Nvidia says the 64GB configuration can run models with as many as 100 billion parameters directly on the device. The company is positioning it for local inference, AI agents, fine-tuning, data science and edge-development workloads that can fit within the smaller memory footprint.
The new configuration arrives as capable open models increasingly fit on systems with less memory. That reduces the need for 128GB in some local AI workloads, while giving buyers a less expensive entry point into the DGX Spark ecosystem.
Nvidia is also retaining the ConnectX-7 networking hardware found in the larger model. Two 64GB DGX Spark systems can be connected directly and combined into a 128GB cluster, allowing developers to expand capacity without moving to a different software environment.
According to Nvidia, a pair of 64GB systems can support models with as many as 200 billion parameters. In the company’s testing with Qwen 3.8 27B, two connected units delivered as much as 1.7 times the performance of a single system.
To simplify that setup, Nvidia is introducing Sync Cluster Assistant. The software detects connected DGX Spark units, checks their configurations and sets up the ConnectX-7 network automatically.
The company is also adding Model Launcher to Nvidia Sync at the end of the month. The feature will let developers download and run Qwen 3.8 27B on either one DGX Spark or a cluster, with Sync handling the model configuration across connected systems. It will also configure OpenCode so the model can be used from a browser-based coding environment.
The 64GB model otherwise retains the same 20-core Arm CPU design and 273 GB/s of shared memory bandwidth as the existing DGX Spark. Nvidia also says the systems will ship with support for tools including Nvidia Agent Toolkit, CUDA-X AI libraries, Nemotron models, Ollama, vLLM and PyTorch with CUDA.
Blender is among the first creator applications adding support for the platform, with Nvidia saying a prebuilt installer is planned.
The original 128GB DGX Spark will remain available for workloads that require more memory, including larger models, longer context windows and more demanding fine-tuning tasks. The new 64GB option gives buyers another starting point while preserving the ability to scale by linking multiple units together.
This analysis is based on reporting from tom's HARDWARE & NVIDIA.
Image courtesy of NVIDIA.
This article was generated with AI assistance and reviewed for accuracy and quality.
