MOSS-Image: 1.3M downloads in four days for a 5B open image model
retained text
MOSI and the OpenMOSS team put out MOSS-Image, an open image model at 5B parameters. Pandaily covered it on Oct 25, citing the team and QbitAI: about 1.3M downloads in the first four days, around 1,400 stars on GitHub, top of the text-to-image trend on Hugging Face. Read the download number with care. It adds up across several repos, Comfy-Org on Hugging Face among them, so no single model page carries that figure on its own. All three numbers come from the team and Chinese media. Nobody independent has checked them. The community moved fast. There are 8-step and 4-step builds already, while the official setting is 50 steps. INT8, GGUF and MLX builds exist, plus style LoRAs. What people seem to want from it: dense text in Chinese and English, structured JSON layout with bounding boxes, prompts up to 4096 tokens. Under the hood it is a 5B diffusion transformer trained from scratch, with a frozen Qwen3.5-9B text encoder and the VAE from FLUX.2. Ostris guessed the more natural look may come from the team dropping AI-generated images from the training data. Treat that as his guess. Testers split. Some praise the realism and the detail. Others say some styles come out shaky. Question for the sub: who has run the 4-step builds, how do they hold against 50 steps, and how does MOSS-Image compare with FLUX.3-klein and Qwen Image 2.1 on text and on busy compositions?
document fetched 2026-10-26 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1xez2bk/mossimage_13m_downloads_in_four_days_for_a_5b