whatmodel.sh / image / Qwen-Image-Edit
Qwen-Image-Edit
↗ share
+ contribute
Qwen-Image-Edit is an open-weights image model from Alibaba, released August 2025. Its published size is 20.43B parameters. Licence: Apache 2.0.
Model information sources
Release date ↗ · Parameter count ↗
MMDiT transformer only: HF Hub safetensors total for the official repo (pinned revision), 20,430,401,088 = the sum of the transformer/ shard headers. The Qwen2.5-VL text encoder and the VAE are not included (convention of qwen-image = 20). The card says the model is 'Built upon our 20B Qwen-Image model'. Observed 2026-10-07.
Score within the LMArena text-to-image comparison set only.
$ how it compares
The image models ranked either side of Qwen-Image-Edit, out of 197 with a score, then more from Alibaba.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 78 |  Z-Image-Turbo | 55±1 | api $0.0050/image +local | ||
| 79 |  LLaDA-Image | 55±3 | local | ||
| 80 | Qwen-Image-Edit | 55±2 | api $0.030/image +local | ||
| 81 |  FLUX.2 [klein] Base 9B | 55±4 | comparison only | ||
| 82 | FLUX.2 [klein] 9B | 55±2 | api $0.015/image +local | ||
| more from Alibaba | |||||
| 11 | Qwen-Image-3.0-Pro | 75±2 | api $0.040/image | ||
| 18 | Qwen-Image-2.1 | 69±2 | local | ||
| 20 | Qwen-Image-2.0-Pro (2026-06-22) | 67±2 | api $0.075/image | ||
$ what it's good at
image to image / editing▁▁·▃▂▃█▃▃▄▁▁ #39 of 62 · 65% of its score · 7 tests
Beats GPT Image 1 Mini, FLUX.2 [dev], P-Image-Edit here, though they rank higher overall. Trails FLUX.2 [klein] 9B here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Arena image edit | 1241.5 ✎ | ▂▁▁▂▆▇█▄▆▅▁▂ beats 20/53 | 16% |
| CEdit-Bench-CN overall (G_O, GPT-4o judge) | 74.6% ✎ | ▃·········█▂ beats 4/7 | 14% |
| CEdit-Bench-EN overall (G_O, GPT-4o judge) | 75.2% ✎ | ▄·▄···▄██··▄ beats 4/7 | 8.5% |
| GEBench GUI state generation (English GE-Score) | 33.0% ✎ | ▄▄····▄▄█▄·█ beats 1/8 | 8.0% |
| UnicBench (EN, overall) | 77.3% ✎ | ▂····▂▄▂▄·█▂ beats 7/11 | 7.3% |
| GEditBench v2 (overall Elo) | 1010.0 ✎ | ▃▂····▂█▂▃▂▃ beats 7/14 | 6.5% |
| MultiTextEdit cross-lingual text-in-image editing (SemAvg, LVM judge) | 41.1% ✎ | ▄·▄▄▄▄··█▄·▄ beats 2/8 | 4.9% |
other tests4 not in a task · 33% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| ChartSync chart editing (Overall) | 50.7% ✎ | ▂··▂█▂▄▂▂·▂▂ beats 4/12 | 15% |
| RealPref-Bench image editing (human user-study score) | 80.1% ✎ | █▄▄····▄···▄ beats 4/5 | 7.3% |
| EdiVal multi-turn editing overall (EdiVal-O after turn 3) | 41.9% ✎ | ▂·▂▃·▂·▃▃·▃█ beats 7/14 | 6.1% |
| WEAVEBench in-context multi-turn editing (Avg) | 66.5% ✎ | █▄▄·▄▄▄▄█·▄█ beats 10/12 | 5.2% |
$ where to run it
Through an API
api · fal.ai · $0.030/image · fal.ai ↗ · price sourceUSD for one 1-megapixel output: fal bills this endpoint $0.03 per megapixel (price = rate × 1 MP; fal's rounding for 1024×1024 ≈ 1.05 MP is not stated; larger outputs cost proportionally more). Product: fal.ai serverless image editing. API model: fal-ai/qwen-image-edit (fal describes it as the endpoint for Qwen's Image Editing model, published 2025-08-18, separate from the -2509 and -2511 endpoints). Settings: one input image, 1 MP output, one image. Price basis: published per-megapixel list rate × 1 MP. Alibaba Model Studio's own qwen-image-edit/-plus API models are not used because Alibaba does not state which open checkpoint they correspond to. Currency: USD. Not an Alibaba direct tariff. fal.ai official model page, observed 2026-09-26.
Other ways to run it
version
- Run it as published · needs ~27.2 GB · from github.com/leejet
sd-cli --diffusion-model models/Qwen_Image_Edit-Q4_K_M.gguf --vae models/qwen_image_vae.safetensors --llm models/Qwen2.5-VL-7B-Instruct.Q8_0.gguf --llm_vision models/Qwen2.5-VL-7B-Instruct.mmproj-Q8_0.gguf --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r input.png -p "change 'flux.cpp' to 'edit.cpp'"
Official weights and documentation: huggingface.co/Qwen ↗
$ how the score was worked out · Meituan LongCat — LongCat-Ima… 22%, LMArena leaderboard dataset 16%, ChartSync: A Benchmark for Vi… 15% and 7 more
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json