What is MOSS-Image?
MOSS-Image is a text-to-image model from OpenMOSS, the lab behind the MOSS models, released on 16 October 2026 as two checkpoints: MOSS-Image-Base and MOSS-Image-SFT, a fine-tune on curated data that is the one most people will run. Its core is a 5-billion-parameter diffusion transformer trained from scratch, paired with a Qwen3.5 text encoder and the small FLUX.2 VAE decoder, and it accepts prompts of up to 4K tokens in English or Chinese.
Besides plain prompts and tag lists, it takes a JSON layout: a scene caption plus a list of items, each with a description, optional text to render and a pixel bounding box. That is how it controls where headlines, labels and objects land, which most open models leave to chance. The technical report also describes an agentic prompt enhancer that turns a short request into such a layout; OpenMOSS ships it as an agent skill.
Weights and inference code are under Apache 2.0, with day-one support in SGLang, vLLM-Omni and ComfyUI, and the models are also sold through the MOSI API Platform.
Is MOSS-Image worth it?
MOSS-Image is built for a job open image models do badly: dense text in a planned layout. Posters, slides, menus and infographics need a headline in one place, a caption in another and the words spelled right, and the JSON layout input lets you state each item's text and bounding box rather than hope the prompt is read that way. The licence is the other draw. Apache 2.0 allows commercial use with no separate agreement, which Qwen-Image-2.1's research licence does not.
The evidence is thinner than the pitch. The technical report scores 96.3% average text coverage on LongLongTextBench, first among open-weight models, and puts the system second end to end on RealLayoutArena, with its backbone first when every generator gets the same enhanced prompts. OpenMOSS designed both benchmarks. We found no independent arena ranking yet, so how it compares on general image quality with Qwen-Image-2.1 or FLUX.3 is not established.
The hardware story is better than the download size suggests. The SFT repository holds about 27 GB: a 10.2 GB transformer, a 16.8 GB text encoder and the VAE. The project's own measurement in its low-memory mode is a 3.65 GiB GPU peak for one 1024 x 1024 image at 50 steps, with 14 GiB of system RAM, at the cost of speed; it has no published timing for any mode. The default is 50 steps with two transformer passes each, so it is not quick, and a community 8-step distillation exists but is not official.
The ecosystem formed fast. Within 17 days of launch the repository had about 2,500 GitHub stars, and community projects added fine-tuning (KohakuMossTrainer claims adapter training in 6 GB of VRAM), FP16 weights and a LoRA to reduce vertical colour banding in the SFT model, a flaw worth checking before you commit. There is no editing mode: it generates images from text and layouts only.
Where it earns its place
A marketing operations lead at a restaurant chain with 40 sites
Produce menu boards and promotion posters with exact prices and dish names
A template script fills a JSON layout with each site's text and box positions and renders at 1024 x 1024 or larger. Apache 2.0 means no licence negotiation; check every output, since text accuracy is reported at 96.3% coverage, not 100%.
A developer building a slide and infographic generator
Let an LLM plan a layout and have an open model draw it
The bundled prompt-enhancer skill turns a request into the layout JSON the model reads, and SGLang or vLLM-Omni serve it. Plan for 50-step generations, or test the unofficial 8-step community distillation.
A bilingual design studio in Shanghai
Render long Chinese and English copy inside product visuals
Prompts run to 4K tokens and the report measures long-form Chinese and English text rendering. Self-host on a single consumer GPU in low-memory mode, or fine-tune a house style with DiffSynth-Studio or KohakuMossTrainer.
Before you commit
- Setup
- Python 3.10+, PyTorch 2.9+, Diffusers 0.38+ and Transformers 5.3+ on an NVIDIA GPU; the custom pipeline loads with trust_remote_code. A Gradio app adds memory profiles, and ComfyUI supports it natively.
- Integrations
- Diffusers, ComfyUI, SGLang, vLLM-Omni, DiffSynth-Studio for training, and an agent skill for prompt and layout enhancement; hosted access through the MOSI API Platform.
- Data and privacy
- Self-hosted under Apache 2.0, so prompts and images stay on your hardware. Training data is not itemised beyond the technical report.
- Cost at scale
- Weights free. Self-hosting needs about 27 GB of disk and, per the project, 3.65 GiB of GPU memory at minimum in low-memory mode; generation speed and MOSI API prices are not published.
- Limits
- Text-to-image only, no editing. Image sides must be multiples of 16; layout boxes must match the output size. Defaults are 50 steps at CFG 3, and community reports note vertical banding in the SFT checkpoint.
What’s good
- JSON layouts set the position of each text block and object
- Apache 2.0, commercial use allowed
- 3.65 GiB GPU peak in low-memory mode, per the project
- Day-one SGLang, vLLM-Omni and ComfyUI support
What’s not
- Headline scores come from benchmarks OpenMOSS built
- About 27 GB to download; 50 steps by default
- Community reports vertical colour banding in the SFT model
- No image editing; API prices not published
Best for
Teams that need legible text and controlled placement in generated graphics and want a licence they can ship under.
MOSS-Image alternatives
MOSS-Image: common questions
- Is MOSS-Image worth it?
- The open image model to try when the words and their placement matter more than the look, with every quality claim still resting on benchmarks its makers built. We rate it 3.8 out of 5.
- How much does MOSS-Image cost?
- Free open weights under Apache 2.0, including commercial use. A hosted version is on the MOSI API Platform; its prices are not published on the pages we checked.
- What are the downsides of MOSS-Image?
- Headline scores come from benchmarks OpenMOSS built; about 27 GB to download; 50 steps by default; community reports vertical colour banding in the SFT model; no image editing; API prices not published.
- Who is MOSS-Image best for?
- Teams that need legible text and controlled placement in generated graphics and want a licence they can ship under.
- How much does MOSS-Image cost for a team or in production?
- Weights free. Self-hosting needs about 27 GB of disk and, per the project, 3.65 GiB of GPU memory at minimum in low-memory mode; generation speed and MOSI API prices are not published.
- Is MOSS-Image safe for company data?
- Self-hosted under Apache 2.0, so prompts and images stay on your hardware. Training data is not itemised beyond the technical report.
- What are the best MOSS-Image alternatives?
- The closest alternatives we have reviewed are Qwen-Image-2.1, FLUX 3 Image Klein and Z-Image-Turbo.
Answers checked Nov 2, 2026.