MOSS-Image API: OpenMOSS’s Open 5B Image Model Has No Hosted API Yet
OpenMOSS released MOSS-Image on 2026-10-16: a 5-billion-parameter text-to-image model with open weights under the Apache-2.0 license, built to render long English and Chinese text in images and to follow layouts set by bounding boxes. As of 2026-10-21, none of the 76 provider price lists we check daily offers it as an API, and its Hugging Face pages show no inference provider. For now it runs only on your own GPU.
What OpenMOSS released
The OpenMOSS team (MOSI.AI) published the model weights and code on 2026-10-16, followed by a technical report on arXiv.
- Two checkpoints: MOSS-Image-Base for fine-tuning and LoRA training, and MOSS-Image-SFT, tuned from Base for everyday image generation.
- A 5B single-stream diffusion transformer trained from scratch, with a FLUX.2 VAE decoder.
- Weights and pipeline code under Apache-2.0, a license that allows commercial use.
- Downloads on Hugging Face and ModelScope, and an online demo on Hugging Face Spaces.
What it does
- Renders long English and Chinese text inside images, such as posters, slides and infographics.
- Accepts prompts of up to about 4,000 tokens, written as plain sentences or as tag lists.
- Can take a JSON layout that places each text block and object in a bounding box.
- Uses 50 sampling steps with guidance by default, which means 100 model passes per image.
In its technical report, OpenMOSS reports 96.3% average text coverage on LongLongTextBench, the best among the open-weight models it compared, and second place in an end-to-end comparison on RealLayoutArena. These are the developer’s own figures; we have not tested the model.
How to run it today
The official repository provides a Diffusers pipeline and a Gradio app, and lists day-0 support in ComfyUI, SGLang and vLLM-Omni. It needs an NVIDIA GPU with CUDA; a lower-memory mode moves parts of the model between CPU and GPU. To try it without installing anything, use the Hugging Face demo.
Is there a MOSS-Image API?
Not yet. As of 2026-10-21, none of the 76 provider price lists and model catalogs we check every day lists MOSS-Image, its Hugging Face pages say no inference provider serves it, and OpenMOSS has not published a hosted API or a price. We update this page daily: when a provider lists MOSS-Image, its entry and price appear here.
Image APIs you can call today
If you need an image API now, these hosted models publish prices. Each row shows the official price per image next to the cheapest published price for the same spec, as of 2026-10-21. Every image model: all model prices.
| Model · spec | Official | Cheapest published |
|---|---|---|
| Nano Banana 21K | $0.0670/imageGoogle Gemini API | $0.009306/imageGPTNB GOAPI~0.14× official |
| Nano Banana Pro1K | $0.134/imageGoogle Gemini API | $0.02778/imageOmniallAI~0.21× official |
| GPT Image 21K (1024×1024) · medium quality | $0.0530/imageOpenAI | $0.008333/imageAIGateHub~0.16× official |
| Seedream 5.0 Proup to 1.5K (≤2.61 MP) | $0.0450/imageBytePlus ModelArk | $0.0340/imageEvoLink~0.76× official |
| FLUX.2 Pro1 MP | $0.0300/imageBlack Forest Labs | $0.0200/imageInfer~0.67× official |
| Qwen Image1 MP / default size | — | $0.00580/imageTogether AI |
| Ideogram 4.0Default speed | $0.0600/imageIdeogram | $0.0600/imageRunware1.00× official |
| Recraft V4.11K · raster | $0.0350/imageRecraft | $0.0350/imageOpenRouter~1.00× official |
Sources
- github.com/OpenMOSS/MOSS-ImageOpenMOSS: official documentation, checked
- huggingface.co/OpenMOSS-Team/MOSS-Image-SFTOpenMOSS: official documentation, checked
- huggingface.co/OpenMOSS-Team/MOSS-Image-BaseOpenMOSS: official documentation, checked
- arxiv.org/abs/2610.23374OpenMOSS: official documentation, checked
- huggingface.co/spaces/OpenMOSS-Team/MOSS-ImageOpenMOSS: official documentation, checked
The numbers in this article are recalculated from our price data at every update; the dates above and in the sources show when each was last checked. See all findings.