NVIDIA Launches 64GB DGX Spark to Run AI Agents Locally

A lighter version of the DGX Spark mini-supercomputer arrives this month from six manufacturers, with the same GB10 chip as the 128 GB model and the possibility of pairing two of them.

NVIDIA is expanding its DGX Spark lineup with a configuration with 64GB of unified memory, sold exclusively by partners Acer, ASUS, Dell, Gigabyte, HP, and MSI. The case retains the GB10 Grace Blackwell chip, ConnectX-7 network card, DGX OS system, and the entire software stack of the 128GB model. It is aimed at developers who want to run agents on their own machine, without going through the cloud, with models of up to 100 billion settings. According to The Neuron, the first shipments are scheduled for October 23 starting at $4,999.

Two machines to double the bet

Two units connected by a QSFP cable pool their memory (128 GB in total) and then accept models up to 200 billion parameters. In an internal test on Qwen 3.8 27B, the duo offered up to 1.7 times the performance of a single case. The NVIDIA Sync application is responsible for detecting machines and configuring the network, and a tool to launch models in a few clicks should follow at the end of the month, with an automatic configuration of OpenCode to code with the local model. Ollama, vLLM and PyTorch are supported out of the box, and Blender is preparing a dedicated installer.