NVIDIA RTX Spark Compact Desktops Launch with Up to 128 GB Unified Memory
NVIDIA’s new RTX Spark compact desktops from ASUS, Dell and MSI arrive with up to 128 GB unified memory.
Source: NVIDIA Press RoomWhat are NVIDIA RTX Spark Compact Desktops and who makes them?
NVIDIA announced that compact desktop systems built around the RTX Spark platform are now shipping from ASUS, Dell and MSI, with additional models from Acer, GIGABYTE, HP and Microsoft slated for later release. The devices are positioned as dedicated Windows PCs for running local AI agents, image and video generation, and model fine‑tuning, according to the NVIDIA Press Room.
What hardware specs enable local AI workloads?
All RTX Spark compact desktops share a common high‑performance core:
- CPU: 20‑core NVIDIA Grace processor
- GPU: Blackwell RTX graphics engine with 6,144 CUDA cores
- Memory: Up to 128 GB of unified memory, providing a single address space for CPU and GPU workloads
- Performance: Up to 1 peta‑flop of AI compute capability
- Connectivity: 10‑gigabit Ethernet for fast dataset and asset transfers
These specifications allow the machines to run very large models locally—NVIDIA cites the ability to host 120‑billion‑parameter large‑language models, render 90 GB+ 3D scenes in real time, and edit 4:2:2 video streams without dropping frames.
Which software and creative tools are supported?
The RTX Spark ecosystem bundles a growing set of applications that take advantage of the hardware:
- Adobe suite: Native Windows‑on‑Arm builds of Photoshop, Premiere, Substance 3D Painter, Illustrator and InDesign now use TensorRT‑RTX acceleration for features such as Remove Tool, Enhance Speech, Auto Reframe and high‑dynamic‑range video processing.
- Blender 5.3: Integrates DLSS 4.5 Ray Reconstruction, delivering near‑final‑quality ray‑traced previews directly in the Cycles viewport.
- ComfyUI: Provides local generation pipelines, including RTX Video Frame Generation and RTX Video HDR, which can expand clips to 60 fps and convert them to HDR for color grading.
- Hermes Agent & NVIDIA NemoClaw: Blueprint‑based agents that can be wired to Photoshop, ComfyUI and other creative apps, enabling natural‑language prompts that drive image creation, animation and asset isolation.
- Unsloth: A fine‑tuning library that lets developers adapt supported models on‑device.
- Switchyard (in Hermes): Dynamically routes tasks between local and cloud models, helping to keep routine work on the desktop while offloading heavy reasoning to the cloud.
A video from Adobe MAX 2026 demonstrates creators turning sketches into fully rendered images and animations on an RTX Spark machine: Adobe MAX 2026 NVIDIA RTX Spark Booth Activation.
How do developers build and run AI agents on these machines?
Developers can leverage several NVIDIA‑provided toolchains:
- CUDA and WSL2: Native CUDA support and Linux development tools via Windows Subsystem for Linux give a familiar environment for building and testing AI pipelines.
- Hermes Agent: Offers pre‑built blueprints such as Hermes‑NemoClaw that connect to local models and external APIs.
- Unsloth: Simplifies fine‑tuning of large language models directly on the desktop’s unified memory.
- Switchyard: Automatically decides whether a request should stay on‑device or be forwarded to a cloud endpoint, reducing per‑token API costs.
- Windows ML updates: Developers can import any GGUF model from Hugging Face and run it with the llama.cpp backend, or use the experimental DxCGC backend to execute models alongside graphics via DirectX 12.
These capabilities turn the compact desktop into a persistent AI workstation that can host always‑on agents, schedule long‑running renders, or serve a local OpenAI‑compatible inference endpoint.
What does this mean for creators, developers and researchers?
The launch signals a shift from cloud‑only AI workflows to truly local, high‑performance pipelines. For creators, the ability to generate assets, edit high‑resolution video and iterate on 3D scenes without leaving the workstation reduces latency and eliminates per‑token cloud fees. Developers gain a sandbox that mirrors production hardware, allowing rapid prototyping of agentic applications and fine‑tuning of massive models without needing a separate server farm. Researchers can experiment with large‑scale models and benchmark real‑time rendering or video‑processing pipelines on a single, compact PC.
In practice, a visual effects artist could run a 120‑B parameter language model to script scene descriptions, feed those prompts into ComfyUI for asset generation, and immediately preview the results in Blender’s ray‑traced viewport—all on the same machine.
How Grafy can complement RTX Spark workflows
While RTX Spark desktops provide the raw compute, Grafy’s multimodal workspace can be used alongside them for cross‑platform collaboration and rapid iteration. The Creator app (/creator) lets users generate mock‑up images or video clips that can later be refined on the RTX Spark machine. For code‑heavy agents, Grafy Code (/code) offers an AI‑assisted IDE where developers can prototype Python scripts that call the local models, then push the final scripts to the desktop for execution. Finally, the Reader app (/reader) makes it easy to ingest NVIDIA’s technical blogs, whitepapers or PDF documentation and ask targeted questions, keeping the knowledge base close at hand while building pipelines.
By pairing Grafy’s flexible, cloud‑agnostic tools with the on‑premise power of RTX Spark desktops, creators and engineers can maintain a fluid workflow that moves from concept to production without being locked into a single environment.
Frequently asked questions
What operating system do RTX Spark compact desktops run?
They ship with Windows 11, including native support for Windows‑on‑Arm applications and Windows Subsystem for Linux, allowing developers to run both Windows and Linux‑based AI tools.
Can I run a 120‑billion‑parameter LLM on a desktop with 128 GB of memory?
According to NVIDIA, the unified memory architecture lets the system keep the entire model in memory, enabling inference of 120‑B parameter LLMs without swapping to disk.
How does Switchyard decide between local and cloud execution?
Switchyard, built into Hermes, monitors the complexity and resource demand of each request. Simple or latency‑sensitive tasks stay on the device, while more compute‑intensive reasoning is automatically forwarded to a cloud model, balancing performance and cost.
Source: https://blogs.nvidia.com/blog/rtx-spark-desktop-local-ai-available-now/
Source: NVIDIA Press Room — NVIDIA RTX Spark Compact Desktops Out Now


