QbitAI reports that the 5B-parameter MOSS-Image image-generation model drew 1.3 million downloads across overseas communities within four days of its open-source release. The model uses a single-stream diffusion Transformer with structured JSON Layout controls, Chinese and English text rendering, and text conditions of up to 4,096 tokens, while community derivatives include accelerated, quantized, and LoRA variants. QbitAI also says MOSS-Image was trained on domestic PPU computing infrastructure after excluding AI-generated images from its training data.