NVIDIA DGX Spark
Hardware details
CPU
NVIDIA GB-class system
system processor
GPU
NVIDIA GB GPU
dedicated accelerator path
Memory
128 GB unified
96-128 GB local AI memory band
Segment
Workstation
NVIDIA device class
Storage
Plan model storage
local model downloads can grow quickly
Fit posture
NVIDIA-aligned workstation experiments and developer demos.
decision summary with stated limitations
Practical notes
Treat as a workstation signal, not a normal consumer mini PC.
Most useful when NVIDIA-aligned hardware matters more than DIY flexibility.
Source-backed model fit
NVIDIA DGX Spark planning boundary
Planning lane
Workstation 32B–70B planning lane
This workstation lane can justify evaluating 32B–70B class workflows, while quantization, context, runtime and workload still need a local check.
Planning band: 64–128 GB. It is not a measured requirement.
Memory boundary
128GB unified is a planning boundary, shared with the platform or runtime where applicable; it is not a measured compatibility guarantee.
Uncertainty to keep visible
Official device facts support this planning lane, but do not measure runtime allocation, throughput, model quality, or long-context fit.