NVIDIA is adding a 64GB version of its DGX Spark, a compact desktop computer built to run AI models locally, starting at $4,999. According to NVIDIA’s blog post, the new configuration becomes available on Friday, October 23, from Acer, ASUS, Dell, Gigabyte, HP and MSI.
The short version is that it has half the memory of the original 128GB model but the same chip and software, and NVIDIA says it can run models of up to 100 billion parameters entirely on the device. For people who want to run AI agents privately without renting cloud capacity, it lowers the entry point. It is still a specialist product, and a second unit doubles the cost.
What changes and what stays the same
The 64GB model keeps the GB10 Grace Blackwell Superchip, the DGX OS operating system and the full NVIDIA AI software stack found in the 128GB version. NVIDIA says it will be sold exclusively through its manufacturer partners.
The memory is unified, which means the processor and graphics unit share one pool. That matters because the size of the AI model you can run is limited mostly by how much memory it can fit in.
The machine ships ready for agent development. NVIDIA lists its Agent Toolkit, CUDA-X AI libraries and Nemotron open models, along with runtimes such as Ollama, vLLM and PyTorch with CUDA. The company says developers can go from power-on to running models in minutes.
What 100 billion parameters means
Parameters are the adjustable numbers inside an AI model, and the count is a rough measure of its size. NVIDIA says the 64GB configuration supports models of up to 100 billion parameters and the agent applications built on them, fully on device.
A general caution applies to a figure like this. What actually fits in memory also depends on how a model is compressed and how much text it must hold in context at once, so a headline number is a ceiling rather than a promise for every model.
How clustering works and what it costs
Every DGX Spark has a built-in ConnectX-7 network adapter. Two units can connect directly with a cable, and NVIDIA’s Sync Cluster Assistant detects them, validates their settings and configures the network. Every node runs the same software, so nothing has to be reconfigured when moving from one unit to two.
Two linked 64GB units pool their memory to 128GB, and NVIDIA says that expands model support to up to 200 billion parameters. The company also reports twice the memory bandwidth and up to 1.7 times the performance in its own test with the Qwen 3.8 27B model. That means a second unit does not double speed, and the figure comes from NVIDIA’s testing rather than an independent benchmark.
The price is the other side of the equation. Two units at the starting price come to $9,998 before any differences between partner systems, so the cluster path is a significant step up from a single machine.
What NVIDIA suggests people use it for
NVIDIA describes three workflows. The first is keeping a coding or research agent running around the clock, reviewing code, analyzing documents or carrying out multistep tasks. The second is running a language or image model on the Spark while using an app on a laptop or desktop, so the Spark handles the heavy inference and the PC stays free. The third is scaling to two units when a task outgrows one.
A feature called NVIDIA Sync Model Launcher is due at the end of the month. NVIDIA says it will let developers download and start Qwen3.8 27B on one unit or a cluster, make it reachable from their laptops and set up OpenCode so they can start coding in a browser.
These are NVIDIA’s own examples, not independent tests.
What this means for you
The 64GB DGX Spark suits developers, researchers and enthusiasts who want to experiment with their own data on their own hardware, or who work with material they would rather not send to a cloud service. It also fits anyone who wants an agent that runs continuously without paying for a cloud instance by the hour.
It is a poor fit for people who only use AI occasionally. For light use, cloud services are simpler and cost far less up front. It also does not suit anyone who needs models beyond the 100 billion parameter range on a single machine, or who does not want to handle the setup of a local system.
Before buying, compare partner configurations, since storage, design and support can differ between Acer, ASUS, Dell, Gigabyte, HP and MSI. Independent benchmarks are worth waiting for, because the performance figures so far come from NVIDIA.
The takeaway is that local AI hardware is getting a lower starting price, not a free pass. The 64GB model makes the entry cheaper, but memory still decides what you can run, and growing beyond one unit costs twice as much.
