What is Ming-Image-0.1-Design?
Ming-Image-0.1-Design was released on Hugging Face on 17 September 2026 by inclusionAI, Ant Group's AI unit, alongside Ming-Image-0.1-Design-Layer, which splits a flat design into editable transparent layers. Both are 6-billion-parameter models under the MIT licence. Design generates UI mock-ups, infographics, posters and other text-rich compositions, and outputs RGBA images with transparent backgrounds when the prompt starts with one of ten fixed phrases.
It samples in 12 steps at guidance 1.0 and works in 1024 or 2048 resolution buckets, with presets up to 2560x1440. The validated deployment is one GPU with at least 80GB of memory in BF16. Best results need a structured, Figma-style JSON prompt written by a second model, Ling-3.0-flash-VL or Qwen3.8-27B, using a system prompt inclusionAI publishes. inclusionAI also ships agent skills that turn its output into UI code or editable PowerPoint slides.
How much does Ming-Image-0.1-Design cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | MIT |
| DeepInfra | $0.01 per 1024x1024 image at 12 steps | Deprecation scheduled for 13 November 2026 |
Free weights under MIT. DeepInfra charges $0.01 per 1024x1024 image at 12 steps, scaling with area and steps, but plans to retire the model on 13 November 2026.
Is Ming-Image-0.1-Design worth it?
Ming-Image-0.1-Design is a narrow tool that is very good at its one job. Artificial Analysis said on 24 September 2026 that it was the top open-weights model for UI and UX design on its text-to-image board and 17th among all models in that category; inclusionAI's chart puts it at 1082 Elo there, against 1052 for Ideogram 4.0 (Quality) and 1000 for FLUX.2 [dev].
Outside design it is ordinary. On the overall text-to-image board it ranks 48th of 108 at 1000 Elo, level with FLUX.2 [dev] and 45 points behind the open leader Qwen-Image-2.1. Artificial Analysis noted that the 88-point gap between its UI and overall scores was the largest across its 10 use-case categories.
Deployment is the hurdle. inclusionAI validates only a single 80GB card in BF16, an A100 or H100 rather than a gaming GPU, and the model expects a long JSON layout description from a separate language model. DeepInfra's guide warns that a bare prompt such as 'a website' returns a stock-style photo of a monitor, and that long body paragraphs degrade into plausible fake glyphs, so real copy belongs in a design tool afterwards.
The hosted route is closing. DeepInfra charges $0.01 for a 1024x1024 image at 12 steps, about $0.04 at 2048x2048, but has scheduled the model's deprecation for 13 November 2026 because of low usage. That leaves self-hosting, which the MIT licence makes legally simple.
Where it earns its place
A product designer at a 40-person SaaS company
Turn a feature brief into realistic screen mock-ups for user testing
Have Qwen3.8-27B expand the brief with the published system prompt, then render at 2560x1440 on a rented 80GB GPU. Headlines and buttons hold up; paste long paragraphs in Figma afterwards.
A consultancy producing client slide decks
Recreate generated slide images as editable PowerPoint pages
Generate the slide, then run inclusionAI's Image-to-Editable-PPT skill, which converts text and simple shapes to native elements; the Layer model can split complex graphics into transparent layers.
A front-end team building an internal coding agent
Let the agent draft UI code from a prompt and check it visually
The Ling UI Design skill gives the agent Ming's images and layer decompositions as visual references; the loop needs the 80GB GPU plus a prompt-rewriting model running alongside.
Before you commit
- Setup
- Clone inclusionAI/Ming-Image, install requirements and run infer.py against the Hub ID on one 80GB GPU; add FlashAttention 2 for speed and a rewriter model for prompts.
- Integrations
- vLLM-Omni recipes, a ModelScope mirror, DeepInfra's API until its deprecation, and inclusionAI's Ling UI Design and Image-to-Editable-PPT agent skills.
- Data and privacy
- MIT weights run offline, so briefs and unreleased designs stay in-house when self-hosted.
- Cost at scale
- DeepInfra bills $0.01 × (width/1024) × (height/1024) × (steps/12), so 10,000 images at 2048x2048 cost about $400 while it lasts; self-hosting means an 80GB data-centre GPU.
- Limits
- Text-to-image runs in 1024 or 2048 buckets with aspect ratios from 1:4 to 4:1; long paragraphs degrade; vague prompts give generic results; it is a 0.1 release.
What’s good
- Top open-weights model on Artificial Analysis's UI and UX board
- MIT licence
- Transparent RGBA output
- Companion layer model and agent skills for UI code and slides
What’s not
- Validated only on a single 80GB GPU
- Needs a second model to write structured prompts
- Long body text degrades into fake glyphs
- Its only hosted API is being retired on 13 November 2026
Best for
Product and design teams with data-centre GPUs who want an MIT-licensed model for app mock-ups, slides and infographics.
Ming-Image-0.1-Design alternatives
Ming-Image-0.1-Design: common questions
- Is Ming-Image-0.1-Design worth it?
- Ming-Image-0.1-Design draws app screens and slides better than any other open model on Artificial Analysis's UI board, but it wants an 80GB GPU and a second model to write its prompts. We rate it 3.8 out of 5.
- How much does Ming-Image-0.1-Design cost?
- Free weights under MIT. DeepInfra charges $0.01 per 1024x1024 image at 12 steps, scaling with area and steps, but plans to retire the model on 13 November 2026.
- What are the downsides of Ming-Image-0.1-Design?
- Validated only on a single 80GB GPU; needs a second model to write structured prompts; long body text degrades into fake glyphs; its only hosted API is being retired on 13 November 2026.
- Who is Ming-Image-0.1-Design best for?
- Product and design teams with data-centre GPUs who want an MIT-licensed model for app mock-ups, slides and infographics.
- How much does Ming-Image-0.1-Design cost for a team or in production?
- DeepInfra bills $0.01 × (width/1024) × (height/1024) × (steps/12), so 10,000 images at 2048x2048 cost about $400 while it lasts; self-hosting means an 80GB data-centre GPU.
- Is Ming-Image-0.1-Design safe for company data?
- MIT weights run offline, so briefs and unreleased designs stay in-house when self-hosted.
- What are the best Ming-Image-0.1-Design alternatives?
- The closest alternatives we have reviewed are MOSS-Image, Ideogram 4.0, Qwen-Image-2.1 and GLM-Image.
Answers checked Nov 7, 2026.