Blog ·

MOSS-Image online: the open image generator that gets text right, no GPU needed

OpenMOSS's MOSS-Image is a 5B open image model that writes clean text inside images and follows long prompts. What it is, its Apache-2.0 license, and how to try it in your browser.

Most open image generators still struggle with one thing: text. Ask for a poster with a headline and you get letters that almost look right. MOSS-Image, released by the OpenMOSS team in October 2026 under the Apache-2.0 license, was built for exactly that: clean text inside images, in English and Chinese, and very long, detailed prompts.

MOSS-ImageOpenMOSS · Apache-2.0A new open 5B image generator that writes clean text inside images, in English and Chinese, and follows very long, detailed prompts.Try it online →

What MOSS-Image is

  • A 5-billion-parameter diffusion transformer, with Qwen3.5-9B as its text encoder and the FLUX.2 autoencoder.
  • Prompts of up to about 4,000 tokens: you can describe a whole poster, item by item.
  • Two official checkpoints: Base, meant for training LoRAs, and SFT, fine-tuned for better-looking images. LiveWeights runs SFT.
  • Apache-2.0 weights and code, so the images can be used commercially.

Why it is hard to run at home

The diffusion model alone is about 10 GB in bf16, and the text encoder another 18 GB. Running it locally means ComfyUI, a 24 GB graphics card and some memory juggling, or quantised files. On LiveWeights it runs on a cloud GPU: you write the prompt and get the image.

How to try it online

  1. Open the MOSS-Image page and sign in.
  2. Describe the image. Put any text you want written in quotes, for example: a poster with the word "EXPLORE".
  3. Pick a format (square, 4:3, 3:4, 16:9 or 9:16) and the quality: Fast uses 28 steps, Best the official 50.
  4. Press Generate. The first run is free. If you like an image, "Redo in best quality" renders it again with the same seed and 50 steps.

Prompt tips

  • Describe the layout: what goes at the top, in the middle and at the bottom.
  • Keep text short and in quotes; one or two lines work best.
  • Name the style explicitly: "flat colours", "photorealistic", "watercolour", "3D render".

The community around it

Within days of the release the community published an image-editing LoRA, an experimental 8-step distillation and fine-tunes such as Kromoss. They are listed on the model page: press Request on the one you would like us to host next.

LiveWeights is not affiliated with the model authors or with the AI Search channel. Facts come from the official pages linked on each model page.