NVIDIA’s 128GB Grace Blackwell desktop - a 1.2kg cube that puts
datacenter-class memory capacity on a desk. Shipped October 2025 at a
$4,699 MSRP (down from the $4,999 announced at GTC in March 2025). GB10
chip: 128GB LPDDR5X unified at 273 GB/s, 170W TDP, NVFP4 acceleration on
the Blackwell GPU, ConnectX-7 networking built in.
What it does well:
-
Capacity at low power: 128GB unified means 70B-class dense models
at 8-bit or 200B-class MoE at 4-bit, on under 240W at the wall.
-
Stacking: ConnectX-7 links two or more Sparks into one pool -
community builds run 1x through 4x units, and the
DGX Spark benchmarks guide has
the measured numbers per stack size.
-
NVFP4: Blackwell’s 4-bit format keeps quality high at the
capacity-focused quants this memory tier wants.
Where it falls short: 273 GB/s is the binding constraint - decode
speed per stream is a quarter of an RTX 5090’s, and the 128GB ceiling
means the biggest open checkpoints do not fit even at 1-bit.
Run it locally: ships with DGX OS and day-one Ollama support; vLLM
and SGLang cover the cluster builds. The
modeldex results rank every seeded
checkpoint by fit and estimated tok/s.