MOSI and OpenMOSS open-source MOSS-Image, a 5B text-and-layout model

MOSI Intelligence and OpenMOSS have released MOSS-Image, a 5-billion-parameter text-to-image diffusion transformer built for dense lettering and precise layout. Base and SFT weights are on Hugging Face and ModelScope under Apache 2.0, with prompts up to 4,096 tokens and JSON bounding-box layout control.

• 5B-parameter model with Apache 2.0 weights on Hugging Face and ModelScope
• Prompts up to 4,096 tokens and JSON layout with bounding boxes
• 96.3% text coverage on Long Long Text Benchmark, 1,698.70 Elo on RealLayoutArena
• Day-zero adapters for Diffusers, ComfyUI, SGLang and vLLM-Omni, plus hosted API