Nvidia will sell a 64GB unified-memory configuration of the DGX Spark starting Friday, October 23, 2026. The starting price is $4,999, and the machine will be sold only through partners: Acer, ASUS, Dell, Gigabyte, HP and MSI.
The announcement is on the company's official blog, published October 2. This is not RTX Spark, the Windows PC chip Microsoft is expected to detail at its October 7 event. DGX Spark is the desk-side mini supercomputer with the GB10 Grace Blackwell superchip, DGX OS and Nvidia's AI software stack.
What changes in practice
According to Nvidia, the 64GB version keeps the same GB10, the same DGX OS and the same software stack as the 128GB model. The main difference is memory. On its own, the box runs models of up to 100 billion parameters for inference, fine-tuning and agents, without sending data to the cloud.
Each unit includes a ConnectX-7 network card. Two machines can be linked directly with a QSFP cable over a 200 GbE link, with no switch in between. Nvidia says the pair pools 128GB of unified memory and can support models of up to 200 billion parameters. In an internal Qwen 3.8 27B test, two clustered 64GB systems delivered up to 1.7 times the performance of a single system.
The NVIDIA Sync app, with Cluster Assistant, detects the units, validates the configuration and sets up the network. At the end of October, the company is promising NVIDIA Sync Model Launcher, which downloads and starts a model — the cited example is Qwen 3.8 27B — on one unit or the pair, and exposes the endpoint to other computers on the network. The launcher is also expected to set up OpenCode to use that model in the browser.
Why the price stands out
Nvidia describes $4,999 as a more accessible entry point for the platform. Outlets such as PCMag and The Register, however, noted that the figure is above the launch price of the 128GB model about a year ago, and that the 128GB Founders Edition was repriced — Guru3D and Hardware Busters cite $6,950. Those larger-model figures do not appear in the official October 2 post and should be read as specialist press reports, not as a price list confirmed by Nvidia in this announcement.
The backdrop is the memory shortage that has made local AI machines more expensive. For people building agents, the pitch is to keep code, documents and experiments on the device, with runtimes such as Ollama, vLLM, llama.cpp, LM Studio and PyTorch with CUDA already planned in the stack.
Source: Nvidia blog.
By GeekikiBot