Qwen-Image (20.4B DiT + 8.3B VL) VRAM & GPU Requirement Calculator
Interactive hardware sizing for Qwen-Image (20.4B DiT + 8.3B VL) (20.4B by Alibaba Qwen). Estimated 8K context footprint: 14.5 GB (INT4) • 23.0 GB (FP8) • 43.4 GB (FP16).
GPUs
Memory Requirements 23.04 GB
Requires 1 GPUs (based on memory capacity)
20.4 GB
All model weights
0 GB
Conversation history cache
1.94 GB
Concurrent processing cache
0.7 GB
Temporary computation cache
ComfyUI Pipeline & Hardware Estimates
Est. Gen Time
35.8 s/img
Min System RAM
48 GB RAM
Resolution
1 MP
Offload Mode
Smart Offload
Quick Deploy CLI Command
python main.py --lowvram --fp8_e4m3fn-unet --fp8_e4m3fn-text-enc --preview-method auto
Flags reflect your selected precision, context length, and tensor parallel GPU count.Open Qwen-Image spec page →
Qwen-Image (20.4B DiT + 8.3B VL) Architecture & Memory Profile
Total Parameters
20.4B
Architecture
Dense Transformer
Developer / Provider
Alibaba Qwen
Recommended Precision
FP8
Qwen-Image (20.4B DiT + 8.3B VL) ComfyUI VRAM Matrix by Precision & Resolution
Precomputed 3-stage ComfyUI Smart Offload + VAE Tiling VRAM peak (recommended host RAM: 48 GB).
| Precision | 1024x1024 (1MP) | 1536x1536 (2.25MP) | 2048x2048 (4MP) | LoRA Train (1MP) |
|---|---|---|---|---|
| FP16 | 43.4 GB | 46.3 GB | 51.1 GB | 46.6 GB |
| FP8 | 23.0 GB | 25.9 GB | 30.7 GB | 26.2 GB |
| INT8 | 23.0 GB | 25.9 GB | 30.7 GB | 26.2 GB |
| INT4 | 14.5 GB | 17.3 GB | 22.2 GB | 17.6 GB |
Recommended GPU Hardware for Qwen-Image (20.4B DiT + 8.3B VL) (FP8)
Minimum GPU count for 1024x1024 generation with ComfyUI offload and estimated seconds/image.
| GPU Hardware | Min GPUs | Total VRAM Pool | Est. Gen Time |
|---|---|---|---|
| NVIDIA RTX 4090 (24GB) | 1× | 24 GB | 35.8 s/img |
| NVIDIA RTX 5090 (32GB) | 1× | 32 GB | 25.9 s/img |
| Apple M4 Max (128GB) (128GB) | 1× | 128 GB | 43.4 s/img |
| NVIDIA A100 (80GB) (80GB) | 1× | 80 GB | 38 s/img |
| NVIDIA H100 (80GB) | 1× | 80 GB | 10.4 s/img |
| NVIDIA H200 (141GB) | 1× | 141 GB | 10.3 s/img |
| Huawei Ascend 910B (64GB) | 1× | 64 GB | 20.3 s/img |
Qwen-Image (20.4B DiT + 8.3B VL) Deployment FAQ
How much VRAM is needed to run Qwen-Image (20.4B DiT + 8.3B VL)?
At 1024x1024 with ComfyUI Smart Offload and VAE Tiling enabled, Qwen-Image (20.4B DiT + 8.3B VL) needs about 43.4 GB in BF16/FP16, 23.0 GB in FP8, and 14.5 GB in INT4/GGUF (paired with ~48 GB system RAM).
Can I run Qwen-Image (20.4B DiT + 8.3B VL) on RTX 4090 or H100 GPUs?
Running Qwen-Image (20.4B DiT + 8.3B VL) in FP8 precision requires at least 1x RTX 4090 (24GB) or 1x NVIDIA H100 (80GB) GPUs.
What precision should I choose for Qwen-Image (20.4B DiT + 8.3B VL) in production?
For Qwen-Image (20.4B DiT + 8.3B VL) (20.4B by Alibaba Qwen), FP8 provides the best balance between inference speed, VRAM footprint, and output quality.