As of: 2026-11-13 13:49 CET.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/m/moss-image/

# MOSS-Image

MOSI / OpenMOSS, MOSS family. Status: current. Type: image generation. Released: 2026-10-16.

## How to call it

| Provider | Model id | Endpoint | Docs |
|---|---|---|---|
| Hugging Face (SFT) | `OpenMOSS-Team/MOSS-Image-SFT` |  | https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT |
| Hugging Face (Base) | `OpenMOSS-Team/MOSS-Image-Base` |  | https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base |
| ComfyUI | `Comfy-Org/MOSS-Image` |  | https://huggingface.co/Comfy-Org/MOSS-Image |
| MOSI API platform |  |  | https://platform.mosi.cn/ |
| GitHub |  |  | https://github.com/OpenMOSS/MOSS-Image |

## Pricing

- Price not published.
- Last checked: 2026-10-26

## Specs

- Input: text
- Output: image
- Context window: not published
- Max output: not published
- Open weights: yes
- Licence: apache-2.0

## Where it stands

- Newer image generation models in the registry since its release: 2:
  - GPT Image 2.6 Flare (2026-11-04) https://postcutoff.com/m/gpt-image-2-6-flare/
  - GPT Image 2.6 Sunburst (2026-11-04) https://postcutoff.com/m/gpt-image-2-6-sunburst/
- Weights are public under a permissive licence (Apache-2.0).

## What stands out

- Dense bilingual text and JSON layouts: Renders long Chinese and English text and accepts structured JSON layouts with bounding boxes and prompts up to 4,096 tokens; the report claims the best open-weight score on its LongLongTextBench (96.3% text coverage). Source: https://arxiv.org/abs/2610.23374
- Trained without AI-generated images (found after launch): The team removed AI-generated images from training data, which community developers (e.g. Ostris) suggest explains its more natural look (Pandaily). Source: https://pandaily.com/moss-image-1-million-downloads-four-days-hugging-face-trending-community-lora-quantized-versions

## Notes

5B single-stream diffusion transformer trained from scratch, frozen Qwen3.5-9B text encoder, FLUX.2 small VAE decoder; trained on a domestic PPU compute platform (Pandaily). Diffusers pipeline needs trust_remote_code. Community 8-step and 4-step distillations, INT8/GGUF/MLX builds.

Open-weight (Apache 2.0) 5B text-to-image model from MOSI and the OpenMOSS team, aimed at text-dense graphics and layouts.

Model page: https://postcutoff.com/m/moss-image/