OpenMOSS introduces MOSS-Image, a 5B model with JSON layout control

The MOSS-Image project page has a news line dated 16 October 2026: introducing MOSS-Image-Base, with code and weights available. It calls the family an OpenMOSS text-to-image pair, a base model and a curated-data fine-tune, with a 5B backbone, prompts of up to 4K tokens, and JSON that places objects, rendered text, and positions.

The Hub repos predate that line. API createdAt is 11 October 2026 at 06:49 UTC for MOSS-Image-Base and 07:01 UTC for MOSS-Image-SFT. Both were last modified on 16 October at 19:30 UTC. Neither is gated. The page’s public introduction is 16 October. The repos existed from 11 October. The API does not say the weight files were complete that day.

Architecture and license

The Hub lists 5,082.3 million parameters for Base and for SFT. The technical report calls the generator a 5B single-stream diffusion transformer, pretrained from scratch. transformer/config.json names MOSSImageTransformer: 42 layers, dimension 3,072, 24 heads, FFN dimension 8,192, text dimension 4,096.

The pipeline is MOSSImagePipeline. The text encoder class is Qwen3_5Model. The VAE class is AutoencoderKLFlux2. The card says the bundled VAE is the FLUX.2 small decoder, with the transformer and text encoder in bfloat16 and the VAE and initial noise in float32. Hub language tags are English and Chinese.

The page and the card credit the OpenMOSS Team. The report’s byline is MOSI.AI, with footnotes for MOSI Intelligence, SII, Fudan University, and Shanghai Jiao Tong University.

The Base LICENSE is Apache License 2.0. The card puts the weights and the bundled inference code under that license. The project page says the weights are Apache-2.0, and that pipeline code and third-party components keep their own licenses. SFT carries the same tag and is marked as a fine-tune of Base.

The layout JSON

Text and layout share one prompt field. A layout is JSON with summary.caption, items, bbox_format of yxyx, and bbox_coord_space of abs_centralized. Coordinates are [y_min, x_min, y_max, x_max], in pixels from the canvas center. The page’s example labels items text and object. It says to pass the JSON string as the prompt at the canvas size the boxes were written for. The card requires image sides in positive multiples of 16. The page’s diagram caption says the figure is the input layout, not a generated image.

The card’s run notes: Python 3.10 or later, an NVIDIA GPU with CUDA, PyTorch 2.9 or later, then diffusers 0.38 or later and transformers 5.3 or later. The project page says the same Python and GPU requirement, a PyTorch build that matches the CUDA install, and that training used heterogeneous GPU platforms. It names no chip. Sampling defaults on the card are flow-matching Euler, 50 steps, CFG 3, shift 5, negative-branch skipping at layer 22, and an empty negative prompt. Default guidance does two transformer forwards per step. Those are instructions, not a timing measured here.

Author scores, and the 16 October ports

The report’s abstract says MOSS-Image is state of the art among open-weight models on LongLongTextBench, with average text coverage of 96.3 percent. The introduction calls 96.3 a Mean-6 score and says it passes Qwen Image 2512 (64.9) by 31.4 percentage points. On RealLayoutArena it says the system ranks second, behind GPT Image 2. Those figures are the authors’. They were not rerun here.

Three community repos, all tagged apache-2.0 and none gated, were created on 16 October. Comfy-Org/MOSS-Image (12:42 UTC) says it repackages ComfyUI files, including BF16 and int8 weights and files named qwen3.5_9b. lodestones/MOSS-Image-SFT-8-steps-dmd (10:28 UTC) calls itself a work-in-progress distillation from 50 steps at CFG 3 to 8 evaluations with no CFG. realrebelai/MOSS-Image_GGUFs (15:02 UTC) titles itself GGUF and W4A8 quantizations of the SFT checkpoint. The project page, separately, says to search ComfyUI Manager for MOSS-Image, and that official ComfyUI integration is in progress.

Quote 16 October for the public page, and 11 October for the Hub creation. The documented control format is the JSON prompt.