MOSS-Image
modelYour notes
An open-weight text-to-image model from the OpenMOSS team at the Shanghai Innovation Institute, released October 16, 2026 as MOSS-Image-Base and MOSS-Image-SFT under Apache 2.0. The generator is a 5B single-stream diffusion transformer (DiT) pretrained from scratch; the shipped config has 42 layers, width 3072 and 24 heads. The 5B figure counts the DiT only: the frozen text encoder is a Qwen3.5 model (Qwen3_5Model, hidden size 4096) and the autoencoder is FLUX.2's. SII says the pretraining ran on domestic PPU accelerators ("依托国产 PPU 算力从零开展预训练"), and the README lists heterogeneous-GPU training as a feature. Prompts can run to 4K tokens and can carry JSON layouts that fix where text and objects go.
The technical report (announced on arXiv October 19, credited to MOSI.AI, the OpenMOSS spin-out) targets dense text and layout, where it says open-weight models trail closed systems. It pairs the DiT with an agentic prompt-enhancement module and introduces two evaluations. Self-reported: average text coverage 96.3% on its LongLongTextBench (240 tasks; 97.1% English, 92.4% Chinese) against 64.9% for Qwen Image 2512. On its RealLayoutArena Elo, the full system with prompt enhancement scores 1698.7, second to GPT-Image2 (1844.4) and ahead of Seedream 5.0 Pro (1661.2). With all generators given the same enhanced prompts, the MOSS-Image backbone ranks first at 1807.3, ahead of ERNIE Image (1748.5). Training-data scale is not given in the SII post.
Adoption was fast. The release had day-0 support in SGLang, vLLM-Omni and ComfyUI, and SII reports more than 2 million downloads of the Comfy-Org distribution within eight days. The HuggingFace API shows about 4.5 million on that repo by November 3, and the GitHub repo has about 2.5K stars. SII counted dozens of community adapters, fine-tunes and quantizations per checkpoint by October 25. The team behind it is Xipeng Qiu's OpenMOSS group, and SII describes it as student-led (12 PhD students). The MOSI API platform began serving it on October 20.
Model Details
Variants
| Name | Parameters | Notes |
|---|---|---|
| MOSS-Image-Base | 5B | Base text-to-image model (DiT backbone; frozen Qwen3.5 text encoder and FLUX.2 VAE not counted) |
| MOSS-Image-SFT | 5B | Fine-tuned from MOSS-Image-Base on curated data |