What is MOSS-Image? The OpenMOSS 5B text-to-image model explained

MOSS-Image is an open text-to-image model family released by OpenMOSS on October 16, 2026. It turns a written description into an image, and it was built around three things many open models still find hard: rendering legible text inside the picture, following long and detailed prompts, and placing elements where you ask for them.

This page sums up what the official model card, user guide and technical report say, checked on October 22, 2026. Where something is our opinion or an estimate, we say so.

The short version

Publisher	OpenMOSS (the MOSS team behind the MOSS language models)
Released	October 16, 2026
Variants	MOSS-Image-Base and MOSS-Image-SFT
Backbone	5B-parameter single-stream diffusion transformer (DiT)
Text encoder	Qwen3.5-9B, kept frozen
Image decoder	The FLUX.2 small VAE decoder
Prompt formats	Natural language, tag sequences and JSON layouts
Prompt length	Up to 4K tokens
Recommended size	1, 1.5 or 2 megapixels; any aspect ratio, sides in multiples of 16
License	Apache 2.0 for the weights and the bundled inference code
Paper	MOSS-Image Technical Report, arXiv 2610.23374

Base and SFT: which one is which

The family ships as two checkpoints that share one architecture.

• MOSS-Image-Base is the pretrained model. OpenMOSS deliberately did not distil or post-train its sampling behaviour, so that fine-tuners and LoRA trainers get the full training signal. It needs well-chosen settings to look its best.
• MOSS-Image-SFT is Base after supervised fine-tuning on curated data. It is the one meant for everyday generation. The model card's own comparison shows SFT grouping text more cleanly and laying out a busy whiteboard scene more tidily than Base.

If you just want good images, use SFT. If you plan to train a LoRA or a full fine-tune, start from Base. Our Base vs SFT article goes deeper.

What it does well

Text inside images. MOSS-Image is trained to render the words you ask for on signs, labels, posters, packaging and whiteboards. It handles English and Chinese. Put the exact wording in quotes; for precise line breaks, write the line breaks into the prompt. Our prompt guide has worked examples.

Long prompts. The text encoder reads prompts of up to 4K tokens. You can describe a whole scene, a product shot with its materials and lighting, or a poster with several text blocks, without the model dropping the end of your description.

Photoreal and anime. OpenMOSS lists both photorealistic and anime-style images as headline strengths, and the official showcase mixes the two.

Layout control. Besides plain language, MOSS-Image accepts a JSON layout: a short scene caption plus a list of items, each with a description and a bounding box in pixel coordinates. That lets you put a headline at the top, a product in the middle and a logo-free badge in a corner. OpenMOSS still labels this mode experimental. See our JSON layout guide.

Fine-tuning. The Base checkpoint is meant to be adapted. Community trainers such as DiffSynth-Studio and KohakuMossTrainer added full fine-tuning and LoRA support within days of release; KohakuMossTrainer reports adapter training on as little as 6 GB of VRAM.

Default sampling settings

These are the defaults from the model card. Hosted services, this one included, choose them for you.

Setting	Default
Sampler	Flow-matching Euler
Steps	50 (the guide recommends 50 or more)
Guidance (CFG)	3, recommended range 3 to 6
Shift	5
Layer skipping	Negative branch, layer 22
Negative prompt	Empty

The user guide suggests raising CFG a little for short prompts; 3 is enough for long ones. Each step runs the transformer twice, which is one reason 50-step generation is not instant even on a large GPU.

What it needs to run locally

OpenMOSS does not publish a single minimum VRAM figure. The bfloat16 files are large: about 10 GB for the diffusion transformer and roughly 18 GB more for the Qwen3.5-9B text encoder. Keeping everything on the GPU therefore needs a card in the high-end data-centre class; consumer cards rely on moving components between CPU and GPU, quantised weights, or both, which works but is slower. The official Gradio app offers GPU-resident, component and sequential memory modes for exactly this reason. Our ComfyUI guide covers the local route step by step.

Where to try it

• In your browser, here. MOSS Image AI runs the model on hosted GPUs. New accounts get free credits, and you choose the style, the aspect ratio, Standard (about 1 MP) or HD (about 2.25 MP) and up to four images per prompt.
• The official demo. OpenMOSS hosts a Hugging Face Space. It is free, but it runs on shared ZeroGPU hardware with per-user quotas and queues.
• The official API. OpenMOSS's MOSI platform added the model on October 20, 2026.
• Your own machine. Download the weights from Hugging Face or ModelScope and run them with Diffusers, the Gradio app, SGLang, vLLM-Omni or ComfyUI.

Is MOSS-Image free for commercial use?

The weights and the bundled inference code are released under the Apache License 2.0, which allows commercial use. Two caveats apply to any image model. A permissive model licence says nothing about the rights in what you depict, so trademarks, real people and copyrighted characters remain your responsibility. And a hosted service adds its own terms on top; ours are in the terms of service.

Sources

• MOSS-Image-SFT model card
• MOSS-Image on GitHub, README and user guide
• MOSS-Image Technical Report, arXiv 2610.23374

MOSS Image AI is an independent service and is not affiliated with OpenMOSS.