Aligned metal movable type with raised letters ready to print

· Translated from the Spanish original

Image models

MOSS-Image, with 5 billion parameters, aims to render readable paragraphs inside an image

On October 16 MOSI.AI published a 5B model that, on its own benchmarks, covers 96.3% of the requested long text. It has not yet released the weights.

What happened

  • On October 16, 2026, MOSI.AI posted to arXiv the technical report for MOSS-Image, a 5-billion-parameter diffusion transformer trained from scratch.
  • The model generates images with long text in English and Chinese and with control over layout. It includes a module that turns the user’s request into a detailed content and layout specification.
  • The authors present two benchmarks of their own: LongLongTextBench, where the model reaches 96.3% text coverage, and RealLayoutArena, where it ranks second in full-system comparison.
  • The abstract says the authors “will release the model, training recipe and evaluation framework,” so the weights were not yet confirmed as available.

Why it matters

  • Writing a whole paragraph inside an image, without warped letters, is what today separates a useful poster from a decorative picture. For anyone producing graphic pieces in Chile, that is where the real cost of manual fixing lies.
  • Language is the limit: the model was evaluated in English and Chinese. Nobody tested accents, tildes or opening punctuation marks, which is exactly where these systems fail in Spanish.
  • The 96.3% comes from a benchmark the authors themselves created, and the lead holds “among open-weight models.” It is a first-party measurement.

The number

96.3% average text coverage on LongLongTextBench, a benchmark created by the authors.

Context

Open-weight image generation has been shrinking in size: FLUX 3 Klein dropped to 5.6B parameters and runs on an 8 GB GPU. At 5B, MOSS-Image falls in the same category of models that can run on a desktop computer. Editing what was generated, as in InstructMesh for 3D models, remains another pending stage.

What’s next

  • No date for publishing the weights, training recipe and evaluation framework.
  • No announced timeline for Spanish or other languages with diacritics.
  • A GitHub repository (OpenMOSS/MOSS-Image) exists where the final license should appear.

Bottom line

The promise of image models has shifted from “painting” to “composing”: it is no longer enough for the image to look good, it has to contain the right text. That bar is more demanding than the previous one and is still measured with home-made rules.

Sources

Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.

← All notes