MOSS-Image
by OpenMOSS · Model · Image
promising
Small 5B image generator that one reviewer rates best for its size
New open model on Hugging Face (text-to-image): 5.08B parameters, 35.9 GB of weights, license apache-2.0
AccessOpen weightsNeeds~35.9 GBDownloadabout 10 GB for the SFT checkpoint; INT8 is roughly halfSize5.08B parametersLicenseApache 2.0
Why we say so
AI Search, the only tester, found it much better than Flux Klein and more flexible than Z-Image. It handles text, infographics and box-based layout control, and runs on small GPUs when quantized. Krea 2, which is over twice the size, was slightly better, and it cannot do natural-language editing. All the praise comes from one reviewer.
Evidence
- Oct 205B text-to-image model with layout control via bounding boxes; quantized versions fit low-end GPUs, even 4 GB VRAM or less
- Oct 20Called the best open image generator in its weight class, with broad knowledge and strong text and infographics
- Oct 20Layout-guided workflow placed every boxed element correctly in a complex scene, including the Coke logo, though rendered in Chinese
- Oct 20Krea 2 produced slightly better quality in initial tests, as expected since it is more than twice the size
- Oct 20Seen as understanding more and being more flexible than 6B Z-Image despite being smaller at 5B
Show all 6
- Oct 18Judged much better than Flux 3 Klein in the reviewer's tests
Cost to run
Alternatives
Built on MOSS-Image