Hardware and configurations
DGX Spark runs on the NVIDIA GB10 Grace Blackwell Superchip and is positioned as a local AI compute device for developers. The developer pages distinguish two memory configurations, with some playbooks applicable to both and others reserved for the larger configuration. Playbooks for DGX Spark 128 GB only include workloads such as training a concept LoRA on FLUX.2 Klein with AI Toolkit.
- NVIDIA GB10 Grace Blackwell Superchip
- DGX Spark 64 GB configuration
- DGX Spark 128 GB configuration
- Playbooks marked for 128 GB only where required
Developer quickstarts and playbooks
The DGX Spark page lists developer quickstarts with indicative completion times, aimed at people setting up the system for the first time. Longer playbooks cover fine-tuning, knowledge graph extraction, agentic workflows and team GPU workspaces. Entries are tagged by skill level, such as Beginner or Intermediate, and by task, such as Run LLM Inference or Build AI Applications.
- Chat with LLMs Using Open WebUI and Ollama — 15 min
- Generate Images and Videos with ComfyUI — 2 hours
- Monitor Your AI Compute with DGX Dashboard — 30 min
- Serve LLMs with vLLM on One Device — 30 min
- Extract triples with Ollama or vLLM into a graph database
- Fine-tuning workflows including supervised, RLHF and parameter-efficient training
Clustering and multi-node workloads
Multiple DGX Spark units can be connected so that larger models and training runs span nodes, using NVIDIA Sync clustering. Published playbooks describe connecting three DGX Spark devices in a ring topology and setting up a cluster connected through a switch. A separate intermediate playbook covers guided and configurable model serving with vLLM across DGX Spark clusters, and another documents benchmarked interconnect bandwidth for distributed training across nodes.
- Connect Multiple DGX Sparks for Distributed Workloads
- Ring topology setup for three devices
- Switch-connected cluster setup
- Serve LLMs with vLLM Across DGX Sparks
Access and support
The DGX Spark page offers a support section providing access to documentation, software and expert assistance. Users can connect from another computer and reach the machine remotely through NVIDIA Sync. A login is available on the page for access to these resources.
- DGX Spark Support for documentation and software
- Remote connection via NVIDIA Sync
- Connect from another computer
Relationship to DGX Station
NVIDIA publishes a parallel developer hub for DGX Station, a deskside system described as powered by the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip. Several playbooks are tagged for both DGX Spark and DGX Station, including serving LLMs with vLLM on one device, generating images and videos with ComfyUI, and running NemoClaw with a local LLM. The DGX Station pages also cover connecting two nodes for distributed workloads.
- DGX Station uses the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip
- Shared playbooks tagged for both DGX Spark and DGX Station
- vLLM single-device serving available on either system
Sources
Last verified 03 Nov 2026. This entry is compiled from the public web pages listed above. Nothing here is stated that those pages do not, and each of them was read on the date shown.