NVIDIA RTX Spark compact desktops launch with 128GB unified memory
NVIDIA RTX Spark compact desktops from ASUS, Dell and MSI launched on 18 November 2026, featuring 20-core NVIDIA Grace CPUs and Blackwell RTX GPUs with 6,144 CUDA cores. These systems offer up to 128GB of unified memory and 1 petaflop of AI performance. (unified memory: a memory architecture where the central processing unit and graphics processing unit share the same pool of memory.) (petaflop: a unit of measurement for massive artificial intelligence speed.) This hardware targets AI developers and 3D artists needing to run 120-billion-parameter models or 90GB+ scenes locally.
Is this for you?
If you need to run 120-billion-parameter models locally, the 128GB unified memory makes these desktops a viable workstation for your workflow.
- If you are an AI researcher, the 128GB unified memory allows you to run 120-billion-parameter models locally.
- If you are a 3D artist, the 128GB unified memory supports rendering 90GB+ scenes in real time.
- If you use Adobe Illustrator, the native Windows on Arm support delivers up to 30% higher performance.
- Skip this if you are a casual office user looking for basic PC performance.
- Skip this if you only use cloud-based AI tools as the value is in local execution.
What you can do now
- If you use Adobe Premiere, check if your version supports TensorRT-RTX to see if you can achieve 50% quicker turnaround on H.264 deliverables.
- If you use Blender, ensure you have Blender 5.3 and the latest NVIDIA Studio Driver installed to use DLSS 4.5 Ray Reconstruction.
- If you use cloud AI, compare the cost of local inference on RTX Spark desktops against your current cloud api pricing for 120-billion-parameter models.
Key figures
- CPU
- 20-core NVIDIA Grace
- GPU
- NVIDIA Blackwell RTX (6,144 CUDA cores)
- Unified Memory
- Up to 128GB (18 November 2026)
- AI Performance
- Up to 1 petaflop
- Connectivity
- Ten-gigabit Ethernet
ASUS, Dell, and MSI launch NVIDIA RTX Spark compact desktops
The new NVIDIA RTX Spark compact desktops provide a dedicated Windows PC environment for local AI agents, image and video generation, and model fine-tuning—the process of training an existing AI on new data. These systems integrate a 20-core NVIDIA Grace CPU with an NVIDIA Blackwell RTX GPU that features 6,144 CUDA cores. By combining these components, the hardware delivers up to 1 petaflop of AI performance (massive AI speed).
The core mechanism enabling these capabilities
The core mechanism enabling these capabilities is the 128GB unified memory—a memory architecture where the CPU and GPU share the same pool of memory. By eliminating the traditional memory wall between system RAM and dedicated video memory, the Spark architecture allows the GPU to access the full 128GB pool. This specific memory capacity is the requirement for loading and running massive 120-billion-parameter large language models (LLMs) that would otherwise exceed the limits of standard consumer GPUs.
Ten-gigabit Ethernet connectivity supports the movement of large datasets and model files between the compact desktops and networked storage. This high-speed interconnect helps ensure that the high-capacity memory is not bottlenecked by slow data transfers during intensive local AI workflows.
128GB of unified memory enables 120-billion-parameter LLMs
The 128GB unified memory allows the RTX Spark to run 120-billion-parameter models locally, which provides a significant advantage for users who need to keep data on the local PC. NVIDIA says running supported models locally keeps processed code and files on the machine and avoids per-token cloud application programming interface (API) charges. (cloud application programming interface (API): a service that allows software to communicate with and use resources on a remote server.) This shifts the economic model from recurring cloud costs to a one-time hardware investment for predictable local inference. (local inference: the process of running a trained model to get results directly on a local machine.)
You should choose the RTX Spark
You should choose the RTX Spark if you need to run models with more than 120 billion parameters locally. This recommendation holds while the system provides 128GB of unified memory; if the memory capacity were lower, the 120-billion-parameter model support would fail. Guessing wrong on memory requirements means you cannot load the models you need, while choosing this for simple cloud-based tasks results in overspending on hardware you won’t fully utilize.
Adobe and Blender software ecosystem optimizations
NVIDIA and Adobe have expanded native Windows on Arm support—applications built for ARM chips—across the creative suite to optimize performance on the RTX Spark platform. (native Windows on Arm: applications built for arm chips that run directly on the operating system without needing extra software layers.) Adobe Illustrator now offers native Windows on Arm versions that NVIDIA and Adobe say deliver up to 30% higher performance than the emulated versions. Adobe Photoshop includes TensorRT-RTX acceleration for the Remove Tool and native Camera Raw support. (TensorRT-RTX: a set of software optimizations used to accelerate artificial intelligence features in creative applications.)
Professional video editors
Professional video editors can use Adobe Premiere to accelerate AI-powered features like Enhance Speech and Object Mask via TensorRT-RTX. NVIDIA says these optimizations allow editors to realize over 50% quicker turnaround on everyday H.264 video deliverables. The software update also improves support for professional formats, including Canon RAW, Red RAW, ARRI RAW, and ProRes RAW.
3D artists can utilize Blender 5.3, which integrates DLSS 4.5 Ray Reconstruction into the Cycles viewport. This feature provides near-final-quality lighting and reflections in the interactive viewport in real time. To use this, you must ensure you have Blender 5.3 and the latest NVIDIA Studio Driver installed.
AI researchers and 3D artists gain local processing power
AI researchers and developers benefit from the 128GB unified memory, which allows them to run 120-billion-parameter LLMs locally to avoid cloud costs. For these users, the primary benefit is the ability to perform model fine-tuning and run agents like Hermes Agent or NVIDIA NemoClaw without sending data to external servers.
3D artists and visual effects creators benefit from the hardware’s ability to render 90GB+ 3D scenes in real time. The Spark architecture allows for immediate visual feedback on high-density environments that would normally require cloud rendering or significantly longer wait times on standard workstations.
Professional video editors benefit from the TensorRT-RTX acceleration, which provides a 50% quicker turnaround on H.264 deliverables. The native support for high-bitrate formats like Canon RAW helps ensure the hardware remains viable for professional production pipelines.
Windows on Arm developers benefit from native support for Adobe Illustrator and InDesign, removing the performance penalties of emulation. This makes the RTX Spark a viable native workstation for those already invested in the ARM ecosystem.
Casual office users should skip the RTX Spark desktops
Casual users looking for basic office PC performance should skip this as these are specialized high-performance compact desktops for AI and creative work; read standard consumer laptop or desktop reviews instead. Users who only use cloud-based AI tools can ignore this as the primary value is local execution to avoid costs; consult cloud AI platform documentation instead.
A 3D artist’s workflow walkthrough with Sven
Sven runs a freelance 3D studio and frequently handles massive environments for indie game studios. Sven previously struggled with cloud rendering costs for 90GB+ scenes that would exceed the memory of his current workstation. He buys an ASUS RTX Spark compact desktop to handle his production load.
Sven uses the 128GB of unified memory to load his entire 90GB+ scene
Sven uses the 128GB of unified memory to load his entire 90GB+ scene and the 6,144 CUDA cores to render in real time during his creative process. Sven finishes his projects faster without incurring cloud rendering fees, and his 3D assets stay on his local machine. If Sven’s scenes were under 90GB, he might have stayed with a smaller system, but the unified memory is the primary reason he chose Spark over a standard workstation.
Practical applications for RTX Spark hardware
| Use Case | Target User |
|---|---|
| Run 120-billion-parameter LLMs locally | AI Researchers |
| Perform real-time 3D scene rendering of 90GB+ size | 3D Environment Artists |
| Execute local AI workflows via llama.cpp and GGUF models | Local AI Developers |
| Generate and animate images on-device using ComfyUI | Generative AI Artists |
| Accelerate Adobe Substance 3D Painter workflows | Game Artists |
Local processing replaces cloud costs and emulation
The RTX Spark replaces per-token cloud application programming interface (API) charges, allowing for predictable costs and keeping data and code local on the PC. It also replaces emulation layers for Adobe software on Arm, providing up to 30% higher performance for applications like Illustrator.
Unconfirmed pricing and regional availability
NVIDIA did not state the regional availability of the RTX Spark desktops. The source did not provide pricing or specific configurations for the models. It is unconfirmed if the Unsloth library supports NVIDIA Blackwell RTX GPUs or which specific Linux development tools are supported in the Windows Subsystem for Linux.
While ASUS, Dell, and MSI offer these systems on 18 November 2026, Acer, GIGABYTE, and HP will offer mini PC versions in the future. Competition from cloud AI providers remains the primary alternative for users who do not require local data privacy or the ability to run 120-billion-parameter models on-device.
NVIDIA RTX Spark compact desktops: what to check, compare and watch
- Cloud api
- Running supported models locally on NVIDIA RTX Spark compact desktops avoids per-token cloud api charges. AI researchers and developers: compare the cost of local inference on RTX Spark desktops against existing cloud api pricing for 120B parameter models.
- Hugging face
- Developers can download GGUF models from Hugging Face to run within Windows ML. AI researchers and developers: watch Hugging Face for new GGUF model releases compatible with the 128GB unified memory of the RTX Spark.
- Annual procurement cycle
- The availability of RTX Spark compact desktops from ASUS, Dell, and MSI impacts the hardware selection for the next cycle. Professional video editors: schedule the purchase of RTX Spark compact desktops during the next annual procurement cycle.
NVIDIA RTX Spark compact desktops: questions the original report leaves open
- What are the key specifications, and how do they differ from the previous generation?
- The original report did not say how these specifications differ from the previous generation.
- How much does it cost, and in which configurations?
- The original report did not say how much it costs or the specific configurations.
- Whose performance figures are these, and have they been independently tested?
- The original report did not say if the performance figures were independently tested.
- Which competing products does it go up against?
- The original report did not say which competing products it goes up against.
- Who should upgrade, and from what?
- The original report did not say who should upgrade or from what.
- Does Unsloth support NVIDIA Blackwell RTX GPUs?
- The original report did not say if Unsloth supports NVIDIA Blackwell RTX GPUs.
- Are there 120B parameter GGUF models available on Hugging Face?
- The original report does not say.
- Which Linux development tools are supported in Windows Subsystem for Linux?
- The original report did not say which specific Linux development tools are supported.
Reported by NVIDIA on 18 November 2026: NVIDIA RTX Spark Compact Desktops Out Now.
Related articles
- NVIDIA RTX Spark: 128GB Unified Memory for Local AI
- Why local AI matters for privacy and hardware control
- NVIDIA expands RTX Spark laptop lineup with 128GB unified memory
ExtremeSpec is reader-supported: some links on this page are affiliate links, and as an Amazon Associate ExtremeSpec earns from qualifying purchases at no extra cost to you. Learn more.
