whatmodel.sh / image / OneCAT-3B
OneCAT-3B
↗ share
+ contribute
OneCAT-3B is an open-weights image model from Meituan (OneCAT team), released September 2025. Its published size is 9B parameters (3B active per token). Licence: apache-2.0.
Catalogue identity notes
onecat-ai/OneCAT-3B (the only released OneCAT checkpoint; HF revision a67e66a, created 2025-09-06, Apache-2.0; report arXiv 2509.03498; contact address at meituan.com). Decoder-only autoregressive unified understanding/generation/editing model. No local command recorded.
Model information sources
Modality-MoE: 9B total, 3B activated, as the makers' report (Table 2) states for OneCAT-3B (Qwen2.5-3B-instruct base with FFNs replicated per modality); whether the image detokenizer is counted is not stated.
Score within the LMArena text-to-image comparison set only.
$ how it compares
The image models ranked either side of OneCAT-3B, out of 197 with a score.
| # | model | score | where |
|---|---|---|---|
| 190 |  HiDream-E1.1 | 26±11 | comparison only |
| 191 |  Stable Diffusion v1.5 | 26±2 | local |
| 192 | OneCAT-3B | 24±5 | local |
| 193 |  LCM Dreamshaper v7 | 23±4 | local |
| 194 |  InstructPix2Pix | 19±3 | local |
$ what it's good at
image to image / editing▃▁·▃▂▃█▃▃▄▁▁ #61 of 62 * · 88% of its score · 1 test
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Edit-Compass (EN, overall) | 24.2% ✎ | ▃▃▂·▃▆▂█·▃·▃ beats 1/20 | 88% |
$ where to run it
Other ways to run it
- Run it as published · needs ~25.1 GB, NVIDIA only · from github.com/onecat-ai
git clone https://github.com/onecat-ai/OneCAT && cd OneCAT && pip3 install -r requirements.txt && hf download onecat-ai/OneCAT-3B --local-dir OneCAT-3B && wget -O infinity_vae_d32reg.pth "https://huggingface.co/FoundationVision/Infinity/resolve/main/infinity_vae_d32reg.pth?download=true" && accelerate launch --num_processes=1 generate_txt2img.py --model_name_or_path="OneCAT-3B" --vae_path="infinity_vae_d32reg.pth" --prompt="Fashion model with iridescent makeup, prismatic light reflections, high-fashion studio setting." --output_dir="output/txt2img"
Weights and documentation: huggingface.co/onecat-ai ↗
$ how the score was worked out · Edit-Compass & EditReward-Com… 88%
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
| source | results | share of the score |
|---|---|---|
 Edit-Compass & EditReward-Compass: A Unified Benchmark for Image Editing and Reward Modeling | 1 | 88% |
| prior · pull toward the middle | 12% | |
| weighted average | = 24.1 |
$ 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