NV NVIDIA Prosumer Unified memory

NVIDIA DGX Spark

Runs downloadable LLMs up to ~195B at Q4, ~44.5-53 tok/s on an 8B model.

NVIDIA DGX Spark
128GB
Unified memory
273
GB/s bandwidth
195B
Max model (Q4)
$3,999
MSRP

What you can run

At 8k context, FP16 KV. Full-speed tok/s holds up to "max ctx"; beyond it the context spills to SSD swap and drops to the degraded figure.

Model sizeFP16Q8Q4
3B 33.2-39.6 t/s
≤1643k ctx · then ~1.2-1.5 (slow)
66.4-79.2 t/s
≤1689k ctx · then ~2.4-2.9 (slow)
118.6-141.4 t/s
≤1709k ctx · then ~4.3-5.2 (slow)
7-8B 12.5-14.8 t/s
≤745k ctx · then ~0.5-0.5 (slow)
24.9-29.7 t/s
≤806k ctx · then ~0.9-1.1 (slow)
44.5-53 t/s
≤833k ctx · then ~1.6-1.9 (slow)
13B 7.7-9.1 t/s
≤535k ctx · then ~0.3-0.3 (slow)
15.3-18.3 t/s
≤615k ctx · then ~0.6-0.7 (slow)
27.4-32.6 t/s
≤650k ctx · then ~1-1.2 (slow)
32B 3.1-3.7 t/s
≤190k ctx · then ~0.1-0.1 (slow)
6.2-7.4 t/s
≤312k ctx · then ~0.2-0.3 (slow)
11.1-13.3 t/s
≤365k ctx · then ~0.4-0.5 (slow)
70B won't run 2.8-3.4 t/s
≤133k ctx · then ~0.1-0.1 (slow)
5.1-6.1 t/s
≤227k ctx · then ~0.2-0.2 (slow)
120B won't run won't run 3-3.5 t/s
≤142k ctx · then ~0.1-0.1 (slow)
400B won't run won't run won't run

Estimate: tok/s ≈ bandwidth ÷ (params × bytes/param) × efficiency (NVIDIA). Cyan = fits ~115.2 GB usable. Beyond "ctx", KV spills to SSD swap (rough degraded speed shown).

Run popular models on the NVIDIA DGX Spark

Videos about the NVIDIA DGX Spark

NVIDIA DGX Spark vs RTX 4090 | LLM inference, training speed and more

Get Started With Your DGX Spark! | Setting Up LLMs & NIM

More hardware like the NVIDIA DGX Spark

About the NVIDIA DGX Spark

NVIDIA DGX Spark is a unified-memory system from NVIDIA, 2025. For local genAI the numbers that matter are its 128 GB (how big a model + context fits) and 273 GB/s bandwidth (generation speed).

Frequently asked questions

What size LLM can the NVIDIA DGX Spark run?

With ~115.2 GB usable it runs up to ~195B at Q4 (~55.5B at FP16) at 8k context. Bigger spills into SSD swap and slows sharply.

How many tokens/sec does the NVIDIA DGX Spark generate?

Generation is bandwidth-bound. At 273 GB/s an 8B model at Q4 runs ~44.5-53 tok/s (estimate).

Why does long context slow it down?

The KV cache for the context must also fit fast memory; once weights+KV exceed 115.2 GB it spills to SSD swap and throughput collapses. Each row shows the max usable context.