NVIDIA and Microsoft introduced RTX Spark for Windows laptops and compact desktops at an event in San Francisco. NVIDIA says the platform brings its full AI software stack to these computers so models and agents can run locally. Microsoft also announced general availability of operating-system infrastructure for agents to run persistently in the background.
Large models running directly on a laptop
NVIDIA points to Qwen 3.8 Flash Next, described as a 125B model with a 51B n-gram, as an example of a model RTX Spark can run locally. The company says it can operate without sending data to the cloud or metered usage charges. NVIDIA also claims its intelligence matches that of many cloud models.
Microsoft’s Pavan Davuluri introduced Surface Laptop Ultra, a laptop built around RTX Spark. According to Davuluri, it will offer up to 128 GB of unified memory and up to one petaflop of AI compute. That combination is intended to accommodate models that would not fit on a traditional laptop.
CUDA connects laptops with DGX Station
Developers get the full CUDA platform as part of RTX Spark. NVIDIA says the shared software stack across RTX Spark and DGX Station lets them move models, tools and workflows between the machines without rewriting them.
A compact desktop for always-on agents
Alongside the laptops, NVIDIA is preparing a compact desktop using the same RTX Spark superchip. The company says its design supports 24/7 operation as a dedicated local AI system with continuously running agents.
Microsoft Execution Containers (MXC) provides the operating-system infrastructure, with Microsoft announcing its general availability at the same event. It allows agents to keep working in the background under Windows’ control.
DGX Station for Windows is designed for even larger models
NVIDIA also previewed DGX Station for Windows, a deskside AI supercomputer bringing GB300 Grace Blackwell-class infrastructure to Windows. According to the company, DGX Station previously ran on Linux, leaving enterprise developers maintaining two environments: one for demanding AI workloads and another for Windows applications and workflows.
The forthcoming version uses the GB300 Grace Blackwell Ultra Desktop Superchip. NVIDIA specifies 748 GB of coherent memory and up to 20 petaFLOPS of FP4 AI compute. The company says those specifications are intended to support local models at up to a trillion-parameter scale.
NVIDIA says development and research teams will be able to build and run persistent agents on the machine, connecting them directly to Windows applications and infrastructure. They will also be able to fine-tune large models and run inference on the same computer. Linux AI tools are expected to remain accessible through WSL.
Laptops are due October 16, desktops in November
RTX Spark laptop preorders opened on the day of the announcement, with availability planned for October 16. Compact desktops are scheduled to go on sale in November.
For DGX Station for Windows, NVIDIA is currently offering registration for notifications about future availability.
