MOSS-Image-SFT: a 5.1B-parameter text-to-image model released on HuggingFace

On October 11, 2026, the OpenMOSS-Team published the MOSS-Image-SFT model on HuggingFace for text-to-image generation. It is a supervised fine-tuned version from the MOSS-Image family that supports natural language prompts, tag sequences, and JSON markup for controlling the placement of text and objects.

What it means

The model has 5.1 billion parameters, and the weights are published in safetensors format. The largest file is 4.6 GB, and the entire repository is 25.3 GB. The license is apache-2.0, and it runs via diffusers. The author recommends Python 3.10 or newer, an NVIDIA GPU with CUDA support, and PyTorch 2.9 or later. The transformer and text encoder run in bfloat16, while the VAE and initial noise run in float32. A native plugin for ComfyUI is available through ComfyUI Manager. The developer does not specify VRAM requirements, and there are no quantized builds in the repository. For those who follow local generation tools, the arrival of a model with open weights and diffusers support is a reason to check whether it fits their current hardware. We collect fresh news on this topic in the CPU3D news feed.