NVIDIA DGX Spark Founders Edition – GB10, 128GB, 4TB
NVIDIA
NVIDIA DGX Spark Founders Edition – GB10, 128GB, 4TB
- Compute NVIDIA GB10 Grace Blackwell superchip
- Memory 128 GB unified LPDDR5X memory
- Storage 4 TB self-encrypting NVMe SSD
- Connectivity NVIDIA ConnectX-7 networking
NVIDIA's own DGX Spark Founders Edition puts the GB10 Grace Blackwell superchip and 128 GB of unified memory into a 1.2 kg desktop unit. Built for AI developers and research teams who want to run and fine-tune large models locally on NVIDIA's reference platform. EU version with a 4 TB self-encrypting SSD.
- Ships across the EU
- Secure payment
In the box
- NVIDIA DGX Spark Founders Edition (GB10, 128 GB, 4 TB)
Shipping
We ship from Belgium to all EU countries. Delivery options, costs and times for your country are shown at checkout.
Manufacturer and safety
- Brand
- NVIDIA
Manufacturer and safety details for this product are available on request.
Highlights
What's inside.
The specs that decide what you can run, straight from the manufacturer's data sheet.
- 01GB10 Grace BlackwellA 20-core Arm CPU and a Blackwell GPU on one superchip, in NVIDIA's own reference design.
- 02128 GB unified memoryCPU and GPU share 128 GB of LPDDR5X at 273 GB/s, so large models load without splitting.
- 03ConnectX-7 at 200 Gb/sTwin QSFP ports link two or more DGX Spark units for models too large for one.
- 044 TB self-encrypting SSDRoom for weights, datasets and checkpoints, with drive-level encryption for sensitive data.
In practice
Built for the work.
Use case
The NVIDIA AI stack, ready on day one
DGX Spark ships with NVIDIA DGX OS and the NVIDIA AI software stack preinstalled, so teams can start fine-tuning and inference straight away. Sensitive data and experiments stay on your own desk instead of in the cloud.
Integration
Built to cluster over ConnectX-7
NVIDIA ConnectX-7 networking provides two QSFP ports at 200 Gb/s for linking two or more units, alongside 10GbE RJ-45 for the office network. The roughly 150 x 150 x 51 mm chassis runs from a 240 W USB-C power adapter.
Video
See it in action.
Performance
The numbers, from the source.
Charts and diagrams published by the manufacturer.
-
DGX Spark performance at launch vs January 2026 software: up to 2.5x faster on Qwen-235B (TensorRT-LLM, NVFP4 + speculative decoding), plus gains for Qwen3-30B, SD 3.5 Large and fine-tuning. Source: NVIDIA. -
DGX Spark llama.cpp generation speed Oct 2025 vs Jan 2026: GPT-OSS-20B 64 to 85 tok/s, GPT-OSS-120B 33 to 55, Nemotron Nano V2 29 to 33, Qwen3 30B 66 to 90 (about +35%). Source: NVIDIA. -
DGX Spark Computex 2026 software update: up to 2.6x overall throughput for Qwen3.6-35B with vLLM (NVFP4 + multi-token prediction). Source: NVIDIA.
Specifications
Every detail, on paper.
Part number 940-54242-0005-000
01Key specifications
- Compute
- NVIDIA GB10 Grace Blackwell superchip (20-core Arm CPU + Blackwell GPU)
- Memory
- 128 GB unified LPDDR5X memory (273 GB/s), shared by CPU and GPU
- Connectivity
- NVIDIA ConnectX-7 networking, 2x QSFP ports (200 Gb/s), to link two or more units
- Connectivity
- 10GbE RJ-45 Ethernet
- Software
- NVIDIA DGX OS (Ubuntu-based) with the NVIDIA AI software stack preinstalled
- Size
- About 150 x 150 x 51 mm, external power adapter
- Storage
- 4 TB self-encrypting NVMe SSD
- Power
- 240 W USB-C power adapter
