# DGX Spark

> OnPrem World · https://onpremworld.com/en/hardware/nvidia-dgx-spark · Last updated: 2026-10-14 · CC BY 4.0

A small CUDA Linux box with 128 GB of unified memory: it holds 120B-class models, but its 273 GB/s bandwidth limits generation speed. It launched at $3,999, and NVIDIA has since raised the price.

## Specifications

|  |  |
|---|---|
| Memory | 64–128 GB LPDDR5X |
| Memory bandwidth | 273 GB/s |
| Power | 140 W |
| Launch | 2025-10 |
| Launch price | US$3,999 |
| Architecture | Grace Blackwell (GB10: Blackwell GPU + 20-core Arm CPU) |
| Software | CUDA |

## What it can run (128 GB, 8K)

Estimates, not benchmarks. See the methodology.

| Model | Parameters | Best quant | Total | tok/s (est.) |
|---|---|---|---|---|
| Qwen3.5-122B-A10B | 122B | Q5_K_M | 85.5 GiB | 23 |
| Mistral Small 4 119B (2603) | 119B | Q5_K_M | 83.4 GiB | 35 |
| gpt-oss-120b | 117B | Q6_K | 95.0 GiB | 39 |
| Llama 4 Scout 17B-16E Instruct | 109B | Q6_K | 89.8 GiB | 12 |
| GLM-4.5-Air | 106B | Q6_K | 87.3 GiB | 17 |
| Qwen3-Coder-Next (80B-A3B) | 80B | Q8_0 | 83.8 GiB | 51 |
| Llama 3.3 70B Instruct | 70B | Q8_0 | 75.7 GiB | 2.2 |
| Qwen3.5-35B-A3B | 35B | FP16 | 69.1 GiB | 27 |
| Qwen3.6-35B-A3B | 35B | FP16 | 69.1 GiB | 27 |
| Qwen3-32B | 32.8B | FP16 | 66.6 GiB | 2.5 |
| DeepSeek-R1-Distill-Qwen-32B | 32.76B | FP16 | 66.6 GiB | 2.5 |
| Gemma 4 31B IT | 30.7B | FP16 | 61.6 GiB | 2.7 |
| Qwen3-30B-A3B (2507) | 30.5B | FP16 | 60.9 GiB | 25 |
| Qwen3-Coder-30B-A3B-Instruct | 30.5B | FP16 | 60.9 GiB | 25 |
| GLM-4.7-Flash (30B-A3B) | 30B | FP16 | 59.6 GiB | 27 |
| Qwen3.5-27B | 27B | FP16 | 53.8 GiB | 3.0 |
| Qwen3.6-27B | 27B | FP16 | 53.8 GiB | 3.0 |
| Qwen3.8-27B | 27B | FP16 | 53.8 GiB | 3.0 |
| Gemma 3 27B IT | 27B | FP16 | 54.3 GiB | 3.0 |
| Gemma 4 26B A4B IT | 25.2B | FP16 | 50.1 GiB | 22 |
| Mistral Small 3.2 24B Instruct (2506) | 24B | FP16 | 48.7 GiB | 3.4 |
| Magistral Small 1.2 (2509) | 24B | FP16 | 48.7 GiB | 3.4 |
| Devstral Small 2 24B Instruct (2512) | 24B | FP16 | 48.7 GiB | 3.4 |
| gpt-oss-20b | 21B | FP16 | 41.8 GiB | 23 |
| Phi-4 (14B) | 14B | FP16 | 29.4 GiB | 5.9 |
| Gemma 3 12B IT | 12B | FP16 | 24.8 GiB | 6.8 |
| Gemma 4 12B Unified IT | 11.95B | FP16 | 24.2 GiB | 6.9 |
| Qwen3.5-9B | 9B | FP16 | 18.4 GiB | 9.1 |
| Qwen3-8B | 8.2B | FP16 | 17.7 GiB | 10.0 |
| DeepSeek-R1-0528-Qwen3-8B | 8.19B | FP16 | 17.6 GiB | 10 |
| Llama 3.1 8B Instruct | 8B | FP16 | 17.1 GiB | 10 |
| Gemma 4 E4B IT | 8B | FP16 | 16.4 GiB | 10 |
| Llama 3.2 3B Instruct | 3.21B | FP16 | 7.7 GiB | 26 |
| Kimi K2.6 | 1000B | Doesn't fit | – | – |
| DeepSeek-R1-0528 (671B) | 671B | Doesn't fit | – | – |
| DeepSeek-V4-Flash | 284B | Doesn't fit | – | – |
| Qwen3.8-Flash-Next | 180B | Doesn't fit | – | – |

## Prices

Latest price at each shop, as listed on the day we checked.

| Region | Shop | Listing | Price | Per GB | Checked |
|---|---|---|---|---|---|
| Taiwan | momo購物網 (摩天商城 store TP0002353) | 【麗臺科技】Leadtek NVIDIA DGX Spark GB10 / 128G / Gen4 4TB SSD桌上型AI超級電腦 (Third-party store inside momo; the page does not name the seller) | [NT$197,899](https://www.momoshop.com.tw/TP/TP0002353/goodsDetail/TP00023530001335) | NT$1,546/GB | 2026-10-14 |
| Taiwan | 原價屋 CoolPC | NVIDIA DGX Spark GB10 / 128G / Gen4 4TB SSD (Line on the shop's public price list (no per-product page); list stamped 2026/10/14 15:47) | [NT$199,500](https://www.coolpc.com.tw/evaluate.php) | NT$1,559/GB | 2026-10-14 |
| US | NVIDIA Marketplace | NVIDIA DGX Spark, 128GB unified memory, 4TB NVMe | [US$6,950](https://marketplace.nvidia.com/en-us/enterprise/personal-ai-supercomputers/dgx-spark/) | US$54/GB | 2026-10-14 |

Taiwan value: about 33.8 tok/s on Llama 3.1 8B Instruct (Q4_K_M, 8K), or 0.2 tok/s per NT$1,000.

US value: about 33.8 tok/s on Llama 3.1 8B Instruct (Q4_K_M, 8K), or 0.5 tok/s per US$100.

All prices: https://onpremworld.com/en/prices.md

## Frequently asked questions

### Which LLMs can the DGX Spark run?

With 128 GB and an 8K context, the largest models that fit include: Qwen3.5-122B-A10B, Mistral Small 4 119B (2603), gpt-oss-120b, Llama 4 Scout 17B-16E Instruct, GLM-4.5-Air. See the full table below.

### How fast is the DGX Spark for local LLMs?

Generation speed is mostly set by memory bandwidth (273 GB/s). For Llama 3.1 8B Instruct at FP16, expect roughly 10 tokens/s; bigger models are slower.

### Can the DGX Spark run 70B models?

Yes. Llama 3.3 70B Instruct fits on the DGX Spark (128 GB) at Q8_0 with an 8K context, using about 75.7 GiB, at an estimated 2.2 tokens/s.

## Sources

- https://www.nvidia.com/en-us/products/workstations/dgx-spark/
- https://docs.nvidia.com/dgx/dgx-spark/hardware.html
- https://nvidianews.nvidia.com/news/nvidia-dgx-spark-arrives-for-worlds-ai-developers
- https://forums.developer.nvidia.com/t/2-23-2026-price-change-announcement/361713