Bringing AI to the Desktop for Agents and Local Workloads
NVIDIA has explained how its DGX Spark platform is now used. The company built the platform to make AI more accessible. Developers, data scientists, researchers, and everyday users can reach AI compute on a desktop with ease.
For a quick overview, see this video overview.
Over the past year, demand for AI agents has surged. These agents run tasks automatically. In doing so, they generate many tokens and also consume a very high share of them.
Meanwhile, open-source models that run locally have improved sharply in quality and accuracy. Therefore, more users now run AI models and agents on local devices. They want to protect data privacy and cut token costs.
A 64GB Version Joins the Lineup
Beyond the original 128GB configuration, NVIDIA announced a DGX Spark with 64GB of memory. NVIDIA stresses that many new open-source models fit easily in 64GB. The space still leaves room for the operating system and large context prompts.
Full Model Compatibility
More importantly, models that run on the 128GB DGX Spark can move straight to the 64GB version. Both share the same underlying system architecture. As long as a model loads into memory, performance is identical. Thus, no compatibility problems arise.
No Founders Edition, Only OEM Builds
The 128GB version came in a gold NVIDIA Founders Edition. The 64GB version will have no NVIDIA-branded model. Instead, OEM partners, including ASUS, will build all the products.
The 64GB version also supports linking systems to scale compute. In addition, users can pair a 128GB unit with a 64GB unit. This flexibility lets them grow compute as needs change.
NVIDIA Sync Model Launcher
To help more people deploy on-device AI, NVIDIA plans to release a tool at the end of October. It is called NVIDIA Sync Model Launcher.
With a few clicks, developers can download and install Alibaba’s Qwen model. The current default is the Qwen3.8-27B size. The tool also helps connect laptops and other networked devices to the DGX Spark. Users can code through a browser or through OpenCode as well.
Supported Tools and Agents
Besides Qwen, users can deploy llama.cpp, Ollama, vLLM, or LM Studio on the DGX Spark. The platform also supports NVIDIA’s NemoClaw, OpenClaw, Hermes Agent, and OpenShell agent tools.
Pricing and Future Plans
The 64GB DGX Spark launches on October 23. The suggested price starts at $4,999.
On the other hand, memory supply limits and rising costs are driving a price change. Starting October 2, NVIDIA raises the price of the existing 128GB version, the NVIDIA FE model, to $6,950.
No 32GB Entry Model Planned
Will a cheaper 32GB version follow? In the Q&A session, NVIDIA answered clearly. It has no such plan and no preparations for a release. For now, it focuses on two options, 64GB and 128GB.
Bigger Clusters With Cluster Assist
For users who need even more compute, NVIDIA also updated the Cluster Assist feature in the NVIDIA Sync app. Through the built-in network interface, users can now link up to four DGX Spark systems into a cluster. As a result, the combined machines deliver higher performance.
Support Our Threat Intelligence
If you find our technology report and cybersecurity news helpful, consider supporting our work.
Buy Me a Coffee
PayPal
