whatmodel.sh / image / Qwen-Image-2512
Qwen-Image-2512
↗ share
+ contribute
Qwen-Image-2512 is an open-weights image model from Alibaba, released December 2025. Its published size is 20B parameters. Licence: apache-2.0.
Score within the LMArena text-to-image comparison set only.
$ how it compares
The image models ranked either side of Qwen-Image-2512, out of 197 with a score, then more from Alibaba.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 74 |  P-Image-Edit | 56±4 | api $0.010/image | ||
| 75 |  FireRed-Image-Edit-1.0 | 56±4 | comparison only | ||
| 76 | Qwen-Image-2512 | 56±1 | api $0.020/image +local | ||
| 77 |  Reve V1.1 Fast | 56±4 | comparison only | ||
| 78 |  Z-Image-Turbo | 55±1 | api $0.0050/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
text to image▂▅▅▃▄▄█▃▁▂▁▁ #51 of 126 · 76% of its score · 8 tests
Beats Wan 2.7 Image, HunyuanImage 2.1, Recraft V4.1 Flash here, though they rank higher overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| GenExam (relaxed overall, GPT-5 judge) | 35.3% ✎ | ▄█▂▂▃▂▆▄·▂·▂ beats 8/16 | 36% |
| LMArena text-to-image | 1125.4 ✎ | ▂▂▃▄▇█▄▂▄▂·▂ beats 35/75 | 24% |
| UltraText Bench (Composite, bilingual dense text rendering) | 68.0% ✎ | ▁▂▂█▃▃▂▃·▁·▁ beats 15/22 | 6.4% |
| ImageTime (prompt-only, overall) | 53.2% ✎ | ▄·····▄█▄·▄█ beats 2/7 | 4.8% |
| Qwen-Image-Bench (Q-Judger, v1) lab | 52.1% ✎ | █▅▅·▅▅█·█··▃ beats 6/17 | 2.6% |
| HPSv3 score of Cosmos 3 report text-to-image evaluation images | 11.0300 ✎ | ▄···█··▄·▄·▄ beats 0/5 | 1.5% |
| CVTG-102ch PNED (Chinese character rendering, Cosmos 3 report) | 71.3% ✎ | ▄·······▄█▄▄ beats 2/5 | 0.4% |
| CVTG-500L PNED (English long-prompt text rendering, Cosmos 3 report) | 90.9% ✎ | █······█████ beats 5/5 | 0.0% |
other tests2 not in a task · 24% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| VVRBench accuracy (10,000 verifiable synthetic-scene tasks) | 5.8% ✎ | ▃·█▅▃·▅▃···▃ beats 6/10 | 20% |
| UniGenBench + Physical-AI prompts, All 1,170 (Cosmos 3 report) | 84.3% ✎ | ▄···█··▄···█ beats 2/5 | 3.5% |
$ where to run it
Through an API
api · Replicate · $0.020/image · replicate.com ↗ · price sourceUSD per output image on Replicate's qwen/qwen-image-2512 model (billing: per output image; hardware time not added). Current version served from replicate-pruna/qwen-image-2512 (Pruna-optimized deployment). Product: Replicate model API. API model: qwen/qwen-image-2512. Settings: one output image. Price basis: published list rate. Currency: USD. Replicate official model page, observed 2026-09-25.
Other ways to run it
- Run it as published · needs ~22.4 GB · from unsloth.ai
sd-cli --diffusion-model models/qwen-image-2512-Q4_K_M.gguf --vae models/qwen_image_vae.safetensors --llm models/Qwen2.5-VL-7B-Instruct-UD-Q4_K_XL.gguf --cfg-scale 2.5 --sampling-method euler -v --steps 40 -H 1024 -W 1024 --diffusion-fa --flow-shift 3 -p 'a cat holding a sign' --offload-to-cpu -o outputs/out.png
Official weights and documentation: huggingface.co/Qwen ↗
$ how the score was worked out · GenExam: A Multidisciplinary… 36%, LMArena leaderboard dataset 24%, Verifiable Visual Rewards Tra… 20% and 4 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