Image generation
Qwen-Image-2.1
Qwen-Image-2.1 has been deployed and tested on one MegaCube. Validation covers text-to-image generation at multiple resolutions and step counts, transparent PNGs, text rendering, resource use, and service stability.
01 / Model parameters
Model architecture & capabilities
Qwen-Image-2.1 generates and edits images using a 7.1B-parameter single-stream DiT, a Qwen3-VL text encoder, VAE, and other components. Deployment, stability, multi-resolution performance, and output quality were validated on one MegaCube, focusing on Chinese text rendering, transparent PNGs, composition across aspect ratios, and commercial imagery.
- Marketing posters and visual content
- Transparent PNG assets and product images
- Style explorations and creative previews across aspect ratios
- Publisher / model
- Alibaba Qwen · Qwen-Image-2.1
- Generation network
- 7.1B-parameter single-stream DiT; 32 layers; BF16; FlowMatchEuler scheduler
- Text encoder
- Qwen3-VL, approx. 8.75B; BF16
- Weight size
- Approx. 33.1 GB; 7 safetensors shards and 29 repository files
- Input / output
- Text input; image output; model capabilities include image editing and native transparent RGBA images
- Default output
- 2048×2048 / 40 steps; supports 1:1, 4:3, 3:4, 3:2, 2:3, 16:9, 9:16, and other aspect ratios
- Inference dependencies
- diffusers 0.41.0.dev0；transformers ≥ 5.17；torch ≥ 2.4
- License
- Qwen Research License; license assessment required before commercial release
Tests primarily cover text-to-image generation, transparent RGBA output, rendering of longer Chinese text, and multiple visual styles. The model also supports image editing, but this page makes no performance claim for editing workflows that have not been tested separately.
Model information checked on 2026-09-21
02 / Deployment
MegaCube deployment configuration
- Compute devices
- 1 × MegaCube · NVIDIA GB10 · 128 GB unified system memory
- Service architecture
- Docker-containerized HTTP image generation service; FastAPI; /health + /generate
- Runtime environment
- Ubuntu 24.04 LTS；ARM64；Docker 29.x；CUDA 13.0 / PyTorch 2.13.0+cu130
- Service port
- 8300; host networking
- Model directory
- /home/test/models/Qwen-Image-2.1; read-only mount
- Concurrency method
- One request generated at a time; a global lock ensures serial GPU execution
Verify the weights
Verify all 29 repository files against the per-file SHA256 manifest and confirm all weight shards are complete.
Build the image offline
Package the preloaded dependencies and diffusers source snapshot into an ARM64 image; the build does not access external package repositories.
Start the service
Start the container and wait for model loading; a cold start takes approx. 80 seconds until /health is ready.
Connect your application
Submit text-to-image tasks through HTTP and choose service tiers by resolution and step count.
03 / Performance tests
MegaCube performance test results
2026-09-21; one MegaCube; NVIDIA GB10; 128 GB unified system memory. Fixed prompts and seed=42, with 40 steps by default; durations measure server-side generation only.
| Tier | Configuration | Time per image | Best for |
|---|---|---|---|
| Fast | 1024×1024 / 20 steps | 26.6 s | Drafts, previews, and rapid iteration |
| Balanced | 1536×1536 / 30 steps | 93.9 s | Routine production use |
| Balanced | 2048×2048 / 20 steps | 134.4 s | Routine production use |
| Default (quality) | 2048×2048 / 40 steps | 249.1 s | Final images and the quality baseline |
| Recommended production limit | 2752×2752 / 40 steps | 538.2 s | Specialized high-resolution output |
| Resolution | Pixel count | Duration | Per megapixel |
|---|---|---|---|
| 256×256 | 0.07 MP | 4.7 s | 71.6 s/MP |
| 512×512 | 0.26 MP | 11.7 s | 44.8 s/MP |
| 768×768 | 0.59 MP | 30.2 s | 51.2 s/MP |
| 1024×1024 | 1.05 MP | 53.7 s | 51.2 s/MP |
| 1280×1280 | 1.64 MP | 85.8 s | 52.3 s/MP |
| 1536×1536 | 2.36 MP | 126.7 s | 53.7 s/MP |
| 1792×1792 | 3.21 MP | 187.5 s | 58.4 s/MP |
| 2048×2048 | 4.19 MP | 249.1 s | 59.4 s/MP |
| 2400×2400 | 5.76 MP | 390.3 s | 67.8 s/MP |
| 2752×2752 | 7.57 MP | 538.2 s | 71.1 s/MP |
| 3072×3072 | 9.44 MP | 734.0 s | 77.8 s/MP |
| Resolution | 20 steps | 30 steps | 40 steps |
|---|---|---|---|
| 1024×1024 | 26.6 s | — | 53.2 s |
| 1536×1536 | 63.8 s | 93.9 s | 124.5 s |
| 2048×2048 | 134.4 s | 198.1 s | 252.2 s |
| Scenario | System memory used | Available headroom |
|---|---|---|
| After model loading (idle) | Approx. 40 GiB | Approx. 81 GiB |
| During 1024²–2048² generation | Peak approx. 77 GiB | Minimum 43.7 GiB |
| After 2752² generation | — | 22.2 GiB |
| During 3072² generation | Approx. 120 GiB | 1.5 GiB, near the limit |
- Time grows approximately linearly with pixel count; aspect ratios with the same pixel count take roughly the same time.
- Composition remains consistent between 20 and 40 steps. The 20-step output provides a practical balance of speed and quality.
- 2752×2752 is the recommended resolution limit for production use. Generation at 3072×3072 completes but leaves only approx. 1.5 GiB of memory headroom, so this resolution is not recommended for customer-facing performance commitments.
- Native transparent RGBA output is supported. Slight halos and stray pixels appear around subject edges, so edge cleanup is recommended for commercial assets.
- Qwen Research License is not a standard open-source license; assess licensing before commercial release and disclose it in distributed artifacts.
Plan model deployment and application integration
Evaluate device configurations and inference services for your model size, workloads, and concurrency needs.