NVIDIA announced a 64GB configuration of its DGX Spark personal AI supercomputer, priced at $4,999 and scheduled for release on Friday, Oct. 23. Hardware partners Acer, ASUS, Dell, Gigabyte, HP and MSI will supply the desktop units exclusively.
The 64GB edition retains the GB10 Grace Blackwell Superchip, DGX OS and complete NVIDIA AI software stack found in the 128GB version. Operating as a standalone system, it runs models containing up to 100 billion parameters and autonomous agents entirely on device without cloud connections.
Clustering and system expansion
Every DGX Spark includes an integrated NVIDIA ConnectX-7 network interface card. Connecting two 64GB units directly using a QSFP cable pools unified memory to 128GB across a 200 GbE fabric, expanding support to models with up to 200 billion parameters. In NVIDIA benchmark testing with Qwen 3.8 27B, two clustered systems delivered up to 1.7 times the performance of a single unit alongside double the memory bandwidth.
The NVIDIA Sync software application configures the multi-node link automatically. Its cluster assistant validates hardware, sets up the network, and distributes workloads without requiring developers to alter software environments across connected nodes.
Software support and ecosystem updates
DGX Spark supports frameworks including PyTorch with CUDA, Ollama, vLLM, llama.cpp and LM Studio out of the box, as well as the NVIDIA Agent Toolkit and Nemotron open models. Creative application developer Blender plans to release a prebuilt downloadable installer for the platform.
At the end of October, NVIDIA will deploy its Sync Model Launcher, enabling browser-based coding through OpenCode and simplifying local execution of Qwen 3.8 27B on single or clustered machines. Separately, manufacturers including Lenovo and Microsoft will ship Windows personal computers featuring NVIDIA RTX Spark processors this month, while Alibaba's lightweight Qwen-Image-2.1 model now runs locally across DGX Spark, DGX Station and RTX GPUs.
