MOSS-IMAGE · OPENMOSS · CHECKED OCTOBER 19, 2026
MOSS Image Download
MOSS-Image is an open 5B text-to-image model from OpenMOSS, released on October 16, 2026 in two versions, Base and SFT. Both are free to download from Hugging Face or ModelScope under Apache 2.0. This page covers which version to pick, what you need to run it, and the community INT8, GGUF and LoRA files.
MossImage is an independent site, not part of OpenMOSS. Community files below are listed from their maintainers' model cards; we have not tested them.
Base or SFT?
| Version | What it is | Pick it for |
|---|---|---|
| MOSS-Image-SFT | Fine-tuned from Base on curated data. OpenMOSS shows clearer text grouping and tidier layouts. | Generating images, posters and text-heavy layouts. |
| MOSS-Image-Base | The pretrained model, with no post-training or distillation of settings like CFG. | Training LoRAs and other fine-tunes. |
Both versions share the same architecture and pipeline: a 5B diffusion transformer, a Qwen3.5-9B text encoder and a FLUX.2 VAE. They take natural-language prompts, tag lists, or a JSON layout that places text and objects. According to the technical report, AI-generated images were filtered out of the pretraining data.
Official downloads
| Version | Hugging Face | ModelScope | License |
|---|---|---|---|
| SFT | OpenMOSS-Team/MOSS-Image-SFT | openmoss/MOSS-Image-SFT | Apache 2.0 |
| Base | OpenMOSS-Team/MOSS-Image-Base | openmoss/MOSS-Image-Base | Apache 2.0 |
Each repository is a Diffusers folder of about 27.2 GB in bf16:
| Folder | Component | Size |
|---|---|---|
| transformer/ | 5B diffusion transformer, 3 files | 10.2 GB |
| text_encoder/ | Qwen3.5-9B, 4 files | 16.8 GB |
| vae/ | FLUX.2 small decoder | 0.25 GB |
Download a full folder with the Hugging Face CLI:
pip install -U huggingface_hub
hf download OpenMOSS-Team/MOSS-Image-SFT --local-dir MOSS-Image-SFTRun it locally
OpenMOSS asks for Python 3.10 or later, an NVIDIA GPU with CUDA, and PyTorch 2.9 or later, plus:
pip install "diffusers>=0.38" "transformers>=5.3" "accelerate>=1.10" "safetensors>=0.6" "Pillow>=10"The pipeline loads with custom_pipeline="pipeline_moss_image" and trust_remote_code=True. The transformer and text encoder run in bf16 and the VAE in fp32. The official Gradio app has three memory modes: GPU resident, Component (moves parts between CPU and GPU) and Sequential (lowest memory, slowest). OpenMOSS does not publish a VRAM figure.
| Setting | Recommended |
|---|---|
| Resolution | 1M, 1.5M or 2M pixels; 512×512 up to 2K works |
| Size step | Width and height in multiples of 16 |
| Steps | 50 or more, Euler sampler |
| CFG | 3.0 (raise toward 6.0 for short prompts) |
| Shift | 5.0 |
| Layer skipping | Negative branch, layer 22 |
Full code and prompt tips are in the GitHub repo and the user guide. MOSS-Image also has day-one support in SGLang and vLLM-Omni.
ComfyUI files
The official ComfyUI plugin is in ComfyUI Manager under MOSS-Image, and the ComfyUI guide includes an example workflow. It can load the Diffusers folder above directly. The ComfyUI team also repackaged single files at Comfy-Org/MOSS-Image:
| File | Put it in | Size |
|---|---|---|
| moss_image_sft_bf16 / moss_image_base_bf16 | models/diffusion_models | 10.2 GB |
| moss_image_sft_int8_convrot / base_int8_convrot | models/diffusion_models | 5.4 GB |
| qwen3.5_9b_bf16 | models/text_encoders | 19.3 GB |
| qwen3.5_9b_int8_convrot | models/text_encoders | 9.8 GB |
| qwen3.5_9b_w4a8_gptq | models/text_encoders | 7.3 GB |
| flux2-vae_small | models/vae | 0.25 GB |
INT8 and GGUF
Community builds shrink the transformer and text encoder for smaller GPUs. They are not official OpenMOSS releases.
| Build | Maintainer | Sizes | Notes |
|---|---|---|---|
| INT8 ConvRot | Comfy-Org | Transformer 5.4 GB, text encoder 9.8 GB | ComfyUI repackage of Base and SFT. |
| INT8 / INT4 ConvRot | SumomoLee | Transformer 5.4 / 3.0 GB | Base only, for ComfyUI. |
| GGUF Q2_K to Q8_0 | Rebel AI | 2.3 to 5.7 GB per transformer | Base and SFT. Needs ComfyUI-MOSS-GGUF plus city96 ComfyUI-GGUF. The maintainer has tested Base Q4_K_S-v2. |
A GGUF file holds only the transformer. You still need the text encoder and VAE, and a smaller file does not guarantee that it fits in 8 GB of VRAM. GGUF files do not run in llama.cpp or LM Studio.
LoRA
Training. OpenMOSS released Base without post-training so the community can train LoRAs on it. Its user guide gives one rule: keep the Qwen text encoder frozen. Ostris AI Toolkit supports LoRA training on both Base and SFT, including edit LoRAs that teach the model to change a reference image. Ostris also publishes ComfyUI nodes to run those edit LoRAs.
Speed LoRAs. Community adapters cut the 50-step, CFG 3 default down to a few steps without CFG:
| Adapter | Steps | Maintainer's numbers |
|---|---|---|
| lodestones · SFT 8-steps DMD | 8, no CFG | Work in progress. LoRA at strength 1.0 (0.7 for the latest file). |
| darask0 · SFT 4-step adapter | 4, no CFG | Transformer time per 1024² image on an H100: 16.2 s → 0.81 s. |
Few-step adapters need their own sampler code, so follow each model card. MossImage does not run custom LoRAs yet.
Questions
Is the MOSS Image download free?
Yes. OpenMOSS publishes the Base and SFT weights and the inference code under the Apache License 2.0, which allows commercial use under its terms. Community files carry their own licenses, so check each model card.
How much VRAM does MOSS Image need?
OpenMOSS does not publish a VRAM figure. The bf16 weights alone are about 27 GB (a 10 GB transformer and a 17 GB text encoder), so smaller GPUs rely on CPU offload or quantized files. Community maintainers note that file size is not peak VRAM.
Should I download Base or SFT?
For generating images, start with SFT: it is fine-tuned on curated data and groups text more clearly. Choose Base to train LoRAs or other fine-tunes, since it keeps the pretraining behavior intact.
Do the GGUF files work in llama.cpp or LM Studio?
No. They are image-model weights. They need a ComfyUI loader that implements the MOSS-Image transformer, such as the community ComfyUI-MOSS-GGUF bridge with city96 ComfyUI-GGUF.
Can I try MOSS Image without downloading it?
Yes. Generate with MOSS-Image in the MossImage studio, from 4 credits per image. We run MOSS-Image-SFT through fal. The free official demo on Hugging Face is also embedded on our home page, with a queue.
Don't have the GPU for it?
Skip the download and generate with MOSS-Image right here, from 4 credits at 0.5K, 10 at 1K and 36 at 2K. Prefer a free queue? Try the official demo. Every image shows the model that made it.