NVIDIA DGX Spark

Overview

DGX Spark is NVIDIA's personal AI supercomputer, packing Grace Blackwell compute, unified memory, ConnectX-7 networking and a CUDA-accelerated AI software stack into one compact system for agents, inference, fine-tuning, data science and edge development. On October 2, 2026, NVIDIA announced a new 64GB unified memory configuration, available from Acer, ASUS, Dell, Gigabyte, HP and MSI starting October 23 at $4,999.

The 64GB model keeps the same GB10 Grace Blackwell Superchip, DGX OS and full NVIDIA AI software stack as the 128GB version, supporting models up to 100 billion parameters entirely on device while holding an accessible price point.

The interesting part is clustering. Every DGX Spark ships with a built-in ConnectX-7 NIC, and two units connected directly with a QSFP cable pool their memory to 128GB, expanding model support to 200 billion parameters while doubling memory bandwidth. In NVIDIA's Qwen 3.8 27B test, two clustered 64GB systems delivered up to 1.7x the performance of a single system. The Cluster Assistant in the NVIDIA Sync app detects connected units, validates configuration and configures the ConnectX-7 network, so the software environment never needs to be reconfigured.

Key Features

• 64GB unified memory: Supports models up to 100 billion parameters running entirely on device; two clustered units pool memory to 128GB and extend model support to 200 billion parameters
• NVIDIA Sync Cluster Assistant: Detects connected units, validates device configuration and configures the ConnectX-7 network, so nothing needs reconfiguring when scaling from one unit to two
• Agent-ready out of the box: NVIDIA Agent Toolkit, CUDA-X AI libraries, Nemotron open models, and runtimes such as Ollama, vLLM and PyTorch with CUDA are all supported from day one, taking you from power-on to a running model in minutes
• 200 GbE interconnect: Every unit ships with a ConnectX-7 NIC; two units connect directly over QSFP, delivering twice the memory bandwidth and up to 1.7x the performance
• Sync Model Launcher: Arriving at the end of the month, it downloads and launches Qwen3.8 27B on a single system or a cluster in a few clicks, and sets up OpenCode so developers can start coding in the browser
• Fully local execution: Agents run on your own device with no cloud dependency, so data never has to leave the machine

Use Cases

• Developers who want a coding or research agent on call around the clock, ready to review code, analyze documents or carry out multistep tasks
• Local deployments where a model is too large for one unit and two machines need to pool memory
• Teams and researchers handling sensitive data whose inference requests cannot go to the cloud
• Individual developers doing fine-tuning and data science experiments without spinning up a cloud instance for every job

Pros

• The 64GB configuration brings a personal AI supercomputer down to a $4,999 starting price
• Two units pool memory to 128GB, extending model support from 100 billion to 200 billion parameters
• Ships with ConnectX-7 built in, so a QSFP cable is all you need to cluster, with no software reconfiguration
• DGX OS and the full NVIDIA AI software stack work from day one, from power-on to a running model in minutes
• Runs fully locally, so data never leaves the device
• Available from six major manufacturers, so supply channels are stable

Pricing

The 64GB configuration goes on sale October 23, 2026, starting at $4,999, sold exclusively through manufacturer partners Acer, ASUS, Dell, Gigabyte, HP and MSI. The 128GB model costs more. It is a one-time hardware purchase with no per-token cost afterwards.

Summary

The 64GB DGX Spark raises the line on what you can run locally: 100 billion parameters on one unit, 200 billion on two, which covers the large majority of open models today. What it actually solves is not compute anxiety but two more practical things, keeping data on premises and not having to open a cloud instance for every task.

At $4,999, cheaper cards cannot hold models this size and setups that can cost far more. If you only run a small model now and then it is clearly not worth it. But if you want an agent sitting beside you all day, or your data simply cannot be uploaded, the math works out.

Version History

• 64GB unified memory configuration (2026-10-02): NVIDIA announced a 64GB unified memory configuration for DGX Spark, available October 23 from Acer, ASUS, Dell, Gigabyte, HP and MSI starting at $4,999, supporting up to 100-billion-parameter models fully on device. It retains the GB10 Grace Blackwell Superchip, DGX OS and the full NVIDIA AI software stack. Two 64GB units connected via ConnectX-7 and QSFP pool memory to 128GB, expanding model support to 200 billion parameters with up to 1.7x performance in NVIDIA's Qwen 3.8 27B test, configured automatically by NVIDIA Sync Cluster Assistant