OpenMOSS, a collaboration between the Fudan NLP team and teams such as Shanghai Chuangzhi, has released MOSS-Image, an open-source image generation model. Built on a 5B-parameter single-stream DiT architecture, the model is pre-trained from scratch. Addressing issues like text spelling errors, paragraph truncation, and element misalignment in traditional text-to-image models for posters, infographics, and similar scenarios, MOSS-Image supports prompt inputs of up to 4K tokens, enabling precise rendering of long Chinese and English texts, along with editable structured JSON Layouts. The model employs the Muon–AdamW training recipe, utilizing Qwen3.5-9B as the text encoder and FLUX.2VAE as the image decoder, trained on approximately 3 billion images with multi-granularity captions, text, and structured bounding box supervision. Additionally, it incorporates an intent-aware prompt enhancement external planning module to expand generation instructions. In multiple evaluations, MOSS-Image excelled in dimensions such as long-text rendering, structured layouts, and end-to-end preferences. Its relevant weights have been open-sourced under the Apache-2.0 license, integrating multiple mainstream ecosystem pathways and supporting LoRA and full-parameter fine-tuning. However, the current model is purely text-to-image and cannot directly accept reference images or perform native image editing, with its prompt enhancement module relying on external large models. Nevertheless, MOSS-Image offers a new open-source option for developers requiring local deployment and precise typesetting.