Nvidia ships 64 GB DGX Spark, a personal AI supercomputer for local inference

The 64 GB version of DGX Spark hits shelves this month. Nvidia's compact box is built to run models and agents on-premises with no cloud dependency. Six partners — Acer, Asus, Dell, Gigabyte, HP and MSI — will sell the new configuration with the same GB10 Grace Blackwell Superchip, the same DGX OS and the same software stack as the pricier 128 GB model. The only difference is unified memory, which lets a single device run models up to 100 billion parameters at a price Nvidia calls "accessible" without quoting a figure.

The interesting part starts when you connect two units. Each box includes a built-in ConnectX-7 card; a single QSFP cable turns the pair into a cluster with 128 GB of shared memory, double the bandwidth and support for models up to 200 billion parameters. In Nvidia's test on Qwen 3.8 27B, the duo delivered up to 1.7× the performance of a single unit — not linear scaling, but a solid jump with zero manual infrastructure setup. The Sync Cluster Assistant discovers the second node, validates the configuration and brings up the network automatically.

The software stack is there from first boot: Nvidia Agent Toolkit, CUDA-X AI libraries, open Nemotron models, and popular runtimes including Ollama, vLLM and PyTorch with CUDA. Nvidia says a developer can go from powered off to a running model in minutes. Blender is expected to be among the first to offer a dedicated installer for the platform, signalling a target audience beyond researchers — three-dimensional creators who want local inference.

By month's end Nvidia will add Sync Model Launcher, an interface that downloads and runs Qwen3.8 27B on a single unit or the cluster, configures the model across the connected devices and exposes it to the developer's laptop. The launcher also spins up OpenCode in the browser so coding against the local model can start immediately. Every node runs the identical software stack, so no reconfiguration is needed when moving from one system to two.

AI agents are graduating from experiments to production, and open models are shrinking to fit more devices. DGX Spark 64 GB sits squarely between a consumer-GPU laptop and a data-center server — private, local and expandable without writing infrastructure code. When the price lands we will learn whether "accessible" means a lab budget or a freelancer budget, but the architecture already proves the home cluster has stopped being a weekend project and become an off-the-shelf product.