MOSS-Image

OpenMOSS's 5B image model draws long English and Chinese text and follows JSON layouts, under Apache 2.0, from about 4 GB of VRAM.

In short. MOSS-Image is an AI image model whose weights you can download. It is built for pictures that carry a lot of text, such as posters, infographics and notes, in English and Chinese, and you can give it a JSON layout that says where each block of text and each object goes. It is released under Apache 2.0, so commercial use is allowed. The developer's low-memory profile runs it on a GPU with about 4 GB of memory, though it then needs about 14 GB of system RAM.

Summary

MOSS-Image is a pair of text-to-image models from OpenMOSS, a research group under the Shanghai Innovation Institute that works with Fudan University and MOSS-AI. MOSS-Image-Base and MOSS-Image-SFT were released with code and weights on 16 October 2026, and a technical report followed on 19 October (arXiv 2610.23374). SFT is Base fine-tuned on curated data, and in the developer's example it groups text on a whiteboard more cleanly; the developer points to Base for adapting the model to your own tasks. The pitch is dense, readable text and controllable layout: prompts can run to 4,096 tokens, long English and Chinese passages are rendered in the picture, and a JSON layout lets you give each item a label, a caption, optional text and a bounding box in pixels measured from the canvas centre.

The generator is a 5.08B-parameter single-stream diffusion transformer with 42 blocks, trained from scratch in three stages of rising resolution and text length with a hybrid Muon and AdamW optimizer. It reads prompts through a frozen Qwen3.5-9B text encoder and works in the latent space of the FLUX.2 VAE, so the "5B" label covers only the generator: the repository is 27.2 GB, of which the text encoder is about 16.8 GB and the transformer about 10.2 GB. The default sampler is flow-matching Euler with 50 steps and a guidance scale of 3, and the user guide recommends outputs of 1 to 2 megapixels and says the model can extrapolate from 0.25 to 4 megapixels. It runs through Diffusers with custom pipeline code, natively in ComfyUI, and in SGLang and vLLM-Omni.

The README lists a measured peak for one 1024×1024 image at 50 steps: 3.65 GiB of allocated GPU memory and 14 GiB of system RAM. That figure comes from the low-memory profile, which keeps the transformer in CPU RAM and loads the text encoder and VAE from disk as needed, so expect it to be slower than a run with everything on the GPU. The site therefore lists about 4.2 GB of VRAM (the 3.85 GiB reserved peak) as the minimum, with the RAM caveat. Third-party projects arrived quickly: trainers that list adapter training from 6 GB of VRAM, an 8-step distillation, an FP16 fix and a C++ and CUDA engine.

The report introduces two benchmarks of its own. On LongLongTextBench, a test of long English and Chinese text, MOSS-Image scores 96.3 (Mean-6) against 64.9 for Qwen Image 2512, with 97.1 and 92.4 on the longest English and Chinese settings. For layout, its mean IoU on detected objects is 0.88 and its layout AUC 0.70, against 0.44 and 0.20 for Ideogram 4. On RealLayoutArena, a blind human comparison of nine full systems, MOSS-Image with a prompt-enhancing agent in front of it scores 1698.70 Elo, second behind GPT Image 2 and ahead of Seedream 5.0 Pro, and its backbone alone has the top rating (1807.34) when every generator gets the same enhanced prompts. The enhancer is not part of the download: the repository ships it as an agent skill designed for GPT-6-family agents. This release makes images from text only; the report says reference-image input and editing are future work.

Good for

  • Posters, infographics, notes and UI mock-ups that need long, correct English or Chinese text inside the picture
  • Layout-controlled generation, where a JSON layout fixes where each piece of text and each object goes, and tools that edit those layouts
  • LoRA and full fine-tuning on a small base, since Base and SFT are both published and third-party trainers support them
  • Commercial work, because weights and code are plain Apache 2.0 with a LICENSE file in each repository

Watch out for

  • The text and layout scores come from OpenMOSS's own benchmarks, and the second-place end-to-end result depends on a prompt enhancer built for GPT-6-family agents that is not part of the weights
  • The "5B" label covers only the generator; the 27.2 GB repository includes a 16.8 GB Qwen3.5-9B text encoder
  • The 3.65 GiB GPU peak is for the low-memory profile, which needs 14 GiB of system RAM and a fast SSD, and should be slower than a full-GPU run
  • It makes images from text only, with no reference-image input or image editing in this release
  • Setup needs recent libraries (the model card lists Diffusers 0.38, Transformers 5.3 and PyTorch 2.9) and custom pipeline code that you must allow with trust_remote_code

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of MOSS-Image. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
MOSS-Image-SFT (BF16) BF16 5.08B 27 GB 4.2 GB Laptop or phone
MOSS-Image-Base (BF16) BF16 5.08B 27 GB 4.2 GB Laptop or phone

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more. The developer states a minimum of 4.2 GB of GPU memory.

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is MOSS-Image?
MOSS-Image is an AI image model whose weights you can download. It is built for pictures that carry a lot of text, such as posters, infographics and notes, in English and Chinese, and you can give it a JSON layout that says where each block of text and each object goes. It is released under Apache 2.0, so commercial use is allowed. The developer's low-memory profile runs it on a GPU with about 4 GB of memory, though it then needs about 14 GB of system RAM.
Can I use MOSS-Image commercially?
Yes. MOSS-Image is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is MOSS-Image open source?
MOSS-Image is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run MOSS-Image on my own computer?
Yes. The smallest variant, MOSS-Image-SFT (BF16), needs about 4.2 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was MOSS-Image released?
MOSS-Image was released by OpenMOSS on October 16, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Z-Image Tongyi MAI 6.15B – Apache 2.0 (Permissive) 16 GB Laptop or phone
Qwen-Image-2.1 Alibaba Qwen 7.1B – Qwen Research License (Non-commercial) 20 GB Gaming GPU
FLUX 3 Image Klein Black Forest Labs 5.6B – FLUX Kommunity (Conditional) 8 GB Laptop or phone
MiniMax H3-Image MiniMax 33.1B – MiniMax H3-Image Community (Conditional) 155 GB Big-memory desktop