Understand this story
Moss Intelligence and the OpenMOSS team open-sourced the 5B-parameter text-to-image model MOSS-Image on October 23, 2026. According to QbitAI, the model surpassed one million downloads within four days of release, gained 1.4k Stars on GitHub, and topped the Hugging Face text-to-image trending list. The model uses a single-stream diffusion Transformer architecture pretrained from scratch, combined with a frozen Qwen3.5-9B text encoder and FLUX.2 VAE, and supports text input of up to 4096 tokens and structured JSON Layout control. In evaluations, it achieved 96.3% Mean-6 text coverage on the Long Long Text Benchmark, and on the 7Bench layout control evaluation it achieved 0.88 target localization mIoU and 0.70 layout AUC. The community has released 8-step and 4-step acceleration schemes as well as INT8 quantized and GGUF versions.
AI-generated from public reports
Developments
1 confirmed developments
Moss Intelligence and OpenMOSS open-source release of MOSS-Image
Source:English text is being prepared.
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 23
QbitAI
Moss Intelligence and OpenMOSS open-source the 5B image generation model MOSS-Image, surpassing one million downloads in 4 days
Moss Intelligence and the OpenMOSS team have open-sourced MOSS-Image, a 5B-parameter text-to-image model, which surpassed one million downloads within 4 days, gained 1.4k Stars on GitHub, and topped the Hugging Face text-to-image trending list. The model adopts a single-stream diffusion Transformer architecture pretrained from scratch, combined with a frozen Qwen3.5-9B text encoder and FLUX.2 VAE, and supports text input of up to 4096 tokens and structured JSON Layout control. On the Long Long Text Benchmark, it achieved a Mean-6 text coverage rate of 96.3%; on the 7Bench layout control benchmark, it achieved 0.88 target localization mIoU and 0.70 layout AUC. The community has already released 8-step and 4-step acceleration schemes as well as INT8 quantized and GGUF versions.