NVIDIA has expanded its personal supercomputing lineup with a 64GB unified memory configuration of the DGX Spark desktop system. Built around the GB10 Grace Blackwell Superchip, the hardware runs local AI models with up to 100 billion parameters without relying on cloud infrastructure. The new model starts at $4,999 through partner manufacturers including Dell, HP, ASUS, Acer, Gigabyte, and MSI.
NVIDIA packages the GB10 Grace Blackwell Superchip alongside 64GB of unified memory and a ConnectX 7 network card in a compact desktop chassis. The machine runs DGX OS and supports standard open source frameworks immediately, including PyTorch, Ollama, and vLLM. Digital creators also gain support in 3D production suites like Blender. The hardware handles data science tasks, inference workloads, and model fine tuning locally, keeping private code and enterprise data on the physical device.
Developers can link 2 separate DGX Spark units together using a QSFP cable to expand memory capacity. The NVIDIA Sync software detects connected hardware automatically and configures the 200 GbE network fabric without manual adjustments. This combined setup pools total unified memory to 128GB, raising model capacity to 200 billion parameters. In benchmark tests running the Qwen 3.8 27B model, a 2 unit cluster delivered up to 1.7x the performance of a single machine while doubling memory bandwidth.
NVIDIA is introducing the Sync Model Launcher to simplify running local models from personal laptops and web browsers. The tool handles model downloads and distributes processing across connected cluster nodes, while preparing environments like OpenCode for browser based programming. Hardware distribution is handled through third party system builders rather than direct retail sales. The base 64GB configuration carries an entry price of $4,999, targeting engineering teams building local software agents.
