# NVIDIA DGX Spark (128GB) for local AI

URL: https://insidethemachine.ai/machines/dgx-spark-128gb
Type: NVIDIA AI box. Chip: GB10 Grace Blackwell. Memory: unified, options 128 GB. Bandwidth: 273 GB/s. Spec source: https://developer.nvidia.com/blog/how-nvidia-dgx-sparks-performance-enables-intensive-ai-tasks/ (2025-10-24).
Price: from about $6,950 (2026-10-15, approximate). Status: current.

## Which models fit

| memory GB | usable GB | models that fit | best 3 |
|---|---|---|---|
| 128 | 96 | 90 | Qwen3.8 27B (22.7 GB), K2 Horizon MoVA 36B-A4B (28.1 GB), Qwen3.5 27B (21.5 GB) |

## Published speeds (not our tests)

- gpt-oss-120b: 55.37 tok/s decode, 1725.47 tok/s prompt (llama.cpp, MXFP4, , vendor-measured) https://developer.nvidia.com/blog/how-nvidia-dgx-sparks-performance-enables-intensive-ai-tasks/ (2025-10-24)
- gpt-oss-20b: 82.74 tok/s decode, 3670.42 tok/s prompt (llama.cpp, MXFP4, , vendor-measured) https://developer.nvidia.com/blog/how-nvidia-dgx-sparks-performance-enables-intensive-ai-tasks/ (2025-10-24)

Notes: Price raised to $6,950 (~75% above launch) on 2026-10-02 per The Register. CUDA + 200GbE ConnectX-7; two units can be clustered. Partner GB10 boxes (Dell, ASUS, etc.) exist at various prices (not captured).

## Questions

**What is the best local LLM for the NVIDIA DGX Spark (128GB)?**
With 128 GB, Qwen3.8 27B (needs 22.7 GB, AA 33.7). 90 of 124 models fit.

**How fast is the NVIDIA DGX Spark (128GB) for local AI?**
Published: gpt-oss-120b at 55.37 tokens/sec (llama.cpp, ); gpt-oss-20b at 82.74 tokens/sec (llama.cpp, ). These are other people's measurements, not ours.

**How much does the NVIDIA DGX Spark (128GB) cost?**
From about $6,950 (as of 2026-10-15). 2026 prices are inflated by the memory-chip shortage; check before buying.