The repository describes it as a full fine-tune of MOSS-Image-Base for 1024 px text-to-image generation, using the same architecture, Qwen3.5-9B text encoder, and FLUX.2 small decoder as MOSS-Image-SFT.
The published checkpoint is the EMA weight set from update 11,000 of a run trained on 704,000 images. The model card provides Diffusers and ComfyUI instructions. It lists a bf16 ComfyUI file and an int8 ConvRot version of 5.4 GB.
The release also includes two 8-step turbo LoRAs. The recommended DMD2 LoRA is listed for 8 inference steps at CFG 1, while the full model is listed for 50 steps at CFG 3. In the publisher's 100-prompt PartiPrompts table, Kromoss scores 27.90 with the full model and 28.13 with the DMD2 LoRA, using CLIP ViT-L/14.
The source does not provide an independent evaluation, comparative human-preference results, or a general hardware requirement for running the model.