Skip to content
localmodel.run

image model · qwen · Linux

Can I run Qwen-Image on Nvidia DGX Spark (128GB)?

Compatibility verdict VRAM check
Yes, it runs fast on this GPU

Yes. Qwen-Image runs on Nvidia DGX Spark (128GB) at Q4_K_M GGUF (~14 GB of ~112 GB usable).

Needs ~14 GB Device usable ~112 GB

Runs at Q4_K_M GGUF using ~14 GB of ~112 GB usable.

Peak VRAM
~14 GB
Usable on device
~112 GB
Device memory
128 GB
Quant
Q4_K_M GGUF
Share on X Share on Reddit

How to run it

Use ComfyUI or Nunchaku (SVDQuant 4-bit) at Q4_K_M GGUF. The big text encoder is loaded to encode your prompt, then offloaded before generation, which is why peak VRAM stays near the backbone size rather than the sum of every file.

Model qwen
Type
image (MMDIT)
Parameters
20B
Peak VRAM
~14 GB at Q4_K_M GGUF
Resolution
1328×1328
License
Apache-2.0
Full Qwen-Image requirements →
Device Linux
Memory
128 GB unified
Usable for weights
~112 GB
Power draw
~140 W
Best runtime
llama.cpp (CUDA) / Ollama (CUDA)
Best models for Nvidia DGX Spark (128GB) →

You could also run

Run Qwen-Image on other hardware

Nvidia GeForce RTX 4060 Ti (16GB)Nvidia GeForce RTX 4080 (16GB)Nvidia GeForce RTX 5080 (16GB)Nvidia GeForce RTX 5070 Ti (16GB)Nvidia GeForce RTX 5060 Ti (16GB)AMD Radeon RX 9070 XT (16GB)

FAQ

Can Nvidia DGX Spark (128GB) run Qwen-Image?

Yes. Qwen-Image runs on Nvidia DGX Spark (128GB) at Q4_K_M GGUF (~14 GB of ~112 GB usable).

How much VRAM does Qwen-Image need?

Nvidia DGX Spark (128GB) has room to spare. At Q4_K_M GGUF the realistic peak is ~14 GB of VRAM, versus ~57 GB with every component kept resident (no offload). With aggressive CPU offload it drops to ~3 GB, much slower.

What do I use to run Qwen-Image locally?

Qwen-Image runs in ComfyUI or Nunchaku (SVDQuant 4-bit). It loads as a diffusion checkpoint plus its text encoder and VAE, not a single chat command.

Sources

VRAM figures are sourced peak-usage anchors at the noted quant; catalog updated 2026-11-01. See methodology.