MOSS-Image: a 5B open image model for long text and JSON layouts

OpenMOSS released MOSS-Image under Apache 2.0: a 5B generator that renders long English and Chinese text and follows JSON layouts, with a low-memory profile near 4 GB of VRAM.

  • New release
  • 2 min read
  • Published

OpenMOSS, a research group under the Shanghai Innovation Institute, released MOSS-Image on 16 October 2026. It is a pair of text-to-image checkpoints, Base and SFT, published with code, weights and a technical report under Apache 2.0. Its pitch is pictures that carry a lot of correct text, in English and Chinese.

What is new

The generator is a 5.08-billion-parameter single-stream diffusion model with 42 blocks, trained from scratch. It reads prompts of up to 4,096 tokens through a frozen Qwen3.5-9B text encoder and draws in the latent space of the FLUX.2 VAE. Besides plain prompts it accepts a JSON layout. Each item has a label, a caption, optional text and a bounding box in pixels measured from the canvas centre, so a headline can be pinned to the top of a poster and an illustration to the bottom.

The report introduces two benchmarks and runs both itself. On LongLongTextBench, which tests long English and Chinese passages, MOSS-Image scores 96.3 against 64.9 for Qwen Image 2512. On layout, its mean IoU on detected objects is 0.88 against 0.44 for Ideogram 4. In RealLayoutArena, a blind human comparison of nine full systems, it ranks second behind GPT Image 2, with an agent-based prompt enhancer counted as part of the system. That enhancer ships as a skill for GPT-6-family agents, not as weights.

Running it

The repository is 27.2 GB, because the 16.8 GB Qwen3.5-9B text encoder comes with the 10.2 GB transformer, so the “5B” label covers only the generator. The default is 50 steps at a guidance scale of 3, and the user guide recommends 1 to 2 megapixels. ComfyUI supports it natively, SGLang and vLLM-Omni had day-0 support, and Diffusers needs custom pipeline code.

The README lists a low-memory profile that keeps the transformer in CPU memory. For one 1024×1024 image it peaked at 3.65 GiB of GPU memory and 14 GiB of system RAM, and we expect it to be slower than a full-GPU run. Third-party projects arrived within days: trainers that claim LoRA training from 6 GB of VRAM, an 8-step distillation and a C++ and CUDA engine. Reference-image input and editing are not in this release.

License

The weights and bundled inference code are plain Apache 2.0, and each repository carries a LICENSE file. There is no revenue cap, territory limit or naming rule beyond the standard notice.

Why it matters

Most of the open image models in our catalog carry license conditions that limit commercial use: a revenue cap on FLUX 3 Image Klein, research-only terms on Qwen-Image-2.1 and territory limits on MiniMax H3-Image. Z-Image is the permissive exception, aimed at speed and photorealism. MOSS-Image adds a second Apache 2.0 option, with long text and layout control as its focus.

Every score comes from OpenMOSS, and its report does not test Z-Image, so try your own prompts. Our MOSS-Image vs Z-Image page lays the two side by side, and the image and video list shows the whole category.

Sources

Models in this story

Model Released Parameters Context License Smallest variant needs
MOSS-Image OpenMOSS 5.08B – Apache 2.0 (Permissive) 4.2 GB Laptop or phone
Z-Image Tongyi MAI 6.15B – Apache 2.0 (Permissive) 16 GB Laptop or phone
Qwen-Image-2.1 Alibaba Qwen 7.1B – Qwen Research License (Non-commercial) 20 GB Gaming GPU
FLUX 3 Image Klein Black Forest Labs 5.6B – FLUX Kommunity (Conditional) 8 GB Laptop or phone
MiniMax H3-Image MiniMax 33.1B – MiniMax H3-Image Community (Conditional) 155 GB Big-memory desktop