Qwen ยท Diffusion transformer
Qwen-Image
Qwen-Image is a pipeline of 3 networks totalling 28.849 billion parameters, of which 20.43 billion do the actual generating. At 1328x1328 that is 6,889 latent tokens per denoising step.
Whole pipeline28.849B
Denoiser20.43B
Download53.7 GB
Default steps30
What the pipeline is made of
This is the part that separates a media model from a language model: these are separate networks, and whether they sit in memory together is your choice, not the model's.
| Component | Parameters | Published as | On disk | What it does |
|---|---|---|---|---|
| transformer | 20.43B | BF16 | 38.05 GB | Does the generating. Runs once per step, so it dominates both memory and time. |
| text_encoder | 8.292B | BF16 | 15.45 GB | Turns your prompt into conditioning. Runs once, then can leave the GPU entirely. |
| vae | 0.127B | BF16 | 0.24 GB | Converts between pixels and the compressed latent space. Small, but its decode pass is a memory spike. |
These figures were measured from the weight file headers.
Memory by precision
Everything resident, at the model's native output size, on a card large enough that precision is the only constraint.
| Precision | Peak VRAM | Quality | What it costs you |
|---|---|---|---|
| BF16 | 58.87 GB | 100% | How the weights are published. No loss, and the largest footprint. |
| FP8 | 32.12 GB | 97% | Halves the denoiser with a small, usually invisible cost. Needs Ada or newer. |
| GGUF Q8_0 | 33.79 GB | 98% | Works on any card, unlike FP8. Slightly slower than native precision. |
| GGUF Q5_K_M | 24.39 GB | 95% | A middle step when Q8 will not fit. |
| GGUF Q4_K_M | 21.53 GB | 91% | The usual way a 12B image model gets onto an 8 GB card. Detail softens. |
| NF4 | 20.08 GB | 89% | Aggressive 4-bit. Fast to load, noticeably looser on fine detail. |
Memory by arrangement
The other lever: the same weights at the same precision, moved around differently. On this model quantising is the stronger lever: the denoiser is 71% of the pipeline and stays resident whatever you rearrange.
| Arrangement | Peak VRAM | Time | How it works |
|---|---|---|---|
| Everything resident | 58.87 GB | 2 min 8 s | All components stay on the GPU. Fastest, and needs the most memory. |
| Text encoder on CPU | 39.82 GB | 2 min 15 s | The prompt is encoded once on the processor, so the encoder never touches the GPU at all. Standard practice on video models, whose encoders are often larger than their denoisers. |
| Component offload | 39.58 GB | 2 min 24 s | One component on the GPU at a time. The text encoder runs, then makes way for the denoiser. Costs a few seconds per generation. |
| Sequential offload | 7.95 GB | 12 min 47 s | Weights stream layer by layer from system RAM. Runs almost anything on almost anything, and is many times slower. |
What a bigger output costs
Memory climbs with the pixel count, and the VAE decode pass climbs with it.
| Output | Latent tokens | Peak VRAM | Time |
|---|---|---|---|
| 512 x 512 | 1,024 | 55.26 GB | 12 s |
| 768 x 768 | 2,304 | 56.05 GB | 31 s |
| 1024 x 1024 | 4,096 | 57.15 GB | 1 min 3 s |
| 1536 x 1536 | 9,216 | 60.3 GB | 3 min 16 s |
| 2048 x 2048 | 16,384 | 64.7 GB | 8 min 4 s |
Measured at BF16 with everything resident on an RTX 4090, so the columns compare with each other rather than with your machine.
Which hardware runs Qwen-Image
117 of 118 consumer devices run it in some arrangement. Each row shows the fastest arrangement that fits on that device.
| Device | Memory | Verdict | Arrangement | Peak | Time |
|---|---|---|---|---|---|
| RTX PRO 6000 Blackwell NVIDIA | 96 GB | Slow | BF16 everything resident | 58.87 GB | 1 min 24 s |
| GeForce RTX 5090 NVIDIA | 32 GB | Barely usable | GGUF Q5_K_M everything resident | 24.39 GB | 1 min 41 s |
| RTX 6000 Ada Generation NVIDIA | 48 GB | Barely usable | GGUF Q8_0 everything resident | 33.79 GB | 1 min 56 s |
| GeForce RTX 4090 NVIDIA | 24 GB | Barely usable | GGUF Q4_K_M everything resident | 21.53 GB | 2 min 8 s |
| GeForce RTX 5090 Laptop NVIDIA | 24 GB | Barely usable | GGUF Q4_K_M everything resident | 21.53 GB | 2 min 43 s |
| NVIDIA DGX Spark 128GB NVIDIA | 128 GB | Barely usable | BF16 everything resident | 58.87 GB | 2 min 49 s |
| GeForce RTX 5080 NVIDIA | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 3 min 17 s |
| GeForce RTX 4080 SUPER NVIDIA | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 3 min 33 s |
| GeForce RTX 4090 Laptop NVIDIA | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 3 min 42 s |
| GeForce RTX 4080 NVIDIA | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 3 min 48 s |
| GeForce RTX 5080 Laptop NVIDIA | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 3 min 53 s |
| GeForce RTX 5070 Ti NVIDIA | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 4 min 11 s |
| GeForce RTX 4070 Ti SUPER NVIDIA | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 4 min 12 s |
| GeForce RTX 3090 Ti NVIDIA | 24 GB | Barely usable | GGUF Q4_K_M everything resident | 21.53 GB | 4 min 24 s |
| RTX A6000 NVIDIA | 48 GB | Barely usable | GGUF Q8_0 everything resident | 33.79 GB | 4 min 33 s |
| GeForce RTX 3090 NVIDIA | 24 GB | Barely usable | GGUF Q4_K_M everything resident | 21.53 GB | 4 min 57 s |
| RTX A5000 NVIDIA | 24 GB | Barely usable | GGUF Q4_K_M everything resident | 21.53 GB | 6 min 20 s |
| GeForce RTX 5060 Ti 16GB NVIDIA | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 7 min 42 s |
| Jetson AGX Orin 32GB NVIDIA | 32 GB | Barely usable | GGUF Q4_K_M everything resident | 21.53 GB | 8 min 17 s |
| Jetson AGX Orin 64GB NVIDIA | 64 GB | Barely usable | GGUF Q8_0 everything resident | 33.79 GB | 8 min 17 s |
| Radeon RX 7900 XTX AMD | 24 GB | Barely usable | GGUF Q4_K_M everything resident | 21.53 GB | 8 min 20 s |
| GeForce RTX 4060 Ti 16GB NVIDIA | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 8 min 24 s |
| Radeon PRO W7900 AMD | 48 GB | Barely usable | GGUF Q8_0 everything resident | 33.79 GB | 8 min 24 s |
| RTX A4000 NVIDIA | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 9 min 35 s |
| Radeon RX 7900 XT AMD | 20 GB | Barely usable | GGUF Q5_K_M text encoder on cpu | 15.3 GB | 10 min 27 s |
| Radeon RX 9070 XT AMD | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 11 min 6 s |
| Radeon RX 7900 GRE AMD | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 11 min 42 s |
| Radeon RX 9070 AMD | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 13 min 48 s |
| Radeon RX 7800 XT AMD | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 14 min 21 s |
| Radeon RX 6900 XT AMD | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 23 min 24 s |
| Radeon RX 7600 XT AMD | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 23 min 55 s |
| Ryzen AI Max+ 395 32GB AMD | 32 GB | Barely usable | GGUF Q4_K_M everything resident | 21.53 GB | 26 min 4 s |
| Ryzen AI Max+ 395 64GB AMD | 64 GB | Barely usable | GGUF Q8_0 everything resident | 33.79 GB | 26 min 4 s |
| Ryzen AI Max+ 395 96GB AMD | 96 GB | Barely usable | BF16 everything resident | 58.87 GB | 26 min 4 s |
| Ryzen AI Max+ 395 128GB AMD | 128 GB | Barely usable | BF16 everything resident | 58.87 GB | 26 min 4 s |
| GeForce RTX 4070 Ti NVIDIA | 12 GB | Barely usable | BF16 sequential offload | 7.95 GB | 26 min 22 s |
| GeForce RTX 4080 Laptop NVIDIA | 12 GB | Barely usable | BF16 sequential offload | 7.95 GB | 28 min 30 s |
| GeForce RTX 4070 SUPER NVIDIA | 12 GB | Barely usable | BF16 sequential offload | 7.95 GB | 29 min 55 s |
| GeForce RTX 3080 Ti NVIDIA | 12 GB | Barely usable | BF16 sequential offload | 7.95 GB | 31 min 1 s |
| Radeon RX 6800 AMD | 16 GB | Barely usable | GGUF Q4_K_M text encoder on cpu | 13.26 GB | 33 min 38 s |
| GeForce RTX 5070 NVIDIA | 12 GB | Barely usable | BF16 sequential offload | 7.95 GB | 34 min 17 s |
| GeForce RTX 3080 12GB NVIDIA | 12 GB | Barely usable | BF16 sequential offload | 7.95 GB | 35 min 26 s |
| GeForce RTX 3080 10GB NVIDIA | 10 GB | Barely usable | BF16 sequential offload | 7.95 GB | 35 min 26 s |
| GeForce RTX 4070 NVIDIA | 12 GB | Barely usable | BF16 sequential offload | 7.95 GB | 36 min 21 s |
| GeForce RTX 2080 Ti NVIDIA | 11 GB | Barely usable | BF16 sequential offload | 7.95 GB | 39 min 3 s |
Source: Qwen/Qwen-Image . Downloaded 0.3 million times in the last month. See how we calculate, or browse every image model.