NVIDIA DGX Spark 64GB Expands Local AI Development and Scaling
NVIDIA DGX Spark 64GB as AI agents move from experimental projects into everyday development workflows, increasingly capable open models are becoming small enough to run on more devices. This shift is giving developers, researchers and AI enthusiasts new opportunities to build, test and deploy AI locally without relying on cloud infrastructure for every task.
NVIDIA is expanding those possibilities with a new 64GB configuration of NVIDIA DGX Spark. Coming this month from leading manufacturer partners Acer, ASUS, Dell, Gigabyte, HP and MSI, the new configuration combines 64GB of unified memory with DGX OS and the NVIDIA AI software stack, providing a ready-to-use local AI platform from the moment the system is powered on.
The new DGX Spark configuration is designed to run capable AI agents directly on the device, enabling private, local workflows without a mandatory cloud connection. When projects require additional memory and compute, developers can also connect two DGX Spark systems using NVIDIA Sync Cluster Assistant, allowing workloads to scale without requiring a complete infrastructure overhaul.
A Personal AI Supercomputer for Local Development
NVIDIA DGX Spark combines the NVIDIA Grace Blackwell architecture, unified memory, NVIDIA ConnectX-7 networking and a CUDA-accelerated AI software stack in a compact system designed for local AI development.
The platform provides developers with a dedicated environment for experimenting with AI models and their own data. Instead of sending every workload to a cloud-based GPU instance, users can run inference, fine-tuning, data science workflows and agentic applications locally.
The new 64GB model retains the same GB10 Grace Blackwell Superchip, DGX OS and NVIDIA AI software stack found in the 128GB configuration. With support for models containing up to 100 billion parameters, the system is designed to accommodate increasingly sophisticated local AI applications while maintaining a relatively accessible entry point for developers and enthusiasts.
The ability to run these models directly on the device can also provide greater control over data and development environments. For workflows involving proprietary code, documents or other sensitive information, local processing can reduce the need to move data to external cloud services.
Scale Local AI by Connecting Two Systems
For developers whose workloads eventually exceed the capacity of a single DGX Spark, the platform is designed to scale beyond one system.
Every DGX Spark includes an NVIDIA ConnectX-7 networking interface. Two systems can be connected directly using a QSFP cable, allowing their memory to be pooled to provide up to 128GB of unified capacity. This expands support for models of up to 200 billion parameters while increasing available memory bandwidth.
NVIDIA testing with Qwen 3.8 27B demonstrates the potential performance gains. Two clustered 64GB systems delivered up to 1.7x the performance of a single system, providing developers with additional capacity without requiring a move to a traditional multi-node data center environment.
NVIDIA Sync Cluster Assistant is designed to simplify the process. The software can detect connected DGX Spark systems, validate their configurations and set up the ConnectX-7 network. Because each system runs the same NVIDIA software stack, developers can scale from one device to two without having to rebuild their software environment.
This approach allows users to begin with the hardware capacity they need today and add another system as workloads become more demanding.
Software Ready for AI Agents
DGX Spark is designed to provide a development environment that works with popular AI frameworks and tools out of the box.
Support includes the NVIDIA Agent Toolkit, CUDA-X AI libraries, Nemotron open models and widely used runtimes such as Ollama, vLLM and PyTorch with CUDA. This enables developers to move from initial setup to running models with minimal configuration.
The platform is also attracting support from creative software developers. Blender is among the first major creator application providers to support DGX Spark, with a prebuilt downloadable installer planned for the platform.
NVIDIA is also expanding the tools available for deploying local AI models. NVIDIA Sync Model Launcher, scheduled to arrive later this month, is designed to simplify model deployment through a graphical workflow. Developers will be able to download and launch Qwen3.8 27B on a single DGX Spark or across a connected cluster.
NVIDIA Sync can automatically configure the model across connected devices, while the launcher also sets up OpenCode to use the model. This gives developers a streamlined way to begin experimenting with local coding agents and AI-assisted development.
Practical Workflows for Developers
The 64GB DGX Spark configuration is aimed at a range of real-world local AI applications.
One example is running an AI agent continuously for coding or research. A developer can keep an agent running locally to review code, analyze documents or complete multi-step tasks. When larger models, longer context windows or multiple simultaneous agents are required, connecting two systems can provide additional capacity.
Another use case involves separating AI workloads from a user’s primary PC. DGX Spark can handle language or image-generation model inference while an agent or creative application runs on a laptop or desktop. This allows the local AI system to handle demanding model workloads while the user’s everyday computer remains available for other tasks.
A third scenario involves scaling an existing workflow as requirements increase. A project that begins with a single DGX Spark can be expanded to two systems when a larger model, longer context window or greater concurrency becomes necessary. With NVIDIA Sync Cluster Assistant handling the underlying configuration, developers can continue using the same software environment while expanding the available resources.
Getting Started With DGX Spark 64GB
The new NVIDIA DGX Spark 64GB configuration will be available from Acer, ASUS, Dell, Gigabyte, HP and MSI starting Friday, Oct. 23, at a starting price of $4,999.
Developers can begin by installing a supported inference framework such as llama.cpp, Ollama, vLLM or LM Studio and downloading a compatible local model for their workflow.
For users who need additional capacity, two DGX Spark systems can be connected through their NVIDIA ConnectX-7 ports. NVIDIA Sync Cluster Assistant then helps configure the network and automatically route workloads across the connected systems.
NVIDIA is also providing dedicated playbooks for agentic AI development through build.nvidia.com, including resources for NemoClaw, OpenClaw, Hermes Agent and OpenShell. Additional playbooks for 64GB systems are planned, covering LLM serving with vLLM, running OpenClaw with a local LLM and connecting multiple DGX Spark systems for distributed workloads.
Expanding the Local AI Ecosystem
The arrival of the 64GB DGX Spark configuration comes as NVIDIA continues expanding its local AI ecosystem across desktops and workstations.
New Windows PCs powered by NVIDIA RTX Spark are also expected this month from Acer, ASUS, Dell, HP, Lenovo, Microsoft and MSI, giving developers and creators additional options for running AI workloads locally.
Meanwhile, Alibaba’s Qwen-Image-2.1 brings image generation and editing together in a lightweight open-weight model that can run locally on NVIDIA RTX GPUs, DGX Spark and DGX Station.
Together, these developments point toward a broader shift in AI computing: powerful models and agents are increasingly moving closer to the developer and the user. With DGX Spark 64GB, NVIDIA is providing a compact platform that can start as a personal AI supercomputer and scale into a two-node system as workloads grow, giving developers more flexibility to build, experiment and run increasingly capable AI applications locally.
Source Link: https://blogs.nvidia.com/