MOSI Intelligence and the OpenMOSS team have open-sourced MOSS-Image, a 5-billion-parameter text-to-image system aimed at posters, product shots and other graphics where dense lettering and precise layout matter as much as raw picture quality. The release covers MOSS-Image-Base and a supervised fine-tuned variant, MOSS-Image-SFT, with weights on Hugging Face and ModelScope under an Apache 2.0 license. The technical report lists MOSI Intelligence together with Shanghai Innovation Institute (SII), Fudan University and Shanghai Jiao Tong University.
Within days of the October 16 launch, Chinese tech media reported that the model sat at the top of Hugging Face's text-to-image trending list and that community download and GitHub star figures moved quickly across Hugging Face mirrors and related hosts, with the official repository climbing past about 1.4 thousand stars. Those download and star counts are community figures rather than audited company metrics, and they shift as mirrors and distilled checkpoints multiply. Developers on Reddit and X have already posted side-by-side tests against other open weights, and community projects have shipped 8-step and 4-step distillations, quantization packs and ComfyUI workflows.
What the team is selling is control. MOSS-Image accepts prompts up to 4,096 tokens and can take a structured JSON layout that names text and objects with bounding boxes, rather than relying only on natural-language composition. On the team's Long Long Text Benchmark, it reports a Mean-6 text coverage of 96.3 percent among open-weight models. On RealLayoutArena, an end-to-end design blind test, its Elo point estimate is 1,698.70, second behind GPT Image 2 with short design briefs, and 1,807.34 with shared enhanced prompts, topping that second group in the team's own scoring.
Architecturally, MOSS-Image is a single-stream diffusion transformer pretrained from scratch. A frozen Qwen3.5-9B text encoder and a FLUX.2 variational autoencoder feed a unified generation path. The training recipe, according to the team, filters out AI-generated images to keep more photographic variety, and the model was trained across heterogeneous GPU platforms. Day-zero adapters are already listed for Diffusers, ComfyUI, SGLang and vLLM-Omni, and a hosted API is live on the MOSI platform.
The open weights give designers and agent builders a compact base for layout-heavy work without waiting on a closed API. Benchmark and ranking figures above come from the team's report and from community trackers; they should be read as self-reported or crowd-sourced signals, not independent lab results.
SEE ALSO: OpenMOSS and MOSI Release MOVA · Tencent's Open-Source Hunyuan Image 3.0