NVIDIA is introducing the DGX Spark HPC Developer Kit at SC26 in Chicago, bringing four DGX Spark systems together into a cluster for local AI development and research. The official announcement puts the four-node configuration at 512GB of combined memory and 20TB of NVMe storage, giving developers a concrete capacity target for workloads spread across the connected machines.
The networking arrangement is a key part of the kit: each system connects directly to two neighboring systems in a ring. Using NVIDIA ConnectX-7 networking, the setup can scale from one node to four without a network switch. That switch-free design removes a separate piece of networking hardware from the configuration — an important practical detail for developers considering how to assemble a small local cluster, rather than simply evaluating an individual machine.
For model deployment, NVIDIA says researchers can use NVFP4 to run models of up to 700 billion parameters locally across four DGX Spark systems. That figure concerns deployment across the cluster; separately, the company says DGX Spark supports fine-tuning models of up to 70 billion parameters. These are distinct capabilities, not interchangeable model-size limits: developers evaluating the platform need to match the figure to whether they intend to deploy a model or fine-tune one.
DGX Spark itself is powered by NVIDIA’s GB10 Grace Blackwell Superchip. NVIDIA positions it as a complete local platform for autonomous agents, with an AI software stack that includes tools for agent orchestration, evaluation and secure local deployment. The company also identifies larger AI models, faster inference and multi-agent workloads as uses for connecting up to four systems through ConnectX networking. The emphasis is therefore on both model capacity and the software needed to put those models to work locally.
The DGX Spark HPC Developer Kit is on display at SC26 and is available for sale. NVIDIA says an installation playbook for the NVIDIA HPC SDK is coming soon.