Zeniteq is now available on mobile for iPhone, iPad and Apple Watch.Download nowLearn more
NVIDIA RTX Spark Desktops Go on Sale
ASUS, Dell and MSI systems are available now, but buyers should check model memory requirements, Windows on Arm compatibility and cloud routing before committing.
NVIDIA RTX Spark compact desktops from ASUS, Dell and MSI went on sale November 18, 2026, with systems from Acer, GIGABYTE, HP and Microsoft still forthcoming. The platform pairs a 20-core Grace CPU with a Blackwell GPU featuring 6,144 CUDA cores and up to 128GB of unified memory. The ASUS ProArt GR1X starts at $4,499 in 64GB and 128GB configurations.
Running 120-billion-parameter models requires heavy quantization, and the 64GB version leaves minimal headroom. The platform runs Windows on Arm, so application compatibility needs verification, and Switchyard's hybrid cloud routing can send data off-device, which matters for sensitive workloads.
NVIDIA’s RTX Spark compact desktops from ASUS, Dell and MSI are now available to purchase, according to the company’s November 18, 2026 announcement. Systems from Acer, GIGABYTE and HP, plus Microsoft’s Surface RTX Spark Dev Box, remain forthcoming. The launch brings up to 128GB of unified memory to dedicated Windows PCs designed for local AI models, agents and creative work. NVIDIA’s announcement makes that availability split explicit.
The practical appeal is a separate machine that can keep processing files, generating assets or running scheduled agent tasks while a laptop is elsewhere. The important qualification is that memory capacity, model speed and software compatibility are separate questions. NVIDIA’s claim that supported configurations can run 120-billion-parameter models is a capacity claim, not a published guarantee of a particular response speed.
Which RTX Spark Desktops Are Available?
NVIDIA lists three manufacturers with purchasable compact desktops and four with systems still coming.
Manufacturer
System identified in the announcement
Status on November 18
ASUS
ProArt GR1X
Available to purchase
Dell
RTX Spark compact desktop
Available to purchase
MSI
EdgeMesa N AI+ Series
Available to purchase
Acer
RTX Spark mini PC
Coming soon
GIGABYTE
RTX Spark mini PC
Coming soon
HP
RTX Spark mini PC
Coming soon
Microsoft
Surface RTX Spark Dev Box
Coming soon
The NVIDIA launch post does not consolidate prices, regional stock or configuration options across those manufacturers. Available to purchase should therefore be read as the announced launch status, not a promise that every memory configuration is immediately stocked everywhere.
ASUS provides the clearest configuration detail in the supplied launch material: the ProArt GR1X comes in 64GB and 128GB unified-memory versions. Its availability announcement says pricing and availability vary by region.
For buyers comparing systems today, I’d keep the forthcoming machines outside the immediate shortlist unless waiting is acceptable. An announced alternative cannot yet resolve a purchase decision about price, delivery or the exact configuration available locally.
What 128GB of Unified Memory Actually Buys
NVIDIA lists the following core hardware for RTX Spark compact desktops:
A 20-core NVIDIA Grace CPU.
A Blackwell RTX GPU with 6,144 CUDA cores.
Up to 128GB of unified memory.
10-gigabit Ethernet for transferring models, datasets and creative assets.
ASUS describes the CPU and GPU sharing one memory pool through NVLink-C2C. That gives the GPU access to a large shared pool, although the operating system, applications and AI runtime also need space. The headline capacity is not all available for model weights. These specifications come from NVIDIA and ASUS.
The significance of the 120-billion-parameter claim becomes clearer with some basic storage arithmetic:
Weight precision
Approximate storage for 120 billion parameters
16-bit
240GB
8-bit
120GB
4-bit
60GB
These are calculated decimal sizes for the weights alone: parameter count multiplied by bits per parameter, divided by eight. They exclude quantization metadata, temporary buffers, context storage and the rest of the running system. Actual model files and memory requirements vary.
That calculation explains why quantization is central to fitting a model of this scale. At 16-bit precision, its weights alone exceed the system’s maximum memory. At 8-bit precision, they leave little headroom. Four-bit weights offer considerably more space, with quality and execution characteristics depending on the model and quantization method.
It also explains why a 64GB configuration should not be treated as interchangeable with the 128GB model. A theoretical 60GB weight payload leaves very little room on a 64GB system before other allocations enter the picture.
Model Size and Context Length Share the Budget
ASUS advertises support for 120B-parameter models and up to one million tokens of context. Its announcement does not provide enough detail to conclude that both maxima are achievable together under a specified model, precision and usable speed. ASUS’s release should be read with that limitation in mind.
Long contexts consume additional memory, commonly through the attention key-value cache used during generation. The amount depends on model architecture, cache precision, context length and how many requests run simultaneously.
For an always-on agent, this matters more than a large parameter number alone. A model that fits while idle may leave insufficient room for a long conversation, a second model or a creative application. My view is that buyers should choose around the complete working set: model, expected context, concurrent tasks and applications.
ASUS also specifies that its advertised one-petaflop peak is at FP4 precision. That is a vendor compute specification, not a measured tokens-per-second result or a measure of answer quality.
What the ASUS Review Actually Measures
An independent Windows Central review of the ProArt GR1X adds useful context beyond the launch specifications. Reviewers Zac Bowden and Cale Hunt tested 64GB and 128GB systems and reported a $4,499 starting price. That starting price should not be assumed to include 128GB.
The review lists a desktop thermal design power of up to 140 watts, compared with an 80-watt maximum for the laptop implementation. Those figures describe the platform’s power envelope, not a measurement of whole-system electricity use at the wall.
Its published 128GB-to-128GB CPU comparison shows modest gains:
Published benchmark
ASUS ProArt GR1X, 128GB
Surface Laptop Ultra, 128GB
Geekbench 7 single-core
2,655
2,710
Geekbench 7 multi-core
24,849
23,571
Cinebench 2026 multi-core
7,248
6,885
The desktop’s multi-core scores are about 5% higher in these tests, while its Geekbench single-core score is slightly lower. Windows Central reported GPU gains of up to 13.3% in its 3DMark comparisons, with smaller improvements in more demanding, higher-resolution tests.
Those are independent measurements of one ASUS design. They are not local-LLM benchmarks, and the graphics-test improvement cannot be translated directly into a percentage increase in model generation speed. Other manufacturers’ cooling and power settings may also differ.
The useful inference is narrower: the desktop’s higher power allowance did not produce a similarly large increase in general CPU performance. I’d evaluate this machine primarily for its GPU-accessible memory, supported AI workloads and dedicated-desktop role, not as an automatic major CPU upgrade over an RTX Spark laptop.
Local Agents Depend on the Software Stack
RTX Spark is a Windows on Arm platform. NVIDIA says it supports native CUDA and Linux development tools through Windows Subsystem for Linux, giving developers both Windows-native and Linux-oriented paths for supported workflows. That does not mean every existing Windows application, extension or CUDA package is automatically compatible. NVIDIA’s software announcement identifies specific supported tools and updates.
The named ecosystem includes:
Hermes Agent and NVIDIA NemoClaw blueprints for setting up agents and connected workflows.
Unsloth for fine-tuning supported models.
ComfyUI for local image and video generation workflows.
Adobe applications, with expanding native Arm support and RTX acceleration.
Each category needs its own compatibility check. Fine-tuning support does not imply that a model advertised as fitting for inference can also be fully trained or fine-tuned within the same memory budget. Training-related allocations and the chosen tuning method change the requirements.
Adobe’s updates are similarly specific. NVIDIA lists TensorRT-RTX acceleration for Premiere features including Enhance Speech, Object Mask, Scene Edit Detection and Auto Reframe. Photoshop gains Remove Tool acceleration and native Camera Raw support. Substance 3D Painter 12.2’s native Arm support is in public beta, a meaningful distinction for production users.
I’d check the complete workflow before buying: the application version, required plug-ins, Python dependencies, model runtime and any custom ComfyUI nodes. A compatible main application is only one part of a working pipeline.
Hybrid Routing Changes the Privacy Calculation
NVIDIA says Switchyard, now available in Hermes, can route agent tasks between local and cloud models. Everyday work can run on-device, while more demanding reasoning uses a cloud model. That can extend an agent’s capabilities, but it also changes what “local AI” means for a particular task. NVIDIA describes the routing approach here.
Locally processed prompts and files can remain on the PC, and those inference tasks avoid per-token cloud API charges. A task routed to a hosted model may send information off the device and incur service charges.
For sensitive code or documents, I’d treat routing configuration as a purchase requirement. The presence of a local model is not enough; users need to understand which tasks and data stay local.
The Purchase Decision Starts With the Workload
The ProArt GR1X puts the hardware in a 150 × 150 × 51mm chassis. ASUS describes dual-fan cooling intended for continuous agent operation, while Windows Central reports cool, quiet operation in its review and criticizes the port selection and placement. Those observations apply to the ASUS model, not the entire RTX Spark desktop family. ASUS and Windows Central provide the respective design and review details.
The 10GbE connection is useful for large model files and network storage. Its nominal rate works out to 1.25GB per second before overhead, but reaching high transfer speeds also requires suitable networking and storage at the other end.
At the ASUS model’s reported $4,499 starting price, avoiding API fees is only part of the economics. Hardware, electricity, administration and any remaining cloud usage still count. There is no supported break-even calculation without knowing the workload and what cloud spending it would actually replace.
Before ordering, establish four things:
Which exact model and quantization will run?
How much context and concurrency are required?
Are the runtime, applications and extensions supported on this platform?
Must every task stay local, or is cloud routing acceptable?
Those answers are more useful than the largest advertised model size. The new desktops offer a dedicated home for local AI, but their value depends on whether the intended workflow fits comfortably and runs at an acceptable speed.
Frequently Asked Questions
3 questions
1
Which RTX Spark Desktops Are Available Now?
NVIDIA says ASUS, Dell and MSI systems are available to purchase. Acer, GIGABYTE, HP and Microsoft’s Surface RTX Spark Dev Box are forthcoming.
2
Can RTX Spark Run a 120B Model Without Quantization?
A 120B model’s 16-bit weights alone require approximately 240GB, exceeding the 128GB maximum. Lower-precision weights reduce that requirement, but runtime and context memory still need headroom.
3
Does Running an Agent on RTX Spark Keep Everything Local?
Only if its models, tools and routing are configured that way. NVIDIA Switchyard supports cloud routing, so some tasks can leave the device and incur cloud charges.