One-line summary
The NVIDIA DGX Spark 64GB (on sale October 23, 2026, from $4,999) and the 128GB version share the same GB10 Grace Blackwell superchip and software stack; the core difference is unified memory and the model scale each can run locally: 64GB runs up to 100 billion parameters on one system (two units can cluster to 200 billion), while 128GB supports up to 200 billion per the official launch specification.
Object introduction
- NVIDIA DGX Spark 64GB: a new configuration NVIDIA began selling on October 23, 2026 through six OEM partners (Acer, ASUS, Dell, Gigabyte, HP, MSI), with a suggested price from $4,999, positioned as a lower-barrier entry to personal local AI.
- NVIDIA DGX Spark 128GB: the standard DGX Spark configuration (successor to Project DIGITS), launched in 2025; official launch material says it can run models with up to 200 billion parameters locally.
- Both ship with DGX OS and the NVIDIA AI software stack, supporting the NVIDIA Agent Toolkit, CUDA-X AI libraries, Nemotron open models, and runtimes such as Ollama, vLLM and PyTorch out of the box.
Core parameter comparison
| Dimension | DGX Spark 64GB | DGX Spark 128GB |
|---|---|---|
| Chip | NVIDIA GB10 Grace Blackwell superchip | NVIDIA GB10 Grace Blackwell superchip |
| Unified memory | 64GB | 128GB |
| Suggested price | From $4,999 (2026-10-23) | From $3,999 at launch (2025); check official site for current price |
| Sales channel | OEM only (Acer/ASUS/Dell/Gigabyte/HP/MSI) | NVIDIA official and OEM channels |
| Local model scale (single unit) | Up to 100B parameters | Up to 200B parameters (official 2025 launch spec) |
| Cluster expansion | 2 units link to 128GB unified memory, up to 200B parameters, up to 1.7x performance (Qwen 3.8 27B test) | Multi-node cluster supported (NVIDIA Sync) |
| Networking | NVIDIA ConnectX-7 NIC | NVIDIA ConnectX-7 NIC |
| Software stack | DGX OS + NVIDIA AI software stack | DGX OS + NVIDIA AI software stack |
| Target users | Individual developers, AI enthusiasts, local AI entry | Users with higher-memory workloads |
Key differences
- Memory and model scale: the 128GB version doubles the unified memory of the 64GB version, raising the single-system model ceiling from 100B to 200B parameters (official launch spec). Two 64GB units linked over QSFP cable pool 128GB of unified memory, extending model support up to 200B parameters.
- Price entry: the 64GB version starts at $4,999 and sells only through OEMs, a lower entry point than the 128GB version (from $3,999 at launch in 2025; current pricing per the official site).
- Clustered performance: NVIDIA's Qwen 3.8 27B test shows two 64GB units clustered deliver up to 1.7x the performance of a single 128GB system, with doubled memory bandwidth.
- Software parity: both versions run the identical software stack; scaling from one system to two requires no software reconfiguration, and the NVIDIA Sync Cluster Assistant automates networking setup.
Buying advice
- Budget-sensitive users running models up to 100B parameters, or those who want to start with one unit and scale later, should choose the 64GB version (from $4,999).
- Users who need to run 100B-200B parameter models on a single machine, or want to avoid cluster complexity, should choose the 128GB version.
- Users who need more scale and can invest in stages can start with two 64GB units clustered (128GB unified memory total, up to 200B parameters, up to 1.7x performance), configured automatically with NVIDIA Sync.