Itdaily - Nvidia brings RTX Spark to compact desktops for local AI

Nvidia brings RTX Spark to compact desktops for local AI

Two compact desktops side by side: a silver ASUS model in front and a black one behind.

Nvidia is now shipping compact RTX Spark desktops from ASUS, Dell and MSI. The systems run AI models with 120 billion parameters locally and have up to 128 GB of unified memory.

The new desktops are designed for AI agents, generating images and videos, and fine-tuning models on a dedicated Windows PC. Systems from Acer, Gigabyte, HP and Microsoft will follow soon. Microsoft is releasing the Surface RTX Spark Dev Box.

According to Nvidia, a fixed desktop can run agents and scheduled tasks while a laptop is on the move. Anyone who runs supported models locally keeps code and files on the PC and pays no per-token costs for cloud APIs.

RTX Spark hardware

The desktops have a 20-core Grace CPU and a Blackwell RTX GPU with 6,144 CUDA cores. Together they deliver up to 1 petaflop of AI compute. A 10-gigabit Ethernet port speeds up the transfer of large datasets and model files to network storage.

Nvidia also cites 3D scenes of more than 90 GB that render in real time and smoothly editable 4:2:2 video. Developers use CUDA and Linux tools via Windows Subsystem for Linux. You set up agents with Hermes Agent or NemoClaw blueprints, and you fine-tune models with Unsloth.

Nvidia Switchyard, now available in Hermes, automatically routes agent tasks to local or cloud models. Routine tasks run on the device, while complex reasoning tasks rely on the cloud.

Software for creative applications

Adobe is releasing native Windows on Arm versions of Illustrator and InDesign. According to Nvidia, Illustrator performs up to 30 percent better than the emulated version. Premiere accelerates features such as Enhance Speech and Auto Reframe via TensorRT-RTX, which delivers H.264 video more than 50 percent faster. Substance 3D Painter 12.2 runs natively on Arm for the first time in a public beta.

Cluster for researchers

At Supercomputing 26 in Chicago, Nvidia is introducing the DGX Spark HPC Developer Kit. It is a desktop cluster of four systems with a combined 512 GB of memory and 20 TB of storage, enough for local models of up to 700 billion parameters.

This lets researchers develop and test workloads without waiting for shared supercomputers. Nvidia has not yet announced pricing or Belgian availability.