Ming-Image-0.1-Design Tops the Open-Weights UI Design Board — and the Number Checks Out

The claim circulating with Ming-Image-0.1-Design is that inclusionAI's 6B model is the best open-weights text-to-image model for UI and UX work, sitting first on the open-weights view of the Artificial Analysis UI/UX Design leaderboard at 1,082 Elo. Vendor leaderboard screenshots are usually the weakest kind of evidence, so the first thing worth doing is reading the live board. As of 24 September 2026 it holds up, with one important qualification that the screenshot does not carry: 1,082 is a use-case slice, not the board, and on the full Text to Image leaderboard the same model sits at rank 45 with an Elo of 995.

Both numbers are real, they come from the same arena, and they describe the same weights. What separates them is which prompts the votes were cast on.

What the live board actually says

Artificial Analysis runs one blind pairwise arena and then filters the same vote pool into use-case slices. The UI/UX Design slice is the one that matters for a model built for dashboards, app screens, posters and infographics, and it is where Ming-Image-0.1-Design does its best work:

• Overall Text to Image board — Ming-Image-0.1-Design at #45, Elo 995, 95% CI ±8, 21,342 samples, listed September 2026, $30.00 per 1,000 images, flagged Open Weights

• UI/UX Design slice — Ming-Image-0.1-Design at #16, Elo 1,084, 2,100 samples in the slice

• Highest-placed open-weights entry in that slice — Ming-Image-0.1-Design; the next open-weights models behind it are Ideogram 4.0 (Quality) at #22 and 1,047, Ideogram 4.0 at #31 and 1,018, and HunyuanImage 3.0 Instruct at #40 and 1,005

• Top of the UI/UX slice — GPT Image 2.5 Flare (max) at 1,226, GPT Image 2.5 Sunburst (max) at 1,215, GPT Image 2 (high) at 1,204, Grok Imagine Image 2.0 at 1,183

• Against the vendor's own screenshot — inclusionAI published 1,082 for Ming against Ideogram 4.0 (Quality) at 1,052 and FLUX.2 [dev] at 1,000; the live slice now reads 1,084, 1,047 and 1,000 respectively

So the headline is accurate on its own terms. Ming-Image-0.1-Design is the highest-rated open-weights model in the UI/UX Design slice, and it is the only open-weights model in that slice's top twenty. It is also 142 Elo behind the leader of that slice, and it is in the middle of the pack when the prompt is not a design prompt. A model that wins on dashboards and lands at 995 overall is a specialist, not an all-rounder, and the two figures are only contradictory if you read one of them as the whole story.

The 21,342 samples on the full board is also worth noting for what it is not. That is a large vote count, which is why the confidence interval is tight at ±8 Elo, but sample count is not the same as independence of method. The arena is blind human preference, so nobody at inclusionAI wrote the score — that part is genuine. What the score measures is "which of these two images did a voter prefer," and voters asked to judge a UI screenshot tend to reward legible text and coherent layout. That is precisely the axis this model was trained on.

What Ming-Image-0.1-Design is

Ming-Image-0.1-Design is a text-to-image model from inclusionAI, the AI lab associated with Ant Group, released under the MIT licence with weights published on 17 September 2026 and the full release package landing 22 September. It generates from a prompt alone — no reference image required — and it is aimed at text-heavy graphic design rather than photography: UI mockups, dashboards, infographics, posters, and anything where rendered text has to stay legible at the size it will be read.

The documented inference configuration is specific and unusually cheap per step:

• Resolution — 2048 x 2048 recommended, or 1024 x 1024 for faster generation, with intermediate sizes snapped to one of those two buckets

• Sampling steps — 12

• Guidance — CFG scale 1.0

• Precision — BF16

• Hardware — one CUDA GPU with 80 GiB of VRAM, described as the validated configuration

Twelve steps at a guidance scale of 1.0 is the signature of a distilled or guidance-free sampler. That is what makes a 2048-pixel output affordable at a step count most diffusion models would not attempt, and it is the main reason this model is interesting to anyone running image generation at volume rather than one image at a time.

The other structural feature is transparency. Ming-Image-0.1-Design can emit native RGBA, so a transparent background comes out of the sampler rather than out of a matting pass, when the prompt begins with one of the documented transparency phrases. A companion model, Ming-Image-0.1-Design-Layer, goes further: it takes a flattened design plus a layer plan and returns separate RGBA PNGs — text, cards, main subject, background — which recompose into the original. That is the difference between a design model and an image model. A flat PNG is a picture of a layout; separate layers are a layout you can still edit.

The 6B label and the 26B download

"6B" is accurate about the part most people mean and misleading about the part that decides whether you can run it. The diffusion transformer is 6.15B parameters. The package you actually download is roughly 52.88 GB, because the multimodal text encoder is 17.01B parameters and the connector adds 3.09B — around 26B parameters at BF16 in total. The Layer model's transformer is double the base model's at 12.31B, and its package is larger again at roughly 65.20 GB.

This matters for two practical reasons. First, the 80 GiB VRAM requirement is not about the 6B transformer; it is about everything that has to be resident alongside it. Second, any comparison against a model described as "6B" needs to check which 6B is meant. A 6B diffusion transformer with a 20B text encoder is a very different deployment problem from a 6B model end to end.

Prompt handling is the other thing the card makes explicit rather than hiding: the recommended recipe runs a rewriting step first, using a vision-language model to expand a short prompt into a structured, Figma-style JSON layout description with coordinates, hierarchy, colour specifications and verbatim text. The model cards name Ling-3.0-flash-VL or Qwen3.8-27B for that job. In other words, the pipeline the vendor benchmarks is a two-model pipeline. If you call Ming-Image-0.1-Design with a one-line prompt and no rewriting step, you are not running the configuration that produced 1,084 in the UI/UX slice.

Where this leaves a design pipeline

The honest summary of Ming-Image-0.1-Design is that it solves a specific and expensive problem — generating text-dense layouts that survive being looked at — and it does so at the top of the open-weights field on the one board that measures that problem. It does not lead the overall image board, it does not remove the need for a VLM in front of it, and it asks for a serious GPU.

What it changes is the middle of a design workflow. If your pipeline currently generates a flat image and then pays a human or a segmentation model to cut it apart, a model that emits RGBA natively and a companion that emits 2–9 named layers removes a stage. That is worth more than an Elo column, and it is the part no leaderboard measures.

For teams that want to test it without committing infrastructure, the split is worth being precise about. Ming-Image-0.1-Design is an MIT-licensed download, so it runs on your hardware or not at all — OrcaRouter routes no inclusionAI model of any version, and saying otherwise would be a claim we cannot honour. What a routing layer does give you is the surrounding surface: one OpenAI-compatible endpoint across 200+ models, automatic failover across providers, and provider list price passed through at 0% markup, so a vendor price change on any model you do call is live on our side the same day. If you are standing up a design pipeline and want to A/B this model against a hosted one you already trust, that comparison runs on one key rather than two contracts.

What would change the picture

Three things would move Ming-Image-0.1-Design from a strong open-weights specialist to a default. A first independent reproduction of the Layer model's Crello figures — the card reports an RGB L1 error of 0.0574 and alpha soft IoU of 0.8923 at 1024, and no outside group has run them. A hosted endpoint that removes the 80 GiB requirement. And a second use-case slice where the model places as well as it does in UI/UX Design, because a model that only wins on one slice of one board is a model whose generality is still unproven.

Until then, the accurate sentence is the narrow one: on the Artificial Analysis UI/UX Design slice, read 24 September 2026, inclusionAI's Ming-Image-0.1-Design is the highest-rated open-weights text-to-image model, at 1,084 Elo on 2,100 votes — and on the same arena's full board it is 45th at 995. Both of those are the model. Neither of them is a launch claim, because this model did not have a launch; it had a repository.