NVIDIA DGX Spark 64GB Brings Local AI Agents to Your Desk

Want to run serious AI agents without shipping your data off to the cloud? NVIDIA just made that a whole lot easier.

This month, the company is rolling out a new 64GB unified memory configuration of its DGX Spark personal AI supercomputer. It's built and sold by some heavy hitters in the PC world — Acer, ASUS, Dell, Gigabyte, HP, and MSI — and comes with DGX OS and the full NVIDIA AI software stack preinstalled. Basically, you plug it in and you're ready to go.

The pitch here is privacy and independence. You can run capable local agents right on the device, no cloud dependency required. And when your workloads start piling up, you can link two units into a cluster through NVIDIA Sync Cluster Assistant without any extra setup headaches.

Under the hood, DGX Spark packs NVIDIA Grace Blackwell compute, unified memory, NVIDIA ConnectX-7 networking, and the CUDA-accelerated AI software stack into one compact box. It's a complete local AI platform for running agents, inference, fine-tuning, data science, and edge development — letting you experiment with models and your own data instead of leaning on cloud instances for every single task.

The new 64GB config is sold exclusively through those OEM partners. It keeps the same GB10 Grace Blackwell superchip, DGX OS, and full NVIDIA software stack as the 128GB model, just at a friendlier price point. On its own, it can fully run models up to 100 billion parameters locally, plus agent apps built on top of them.

Here's where it gets interesting: pairing two 64GB units isn't just about doubling memory. In NVIDIA's testing with Qwen 3.8 27B, a two-unit cluster delivered up to 1.7x the performance of a single 128GB system — and it keeps scaling as your needs grow. Each unit has a ConnectX-7 NIC built in, and a direct QSFP cable link pools memory to 128GB, pushing model support up to 200 billion parameters with double the memory bandwidth.

Out of the box, DGX Spark supports NVIDIA Agent Toolkit, CUDA-X AI libraries, Nemotron open models, and popular runtimes like Ollama, vLLM, and PyTorch with CUDA. Blender is one of the first major creative apps on board, with a prebuilt downloadable installer on the way. The NVIDIA Sync Model Launcher, landing at the end of this month, makes running local AI as easy as a few clicks — it can download and launch Qwen3.8 27B across connected devices and even set up OpenCode so you can start coding right in your browser.

Real-world use cases? Running coding or research agents around the clock, offloading model inference from your daily PC, or scaling up when a single task outgrows one device. The DGX Spark 64GB launches October 23 starting at $4,999. Thinking about picking one up, or waiting to see how the cluster setup performs? Let us know in the comments.