whatmodel.sh / image / Emu3.5
Emu3.5
↗ share
+ contribute
Emu3.5 is an open-weights image model from BAAI, released October 2025. Its published size is 34.1B parameters. Licence: apache-2.0.
Model information sources
Release date ↗ · Weights licence ↗ · Parameter count ↗
Whole Emu3.5 autoregressive model (31.2B transformer layers + 2.9B embeddings), per the makers paper; excludes the separate Emu3.5-VisionTokenizer.
Score within the LMArena text-to-image comparison set only.
$ how it compares
The image models ranked either side of Emu3.5, out of 197 with a score, then more from BAAI.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 67 |  Recraft V4.1 Flash | 57±3 | api $0.0070/image | ||
| 68 |  HunyuanImage 3.0 | 57±2 | api $0.080/image +local | ||
| 69 | Emu3.5 | 57±5 | local | ||
| 70 |  JoyAI-Image-Edit | 57±3 | comparison only | ||
| 71 |  MAI-Image-1 | 57±3 | comparison only | ||
| more from BAAI | |||||
| 48 | Emu3.5-Image | 60±5 | local | ||
| 136 | OmniGen2 | 42±2 | api $0.15/image +local | ||
| 164 | Emu3 | 36±2 | local | ||
$ what it's good at
text to image▂▅▅▃▄▄█▃▁▂▁▁ #54 of 126 * · 91% of its score · 1 test
Beats HunyuanImage 2.1 here, though they rank higher overall. Trails Qwen-Image-2512 here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| TIIF-Bench v3 (short prompts, TIIF-Evaluator) | 85.9% ✎ | ▂█····▄▄▄··▂ beats 5/11 | 91% |
$ where to run it
Other ways to run it
- Run it as published · needs ~84.5 GB, NVIDIA only · from huggingface.co/BAAI (official)
git clone https://github.com/baaivision/Emu3.5 && cd Emu3.5 && pip install -r requirements/transformers.txt && pip install flash_attn==2.8.3 --no-build-isolation && CUDA_VISIBLE_DEVICES=0,1 python inference.py --cfg configs/example_config_visual_guidance.py && python src/utils/vis_proto.py --input outputs/emu3p5
Official weights and documentation: huggingface.co/BAAI ↗
$ how the score was worked out · TIIF-Bench: How Does Your T2I… 91%
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 |
|---|---|---|
 TIIF-Bench: How Does Your T2I Model Follow Your Instructions? (arXiv v3) | 1 | 91% |
| prior · pull toward the middle | 9% | |
| weighted average | = 57.0 |
$ 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