Ming-Image-0.1-Design vs Meta Muse Image: The Model With a Paper Trail and the Model With a Press Release

Two image models arrived in 2026 in opposite ways. Ming-Image-0.1-Design is a 6.15-billion-parameter text-to-image model from Ant Group's inclusionAI team, uploaded to Hugging Face on 17 September 2026 under an MIT licence with no announcement, no blog post and no benchmark table — a 71.5 GB repository you can read file by file. Meta Muse Image is Meta Superintelligence Labs' first in-house image model, launched on 7 July 2026 with a press cycle, vendor-reported benchmark claims and no downloadable weights of any kind. The interesting question is not which one draws a better poster. It is that one of these models can be audited today and the other can only be taken on faith, and the auditable one is sitting first on the only independent board it appears on.

One shipped as files. One shipped as a claim

Start with what each release actually consists of, because the difference is not subtle.

• Ming-Image-0.1-Design — 6,154,901,056 parameters, MIT licence, a diffusers pipeline, all weights in BF16 safetensors, roughly 71.5 GB across six component folders: connector/ (2 shards), mllm/ (7 shards), mlp/ (124.8 MB), scheduler/, transformer/ (5 shards) and vae/ (253.8 MB). No model_index.json, which is why the recommended path is a git clone of the inference repo and python infer.py rather than a one-line pipeline call.

• Meta Muse Image — no weights, no published parameter count, no architecture document, no licence file to read. It is reachable as a service and, since August 2026, through Meta's own Model API on an OpenAI-compatible endpoint at a flat $0.01 per image. Everything you know about its internals comes from Meta's own description of them.

That asymmetry has a practical consequence that has nothing to do with image quality. With the Ming release you can answer questions the vendor never addressed: what the text encoder is, how many sampling steps the scheduler expects, whether the VAE emits four channels or eight, whether the MLP connector is doing anything you could replace. With Muse Image you cannot answer any of them, and you will not be able to next quarter either, because there is no artifact to open.

The one number worth quoting, and where it comes from

Ming-Image-0.1-Design is ranked first on the Artificial Analysis Text to Image Leaderboard for UI/UX Design, in the open-weights view, at an Elo of 1,082. The field behind it reads: Ideogram 4.0 (Quality) at 1,052, Ideogram 4.0 at 1,015, HunyuanImage 3.0 Instruct at 1,005, FLUX.2 [dev] at 1,000, FLUX.2 [dev] Flash at 999 and FLUX.2 [dev] Turbo at 994.

Three honesty notes on that, because it is the strongest evidence in this comparison and it deserves to be handled carefully. First, the copy of that leaderboard we are reading is the one published on the model's own card, and it carries Artificial Analysis branding — it is an Artificial Analysis board reproduced by the vendor, not a figure we generated. Second, nobody has reproduced the score independently, and a blind-preference Elo board is a population statistic, not a per-prompt test: it tells you which image a crowd preferred on average, not whether the model will render your specific infographic correctly. Third, and this is the part that explains a lot, the board is filtered to open weights. Meta Muse Image cannot appear on it, so its absence is not a result.

For Muse Image the numbers are Meta's own and Artificial Analysis's, and they should be labelled as such. Meta reported a 94% multi-image character consistency figure, support for up to 10 reference images, and output up to 1600 pixels on the long edge. Artificial Analysis places the model inside the top five for text-to-image and third on its image editing board at an Elo of roughly 1,105. Those are vendor-reported and independently-measured figures respectively, and they are not the same kind of thing.

What Ming-Image-0.1-Design is built to do

The model card is specific about the target, and it is not general-purpose photography. It is aimed at UI, infographics, posters and text-heavy design work — the class of image where the hard part is legible type and clean layout rather than photoreal skin. The recommended settings reflect that: 2048×2048 output, or 1024 when you want speed, 12 sampling steps, classifier-free guidance at 1.0, BF16 precision, and a single 80 GiB CUDA GPU. Prompt enhancement is expected to come from a separate vision-language model — the card names Ling-3.0-flash-VL and qwen3.8-27B — and deployment is expected to go through vLLM-Omni.

There is also a companion, Ming-Image-0.1-Design-Layer, which is a separate repository with a different pipeline tag (image-text-to-image), 6,154,908,736 parameters and the same MIT licence. Its job is layer decomposition: you hand it a flattened design image plus a layer plan and it returns RGBA layers. It runs at 1024 in its default bucket, or 512 for speed, at 12 steps with guidance at 2.0 and flash_attention_2. The card shows results on a Crello test set as an image rather than as a table, which is worth noting if you were hoping to compare it numerically to anything.

What is genuinely not knowable yet

This is where the framing has to be blunt. Ming-Image-0.1-Design has no release note, no announcement, no stated limitations section and no third-party coverage at the time of writing. The download counter on the repository reads zero. There is no hosted endpoint from the vendor. The only independent measurement is the one leaderboard position, and everything else — throughput, prompt adherence on your data, how it behaves at reference-image counts above a handful, whether the RGBA channel is clean at the edges — is unmeasured in public.

What that buys you is a short, well-defined evaluation. You can read the architecture, run it, and decide for yourself inside a day. It does not buy you a benchmark you can quote in a procurement document.

Meta Muse Image is the mirror image. It has been measured on two independent boards and it has a published price, and neither of those things will let you inspect a single weight. If your constraint is a compliance review that asks where the training data came from and what the model does with a prompt, "Meta says" is the whole of the answer available to you.

The version of this comparison a router can act on

Neither of these two models is one we route, and this article is not going to imply otherwise. Meta Muse Image is closed and served by Meta; Ming-Image-0.1-Design is a weights download with no hosted route at all. The image models we do front today are OpenAI's GPT-Image family, Google's Imagen 4 tiers and Gemini image previews, and xAI's Grok Imagine image endpoint — all behind one OpenAI-compatible endpoint at provider list price with zero markup, with automatic failover across providers and a routing DSL that composes several models into a single call.

That matters here for one specific reason. An unannounced open-weights release with a single leaderboard position is exactly the kind of model a team wants to evaluate without betting a production path on it. Running it beside a measured model on the same key, rather than in front of it, is how you get the evaluation done without a second contract and without a code change.

Pick by what you can verify

• If you need a model you can inspect, modify, fine-tune or run on your own hardware — Ming-Image-0.1-Design is the only option of the two, and its MIT licence is unusually permissive for a 2026 image release.

• If you need a model with a published price and independent scores you can quote — Meta Muse Image is the only option of the two, and $0.01 per image is a number that survives a budget review.

• If you need transparency as an output — Ming-Image-0.1-Design's RGBA path and its Layer companion are aimed at exactly that; Meta's transparency story is not published.

• If you need it to work this afternoon — neither. One has no hosted endpoint, the other has no download.

The open question is narrower than it looks, and it is a date rather than a benchmark. Ming-Image-0.1-Design was uploaded on 17 September 2026 and nobody has announced it. Either that announcement comes and the leaderboard position gets a second, independent reading, or it does not and this stays a repository that a small number of people have run. The first third-party evaluation of it is the thing to watch, because that is the moment the paper trail stops being a promise and starts being evidence.