23 September 2026
Ant Group’s inclusionAI lab published Ming-Image-0.1-Design and Ming-Image-0.1-Design-Layer on Hugging Face on 22 September 2026. Both are 6 billion parameter image models released under the MIT licence, which allows commercial use without conditions. The pair is aimed at text-rich design work rather than photographs.
What does Ming-Image-0.1-Design do?
It is a text-to-image model built for interface screens, infographics, posters and other designs where the words have to be readable. Most open image models still garble small type. Ming-Image-0.1-Design generates the whole composition, including the text, and can output RGBA images, which means the background comes out transparent instead of white.
The model card recommends 2048 by 2048 output, or 1024 by 1024 when you want it faster, with 12 sampling steps and a CFG scale of 1.0. Twelve steps is low for a diffusion model, so generation is quick once the weights are loaded. It runs through the standard diffusers library and also supports vLLM-Omni for serving.
What you need to run it:
- Licence: MIT, so commercial use is allowed with no extra terms
- Hardware: one CUDA GPU with 80 GiB of VRAM is the only setup Ant Group validated
- Precision: BF16
- Download: the inclusionAI organisation on Hugging Face
- Code: the Ming-Image repository on GitHub
- Community INT4, INT8, FP8, GGUF and ComfyUI builds appeared within hours of release
What is Ming-Image-0.1-Design-Layer?
The second model is the more unusual one. Ming-Image-0.1-Design-Layer takes a finished, flattened design and splits it back into a set of transparent layers, using a layer plan you supply to say how many you want. You get something closer to a working design file than a single flat PNG.
That matters because a flat image is the reason most image models stop short of real design work. If a client wants the headline moved and the logo swapped, a PNG forces a regenerate and you lose everything else on the canvas. Layers let you edit the one thing. Ant Group evaluates this on the Crello test set, measuring how close each recovered layer is in colour and in shape.
What Ant Group is not saying about the benchmarks
There are no numbers you can quote. The model card points at the Artificial Analysis UI/UX Design leaderboard and at Crello results, but publishes both only as chart images, with no figures in the text and no named rivals. You cannot tell from the release how Ming-Image-0.1-Design compares to Qwen-Image, Seedream or any commercial image model, and neither claim can be checked without rerunning the evaluation yourself.
The hardware guidance has a gap too. An 80 GiB GPU is the only validated configuration for a 6 billion parameter model, which is far more headroom than the size suggests. Ant Group gives no guidance for 24 GB consumer cards, even though community quantisations for exactly those cards were up within hours.
What this means
Worth testing now if you produce design assets at volume and you need the text to be legible. A three-person marketing team turning out social cards, ad variants and internal decks is the clearest fit, and the MIT licence means nothing in the terms stops you using the output commercially. Ming-Image-0.1-Design-Layer is the part to try first, because layered output is the thing your existing image tool almost certainly cannot do.
Treat the quality claims as unverified until you run your own comparison. Ant Group has published a leaderboard position as a picture and nothing else, and that is not evidence. What we would test first is text rendering at small sizes in a language other than English, since that is where models in this class usually fail and the release says nothing about it. If your output is photographic rather than typographic, this is safe to ignore.
For more information, visit the official release in the Ming-Image-0.1-Design model card.
