Alibaba QwenDiffusion TransformerCommercial use permitted
Qwen-Image
The largest open image model here at 20.4B, and Apache 2.0. At BF16 it is a multi-GPU or heavy-offload proposition.
Parameters20.43BImage model
Weights at BF1657.7 GBAll components
Native resolution1328x1328Trained for
Minimum VRAM15.3 GBNF4 + sequential CPU offload
What you actually download
A diffusion pipeline is three separate models. These sizes are read from the real .safetensors files in the Hugging Face repository, not estimated.
Transformer / UNetThe image model itself40.86 GB
Text encoderTurns your prompt into embeddings16.58 GB
VAEDecodes latents into pixels0.25 GB
Peak VRAM at 1024×1024, one image
Every combination of precision and offload strategy. Lower precision costs image quality; offloading costs speed.
| Precision | Everything in VRAM | Text encoder on CPU | Sequential CPU offload |
|---|---|---|---|
| BF16 / FP16Reference quality. What the repos ship by default. | 60.7 GB | 44.1 GB | 43.9 GB |
| FP8Roughly half the weight memory. Small quality cost, widely used for Flux and SD 3.5. | 32.0 GB | 23.7 GB | 23.4 GB |
| NF4 / INT4Aggressive 4-bit. Visible quality loss, but it puts big models on small cards. | 23.8 GB | 15.5 GB | 15.3 GB |
Which machines can run it
48 of the 75 devices we profile can run Qwen-Image at 1024×1024 in some configuration. Showing the best-quality setup each one supports.
| Device | Best setup | Peak | Status |
|---|---|---|---|
| Apple Mac Studio (M5 Ultra, 512GB, 2026) | BF16 / FP16 | 60.7 GB | Comfortable |
| Apple Mac Studio (M3 Ultra, 512GB, 2025) | BF16 / FP16 | 60.7 GB | Comfortable |
| Apple Mac Studio (M5 Ultra, 256GB, 2026) | BF16 / FP16 | 60.7 GB | Comfortable |
| Apple Mac Studio (M2 Ultra 192GB) | BF16 / FP16 | 60.7 GB | Comfortable |
| NVIDIA DGX Spark (GB10, 128GB) | BF16 / FP16 | 60.7 GB | Comfortable |
| Apple MacBook Pro 16-inch (M4 Max 128GB) | BF16 / FP16 | 60.7 GB | Comfortable |
| Apple MacBook Pro 16-inch (M3 Max 128GB) | BF16 / FP16 | 60.7 GB | Comfortable |
| Next-Gen AI Laptop (Ryzen AI Max+ 395 128GB) | BF16 / FP16 | 60.7 GB | Comfortable |
| Apple Mac Studio (M5 Max, 128GB, 2026) | BF16 / FP16 | 60.7 GB | Comfortable |
| Framework Laptop 13 (Ryzen 7 7840U) | FP8 | 32.0 GB | Comfortable |
| Apple MacBook Pro 14-inch (M4 Pro 48GB) | FP8 | 32.0 GB | Comfortable |
| Apple Mac mini (M5 Pro, 48GB, 2026) | FP8 | 32.0 GB | Comfortable |
| Flagship AI Workstation (RTX 5090 32GB) | FP8 | 32.0 GB | Tight Fit |
| Apple MacBook Pro 16-inch (M4 Max) | BF16 / FP16Text encoder on CPU | 44.1 GB | With offload |
| Apple Mac Studio (M1 Ultra 64GB) | BF16 / FP16Text encoder on CPU | 44.1 GB | With offload |
| Homelab AI Server (Dual RTX 3090 48GB VRAM) | BF16 / FP16Text encoder on CPU | 44.1 GB | With offload |
| Apple Mac mini (M5 Pro, 64GB, 2026) | BF16 / FP16Text encoder on CPU | 44.1 GB | With offload |
| ASUS Zenbook A16 (Snapdragon X2 Elite Extreme, 48GB) | FP8Text encoder on CPU | 23.7 GB | With offload |
| Apple MacBook Pro 14-inch (M3 Pro 36GB) | FP8Text encoder on CPU | 23.7 GB | With offload |
| Custom Workstation PC (RTX 4090 24GB) | FP8Text encoder on CPU | 23.7 GB | With offload |
| Used AI Powerhouse PC (RTX 3090 24GB) | FP8Text encoder on CPU | 23.7 GB | With offload |
| Custom Workstation PC (Radeon RX 7900 XTX 24GB) | FP8Text encoder on CPU | 23.7 GB | With offload |
| Apple Mac mini (M6, 32GB, 2026) | NF4 / INT4Text encoder on CPU | 15.5 GB | With offload |
| Apple MacBook Pro 14-inch (M5, 32GB, 2025) | NF4 / INT4Text encoder on CPU | 15.5 GB | With offload |
| HP Omnibook 5 OLED | NF4 / INT4Text encoder on CPU | 15.5 GB | With offload |
| Lenovo ThinkPad T14s Gen 5 (Snapdragon) | NF4 / INT4Text encoder on CPU | 15.5 GB | With offload |
| ASUS Zenbook S 16 (Ryzen AI 9 HX 370) | NF4 / INT4Text encoder on CPU | 15.5 GB | With offload |
| Dell XPS 13 (Intel Core Ultra 7 258V) | NF4 / INT4Text encoder on CPU | 15.5 GB | With offload |
| Lenovo Yoga Slim 7x Copilot+ | NF4 / INT4Text encoder on CPU | 15.5 GB | With offload |
| Microsoft Surface Pro 11 Copilot+ (OLED) | NF4 / INT4Text encoder on CPU | 15.5 GB | With offload |
Showing the 30 devices with the most memory. The full hardware matrix covers all 75.
Licence: Apache 2.0
Yes. Qwen-Image is released under Apache 2.0, which permits commercial use. Keep the licence and attribution notices with any redistribution.
Compare model licences →Qwen-Image: common questions
- How much VRAM does Qwen-Image need?
- At 1024x1024 and batch 1, Qwen-Image peaks at about 60.7GB with everything resident at BF16, and as little as 15.3GB at NF4 + sequential CPU offload. The weights alone are 57.7GB at BF16 — 40.9GB of transformer, 16.6GB of text encoder and 0.25GB of VAE.
- Which GPUs can run Qwen-Image?
- 48 of the 75 devices we profile can run it at 1024x1024 in some configuration. The Apple Mac Studio (M5 Ultra, 512GB, 2026) handles it best, at about 60.7GB. Lower precision and text-encoder offload widen that list considerably.
- Can Qwen-Image be used commercially?
- Yes. Qwen-Image is released under Apache 2.0, which permits commercial use. Keep the licence and attribution notices with any redistribution.
- Does offloading the text encoder help on Qwen-Image?
- Yes, substantially. Qwen-Image uses a large language model as its text encoder — 16.6GB at full precision. Encoding happens once per prompt, so moving it to CPU costs a couple of seconds and frees that 16.6GB for the rest of the run, taking peak VRAM at BF16 from 60.7GB down to 44.1GB.
Weight sizes read from the model repository on 2026-09-12. Activation memory is an approximation, not a benchmark — BFWR does not run a hardware test lab.
Other open image models
Stable Diffusion 1.50.86B · 2.1GB at BF16 · conditions applyStable Diffusion XL 1.02.57B · 6.9GB at BF16 · conditions applyStable Diffusion 3.5 Medium2.47B · 16.5GB at BF16 · conditions applyStable Diffusion 3.5 Large8.15B · 27.9GB at BF16 · conditions applyFLUX.1 [schnell]11.89B · 33.7GB at BF16 · commercial OKFLUX.1 [dev]11.9B · 33.7GB at BF16 · non-commercialQwen-Image 2.17.12B · 33.1GB at BF16 · non-commercialFLUX.2 [klein] 4B3.88B · 16.0GB at BF16 · commercial OKFLUX.2 [klein] 9B9.08B · 34.7GB at BF16 · non-commercialFLUX.2 [dev]32.22B · 112.8GB at BF16 · non-commercial
Guides that go deeper
Which GPUs have the VRAM these models need, and what they cost right now.
- How to Run Stable Diffusion Locally: Hardware, VRAM and SetupCurrent models from SD 1.5 to Qwen-Image, and the VRAM each needs.Read the guide
- Stable Diffusion LoRA: Base Models, Weights and StackingMatching the base model, setting strength, and combining adapters.Read the guide
- How to Increase VRAM: What Actually Works on PC, Laptop and MacIt depends where the memory physically lives.Read the guide