On October 29, 2026, OpenMOSS released the open-weight image generation model MOSS-Image on its official blog. Built on a 5-billion-parameter single-stream DiT and pre-trained from scratch, MOSS-Image targets text-heavy, layout-sensitive work such as posters, infographics, comics, UI mockups, and teaching materials. Its weights are released under the Apache-2.0 license in Base and SFT versions, with Diffusers, Gradio, and ComfyUI paths provided.
Two inputs for two ways of working
MOSS-Image accepts two kinds of input. A plain natural-language prompt suits quick drafts; a JSON Layout suits images that must be reproduced and revised, because it spells out the canvas, text, objects, and bounding boxes, fixing each element's position, size, and spatial relationships. Both inputs share one text-conditioning path, so a user can move from describing an idea to specifying a layout without switching models. The model supports prompts of up to about 4K tokens, preserves line breaks and paragraph structure, and renders long text in Chinese and English, fitting titles, captions, dialogue, tables, and other information-dense visuals. An optional intent-aware prompt-enhancement module can expand a short request into a content and layout specification, but it is an external planning module; the local backbone also accepts ordinary prompts or a JSON Layout on its own.
Read the benchmark numbers separately
The technical report gives three sets of numbers. On the LongLongTextBench long-text benchmark, MOSS-Image scores a Mean-6 of 96.3, the highest among open-weight models in that evaluation; on the 7Bench layout benchmark, its detection-box mIoU is 0.8786; on the end-to-end preference benchmark RealLayoutArena, its Elo is 1698.70, second behind GPT Image 2 at 1844.35. These numbers come from different evaluation protocols and cannot be combined into one cross-task leaderboard. The long-text score measures text coverage and rendering quality, not factual accuracy, and the end-to-end ranking also reflects each system's prompt planning and product pipeline; when backbones alone are compared under the same enhanced prompts, MOSS-Image scores 1807.34.
Who it fits, and where its limits lie
MOSS-Image fits teams that need local deployment, reproducible layouts, or further training on their own data. The Base version is a starting point for continued pre-training, domain fine-tuning, or LoRA, while the SFT version offers more stable text grouping and whiteboard-style layouts out of the box. Its limits are also clear: the current backbone is a text-to-image model only, so it takes no reference image and offers no native image editing, and prompt enhancement depends on an external language model, so planning quality affects the result. Its value is not general-purpose image generation, but turning where text should appear and where each element should sit into explicit, checkable inputs.
