A 256M model that answers typed questions about an image with calibrated odds
Emerged finds new products worth a look, daily — what’s new today
Weights are Apache-2.0 and already native in ComfyUI, so a small team can put typography-heavy image work on its own hardware without a licence conversation.
MOSS-Image is a diffusion transformer released as a base checkpoint and an instruction-tuned one. Two things separate it from the usual open release: it is built to put legible words inside a picture, which most open models still fumble, and it takes a JSON prompt naming objects, the text they carry and where each sits. Prompts run to 4K tokens, and the 5B backbone is small enough to fine-tune. Weights sit on Hugging Face and ModelScope, drew 8,557 downloads last month, and need no plugin on recent ComfyUI builds. There are no benchmark tables, only a technical report and samples.
More like this
Render documents whose layout you already know, so labels come for free
A free macOS image editor that copies Photoshop's layout, keys and blend maths
Name the object you want to keep, and it removes the rest
Apache-licensed Italian LLMs for five European languages, 2B and 15B
Pull clean text from images and screenshots with a pip install and no GPU