The OpenMOSS team has released and open-sourced an image generation model, MOSS-Image, with 5B parameters. Weights for both the Base and SFT versions are published on Hugging Face under the Apache-2.0 license, the GitHub repository comes with Chinese documentation and a ComfyUI workflow, and a technical report was released alongside. In its official blog, the team describes the model like this:
“A 5B open-weight image generation system that brings long text, structured layout and fine-tunable deployment into a single generation path.”
Long text and editable layouts
Text-to-image models have long struggled to render large blocks of text, and that is exactly what MOSS-Image is built for. It accepts prompts of up to roughly 4K tokens, and long passages in both Chinese and English keep their line breaks and paragraph structure.
The other headline feature is structured layout: a JSON file describes the canvas, text blocks, objects and bounding boxes, directly specifying their positions, sizes and relationships. Designers can revise the structure as many times as they like before rendering, then generate once the layout is settled, which saves the back-and-forth of trying to coax the right result out of a prompt.
Technically, it is a single-stream Diffusion Transformer pretrained from scratch, optimized with a combination of Muon and AdamW. The text encoder is Qwen3.5-9B, and image decoding borrows the VAE from FLUX.2. Training used about 3 billion images carrying multi-granularity captions, text supervision and bounding-box annotations, with resolution raised step by step to 1024×1024 and text context extended to 4096 tokens.
Benchmark numbers
The team's own long-text benchmark, LLTB, contains 120 questions each in Chinese and English, with an average target text of about 1,340 characters. MOSS-Image averages 96.3 across the six metrics, which the team says is the highest among open-weight models; Qwen Image 2512 in the same table scores 64.9. On the longest English task it scores 97.1, and on the longest Chinese task 92.4.
The gap is wider on the layout benchmark 7Bench:
| Model | mIoU | Layout AUC |
|---|---|---|
| MOSS-Image | 0.8786 | 0.7031 |
| Ideogram 4 | 0.4449 | 0.2046 |
| Nano Banana Pro | 0.3296 | 0.1628 |
| GPT Image 2 | 0.3042 | 0.1554 |
In RealLayoutArena, however, an end-to-end human preference test, GPT Image 2 ranks first with about 1844 points, MOSS-Image second with about 1699, and Seedream 5 Pro third. The team also notes that the three benchmarks use different protocols, so their scores are not interchangeable and do not add up to a unified ranking. These figures come from the team alone, and no third-party reproduction has been seen yet.
What teams in China can do with it
For teams building Chinese-language design tools, several conditions look favorable. Apache-2.0 permits commercial use, and at 5B the model is far lighter than image models with tens of billions of parameters. By rule-of-thumb estimates, a single consumer GPU with ample VRAM can handle inference and LoRA fine-tuning, though the exact memory requirements should follow the official documentation. It supports Diffusers, Gradio and ComfyUI, as well as LoRA and full fine-tuning, and those who would rather not self-host can use MOSI's hosted API.
As for use cases, long Chinese text is a daily need for e-commerce posters, long-form images, course slides and WeChat official account headers, an area where open-source models have generally been weak and where text often has to be re-typeset by hand after generation. Being able to lock the layout with JSON also makes it easier to plug the model into existing template systems.
The limitations are stated just as plainly. The current version does text-to-image only: it cannot take reference images and has no native image editing. Its prompt-enhancement module relies on an external language model, so the quality of the planning directly affects the output. Workflows that require editing images or swapping backgrounds will still need other models alongside it.
Sources: OpenMOSS official blog, CocoLoop, MOSS-Image technical report, Hugging Face model page; the technical report was used to verify the parameter count, license, and the LLTB and 7Bench scores.