Ant Group's Bailing (Ming) model team open-sourced the Ming-Image-0.1-Design series on September 23 — two models, both with 6 billion parameters, with code and weights hosted on Hugging Face under the inclusionAI organization and released under the MIT license.
- Ming-Image-0.1-Design handles "text to design": it takes a written brief and generates full, layout- and text-heavy designs such as UI screens, dashboards, infographics and posters.
- Ming-Image-0.1-Design-Layer handles "design to layers": it splits a flattened design image into 2 to 9 separate, editable transparent layers.
The same release also includes two agent skill packs: the Ling UI Design Skill, and the Image-to-Editable-PPT Skill, which converts an image into an editable PowerPoint file. The model API is available for free on OpenRouter for two weeks.
What kind of images it targets
General-purpose text-to-image models are good at pictures, but tend to stumble on content that needs precise text and tight alignment — buttons, cards, multi-column information blocks — where misspelled labels, misaligned elements and inconsistent colors within the same image are common. Ming-Image-0.1-Design is built specifically to address that.
According to the official documentation, the Design model supports structured prompts of up to 8K tokens, so a long brief that spells out titles, button copy, card content and color scheme can be fed in as a single block. Text rendering and layout stability for titles, buttons and multi-region information have been specifically reinforced, and the VAE natively supports RGBA, so elements with transparent backgrounds can be generated directly. The recommended output resolution is 2048×2048, with 1024×1024 available for faster generation.
The Layer model addresses a different problem. When a designer receives an AI-generated image and wants to change the color of one button, the only options used to be regenerating the whole image or manually cutting it out. Once the image is split into layers, text, background and individual elements can be moved and replaced separately.
Results and caveats
On the UI/UX design leaderboard that third-party benchmark firm Artificial Analysis updated on September 18, Ming-Image-0.1-Design scored an Elo rating of 1082, ranking first among open-weight models. Ant Group also published three sub-scores: layout stability at 67.4%, complex composition at 67.0%, and text rendering at 66.7%. The public materials do not spell out the methodology behind those percentages or which models they were benchmarked against, so the Elo ranking is the more reliable reference point.
The Hugging Face model page does not disclose base architecture details, nor does it offer a direct comparison with closed-source models. The leaderboard ranking is limited to open-weight models, and there is no public data on how it compares with closed-source commercial design-generation tools.
What it means for teams in China
Start with the hardware bar. The officially verified setup is a single CUDA GPU with 80GB of VRAM at BF16 precision — data-center-class hardware in the A100/H100 tier. The model page says nothing about whether consumer GPUs can run it or how much quality drops after quantization, so teams deploying locally will need to test that themselves. For anyone who just wants to see the results first, the two-week free OpenRouter API is the lowest-cost way in.
On licensing, MIT permits commercial use and derivative work, which means little added legal burden for companies that want to embed the generation capability into their own design tools or marketing-material systems.
The piece closest to everyday office work is the PPT skill pack. Plenty of report decks are still assembled by "screenshot and manually edit"; if an image can be split back into editable page elements, that saves the repeated rework. How faithful the result actually is — whether the extracted text can be edited directly, or whether fonts get swapped out — depends on testing that hasn't been officially benchmarked either.
Chinese-language typography is another item still to be verified. Ant's skill-pack materials include Chinese design examples, but the model page does not list separate test results for Chinese text rendering, so users producing Chinese-language posters or interfaces will need to run their own trials before drawing conclusions.
Sources: Hugging Face model page, IT Home, CocoLoop, Artificial Analysis; parameter counts, license terms and VRAM requirements follow the Hugging Face model page, while leaderboard rankings and sub-scores follow Artificial Analysis and the official release.