Nvidia has put its first RTX Spark compact desktops on sale, turning the promise of a Windows PC that runs serious AI models on the desk into a product buyers can order today. Systems from ASUS, Dell and MSI went on sale on November 18, and more makers are set to follow.
The pitch is simple. Instead of sending prompts, files and code to cloud services, owners run large models and always-on agents on a small box beside the monitor. Nvidia describes the machines as a dedicated Windows PC for local models and always-on agents, and says they can run language models with 120 billion parameters locally.
What is inside the box
Every RTX Spark desktop pairs a 20-core Grace processor with a Blackwell graphics processor carrying 6,144 CUDA cores. The two chips share one pool of up to 128GB of unified memory, and Nvidia rates the system at up to 1 petaflop of AI performance. Ten gigabit Ethernet is built in for moving large model files and datasets quickly.
The shared memory pool is the detail that matters most. A 120 billion parameter model needs far more memory than a typical gaming PC offers, and 128GB gives developers room to load large models without splitting them across machines or renting cloud servers by the hour.
The first three systems
ASUS leads with the ProArt GR1X, a mini PC that measures 150 by 150 by 51 millimetres and weighs 1.48 kilograms. It comes in 64GB and 128GB versions, runs Windows 11, and can drive four 4K displays. ASUS lists the 64GB model from 4,499 dollars in the United States.
MSI offers the EdgeMesa N AI+ in a 1.3 litre aluminium chassis, with prices from 4,399 dollars for the 64GB model and 7,499 dollars for the 128GB model. MSI has also launched the series in India, where it is now available through retail and commercial channels. Dell is selling a Creator Edition desktop too, although Nvidia has not named the model.
More systems are coming. Mini PCs from Acer, GIGABYTE and HP are on the way, along with the Microsoft Surface RTX Spark Dev Box. Nvidia unveiled RTX Spark at COMPUTEX 2026, and laptops using the same chip shipped earlier this year.
Software for agents and creators
Hardware is only half the story. The desktops support Hermes Agent and Nvidia NemoClaw blueprints for building agents, plus Unsloth for fine tuning models on local data. A tool called Switchyard routes each task to the local model or a cloud model, so everyday work stays on the device while harder reasoning still uses the cloud.
Creative apps are part of the package too. Adobe is bringing native Windows on Arm versions of its apps to RTX Spark, and Blender 5.3 adds DLSS 4.5 Ray Reconstruction to its viewport. ComfyUI workflows can generate images and video locally, and creators can connect Photoshop to local tools through Hermes Agent.
Why local AI is having a moment
The launch arrives as AI models keep getting larger and more expensive to run in the cloud. OpenAI introduced GPT-6.2 Astra and Sol on November 17, and Google published its Gemini 4 Neon model card on November 18. Each new model raises the same question for developers, which is where the work should run.
Nvidia has just reported record quarterly revenue of 115.8 billion dollars, up 103 percent from a year earlier, as demand for AI hardware keeps climbing. Its shares still slipped after the results, a reminder that even record sales do not guarantee a calm market. The RTX Spark desktops are a bet that some of that demand now lives on desks, not only in data centres.
Who should look at one
The price puts these machines firmly in professional territory. A developer who runs agents around the clock, a creator who generates media locally, or a small team that wants models and data to stay in the office will get the most from one. Casual users who only chat with a cloud model will not. For the right buyer, though, the appeal is clear. A small box can now hold a model that once needed a server room.