX, 6,144 CUDA cores |
| Unified memory | Up to 128GB |
| AI performance | Up to 1 petaflop |
Switchyard splits the bill
The software stack is where the economics get interesting. NVIDIA Switchyard, now available in Hermes, automatically routes agent tasks between local and cloud models — everyday work on device, heavy reasoning in the cloud.
Developers get native CUDA support, Linux tooling through Windows Subsystem for Linux, agent setup via Hermes Agent or NVIDIA NemoClaw blueprints, and fine-tuning through Unsloth. The demo NVIDIA’s blog touts connects Photoshop and ComfyUI through Hermes Agent: a sketch becomes a rendered image or an on-device animation from natural-language instructions.
ComfyUI pipelines can then stretch a generated clip to 60 frames per second via RTX Video Frame Generation and convert it to HDR — all in one workflow.
Who gains, and the fine print
Creators and developers gain a machine that earns its desk space between editing sessions. And emulation — the historical tax on Windows on Arm — loses ground as more of Adobe’s stack goes native. TechPowerUp carried NVIDIA’s announcement essentially in full, while GPN Central’s shorter summary omits some detail, so treat those items as coming from NVIDIA directly.
Among the additional announcements, developers will be able to import Hugging Face models into Windows ML or execute machine learning tasks via DirectX 12. NVIDIA’s local AI roundup adds llama.cpp v1.0.0 with NVFP4 inference, Ricursive Intelligence’s Ripple-65B Preview trained on B300 GPUs, and Runway’s Praxis-1 — a 27-billion-parameter world action model that runs locally on everything from a GeForce RTX 5090 to DGX Spark.
Acer, GIGABYTE and HP are next
The launch lineup is only the first wave. In the near future, Acer, GIGABYTE and HP will each release mini PCs, as will Microsoft with its Surface RTX Spark Dev Box — and Microsoft confirmed at Ignite that Windows Server 2028 is coming to DGX Station for Windows on the GB300 Grace Blackwell Ultra Desktop Superchip.