MOSS-Image 5B tops open-weight long-text rendering at 96.3%

MOSS-Image, a 5B single-stream diffusion transformer pretrained from scratch, hits 96.3% average text coverage on LongLongTextBench, which the authors' technical report describes as state-of-the-art among open-weight models. The paper frames the gap plainly: "open-weight text-to-image models still struggle with dense text and intricate layouts."

On RealLayoutArena, a 61-task evaluation spanning 20 posters, 20 infographics, and 21 user interfaces, MOSS-Image's end-to-end system ranks second at 1,698.70 Elo behind GPT Image 2 at 1,844.35. Its backbone generator ranks first at 1,807.34 Elo among models evaluated with shared enhanced prompts.

LongLongTextBench packs 240 bilingual tasks averaging 1,339.8 characters per target, split evenly between English and Chinese. The report places Qwen Image 2512 at 64.9 Mean-6 on the same benchmark, a 31.4-point gap. The authors acknowledge the system has "no direct image-conditioning or reference-image support," and the prompt-enhancement module leans on GPT-6 as its external language model.

The release lands in a dense stretch for open-source model releases that our tracker has logged 356 times in the past 90 days, and extends a steady run of open-weight text-to-image work covered on our AI photo desk this fall.