Developers who want to run AI agents on their own desk now have a cheaper option from Nvidia. The company is adding a 64GB configuration of its DGX Spark desktop, on sale October 23 from Acer, ASUS, Dell, Gigabyte, HP and MSI at a starting price of $4,999.
The new version keeps the GB10 Grace Blackwell superchip, the ConnectX-7 networking and the DGX OS software stack of the 128GB model. What changes is memory: 64GB of unified LPDDR5x instead of 128GB. Nvidia says that is enough for the most capable 30B to 35B class open models, and the GPU delivers up to 1 petaFLOP of FP4 compute with sparsity.
The pitch is economic. Agents burn tokens continuously through tool calls, retries and long context, and Nvidia puts token consumption growth at 14x since early 2026. On owned hardware there is no per-token bill.
Two units can be linked with a QSFP cable and configured through the NVIDIA Sync Cluster Assistant. That pools memory to 128GB and lifts model support to 200 billion parameters. On a Qwen 3.8 27B test, Nvidia measured up to 1.7x the performance of a single system. A forthcoming Sync Model Launcher will install and run models across a cluster with a few clicks.
DGX Spark supports Ollama, vLLM and PyTorch out of the box, part of a push to keep agent workloads private and local instead of metered in the cloud.
