A new open image model gets 96.3 on long text in pictures

On 16 October 2026, the OpenMOSS team, a research group in China, released MOSS-Image. The group is part of the Shanghai Innovation Institute. The model makes pictures from text and is for correct text on menus, posters and slides. The code and weights are free in the Apache 2.0 license. The team gives it 96.3 on a long text test.

How it works

MOSS-Image is a diffusion transformer, an image model that makes a picture step by step from noise. A Qwen model with 9 billion parameters reads the input. A designer can also give a JSON layout, a set of data that puts each text part in its box.

The numbers

• The next open model gets 64.9 on the same test.
• GPT Image 2 gets 99.0 and Nano Banana Pro from Google gets 96.6.
• Human judges compared pictures. The system is number 2, behind GPT Image 2.

What is not known

The team did all the tests. No other group has done them again. Number 2 is for the full system with its input tool, not for the model only. The team and news reports give the count of 1.3 million downloads.

Insight: Free image models spread fast, and so does the trust problem

MOSS-Image has free weights. Other persons made versions with 4 or 8 steps, and LoRA add-ons, which are small add-on weights. In October 2026, researchers showed a LoRA add-on on an open Qwen text model. It could send project files to a server. The price was less than $50. This reads as a fast move to get users with free weights, and the maker of each version is not easy to find.