Ming-Image-0.1-Design vs FLUX 3: Two Ways for an Image Model Not to Be Available

Both Ming-Image-0.1-Design and the image tier of FLUX 3 are, right now, models you cannot call. They got there by opposite routes. Ming-Image-0.1-Design is a finished artefact with no front door: 6B of MIT-licensed weights published to Hugging Face by inclusionAI on 17 September 2026, documented in a model card whose Quick Start tells you to clone a repository that returns 404. FLUX 3 is a front door with nothing behind it: Black Forest Labs announced the family on 23 July 2026 as four tiers — Video, Image, Action and Dev — shipped the video tier, and has not published a FLUX 3 image endpoint, model identifier, request schema or price since. This is a comparison between a model that exists and cannot be run and a model that is promised and cannot be priced, which is a more useful thing to know than either headline on its own.

What each side has actually published

The distinction between a repository and an announcement matters, so here it is side by side.

• Artefact — Ming-Image-0.1-Design has real weights: five transformer shards, a VAE, a scheduler, and separate mllm, connector and mlp components, all under an MIT licence vs FLUX 3, where the only shipping component is the video tier, generally available since early August 2026.

• Identifier — inclusionAI/Ming-Image-0.1-Design, callable as a Hugging Face repo ID vs no FLUX 3 image model ID at all. Black Forest Labs' endpoint documentation lists /flux-2-max, /flux-2-pro, /flux-2-flex and the /flux-2-klein-* routes; a /flux-3 image route does not appear.

• Recommended configuration — 2048 x 2048 at 12 sampling steps, CFG scale 1.0, BF16, one 80 GiB CUDA GPU vs nothing published for the image tier.

• Price — none. The weights are free to download under MIT, but there is no hosted endpoint to buy vs a published rate card that covers video only, at roughly $0.06 per second at 720p draft, $0.17 at HD and $0.29 at FHD.

• Code — a Quick Start that does not run, because the companion inclusionAI/Ming-Image repository is absent from the organisation's public repository list vs an early-access application form, opened video-first, with no image-tier documentation behind it.

• Independent score — none. Ming-Image-0.1-Design appears on no Artificial Analysis board, in no category view, and on no open-weights view as of 23 September 2026 vs also none for any FLUX 3 tier. The nearest BFL entries on the live board are from the previous generation: FLUX.2 [flex] at 1,065 Elo in the UI/UX Design category and 1,026 on the general text-to-image board, with FLUX.2 [max] at 1,053 and FLUX.2 [dev] at 1,000.

Why the video tier shipped and the image tier did not

Black Forest Labs has given a reason, and it is worth stating precisely because it shapes what to expect. The company has said that more than 95% of FLUX 3's training compute went to video prediction. That is a coherent explanation for why the video tier is the one with a product behind it, and it is also the reason to be sceptical of the assumption that the image tier will arrive as a small delta when it does. The capability was trained second, and the roadmap promises attached to it are not small: improved complex-prompt handling, high-accuracy multilingual text rendering, style transfer, character and brand consistency without fine-tuning, and non-destructive editing. Each of those is a claim rather than a feature until there is an endpoint to test.

What BFL does sell for still images today is the previous line: FLUX.2 in its max, pro, flex and klein variants, plus the older FLUX Pro, Dev and Schnell releases, at roughly a cent and a half to seven cents per image depending on tier and resolution. Those are real, callable products and they are the honest benchmark for where BFL's image quality sits while the FLUX 3 image tier waits.

The deployment gap is the interesting asymmetry

Here is where the two models stop being mirror images. Ming-Image-0.1-Design has no working code but it does have a serving story: the model card recommends vLLM-Omni, and on 22 September the repository gained links to that framework's installation guide and to a model-specific recipe. The installation guide resolves. The recipe does not — it 404s as of 23 September — so what you have is a framework that has acknowledged the model and a model-specific page that has not been written. That is still further along than an application form.

FLUX 3's image tier has the opposite profile. Black Forest Labs is a company with a working API, a documented endpoint list, and a video tier in production on it. The infrastructure for a FLUX 3 image endpoint plainly exists. What does not exist is the decision to expose it. There is no gated model ID to apply for, no waitlist confirmation date, and no published capability list that survives contact with a request schema.

The practical consequence: a team that wants to serve Ming-Image-0.1-Design can start today by reading the vLLM-Omni quickstart and wiring the model up themselves, on their own hardware, at their own cost, with a permissive licence that permits it. A team that wants to serve FLUX 3 Image cannot start at any price, because there is nothing to serve.

Reading the one performance claim in the set

Ming-Image-0.1-Design ships a leaderboard graphic inside its repository: a UI/UX Design board headed "Text to Image Leaderboard: UI/UX Design", footed with "Elo scores from blind preference votes in our Image Arena", carrying an "Open Weights Leaderboard" badge. In it the model is first at 1,082 Elo, above Ideogram 4.0 (Quality) at 1,052 and the FLUX.2 dev variants between 994 and 1,000.

Two problems with reading that as a result. The first is provenance: it is vendor-published artwork in the vendor's own repository, and we are describing the file rather than verifying the study. The second is that we could not find the model on the live Artificial Analysis boards at all on 23 September, and the live UI/UX Design category is a different scale — its leader is GPT Image 2.5 Flare at 1,227 Elo, and the FLUX.2 entries the card lists read 1,065 for [flex], 1,053 for [max] and 1,000 for [dev] on that same live board. A ranking expressed on a board we cannot locate, roughly 150 Elo below the board we can, is not a benchmark result. It is a claim awaiting an outside run.

On the FLUX 3 side there is nothing to check, which is the point. No resolution target, no step count, no reference-image cap, no parameter count for the image tier. The roadmap is a list of adjectives.

What to do while both are unavailable

Neither side of this comparison is routable through OrcaRouter today, and it is worth being blunt about that rather than letting a model page imply otherwise: our catalogue carries no inclusionAI model and no FLUX tier, so the figures above are the vendors' own and nothing here is a route you can call through us. What the routing layer is genuinely worth in a gap like this is the day after the gap closes. Because OrcaRouter passes provider list prices through at 0% markup, a vendor rate revision is live on our side the same day rather than after a repricing cycle, and because a generation call sits behind one endpoint across 200-plus models, swapping in whichever of these two ships first is a model-ID change rather than a migration. Automatic failover matters for the same reason it always does with an unproven model: if a first-release image endpoint wobbles under load, the request moves instead of failing.

The decision today is narrower than the framing suggests. If you need image generation in production this quarter, neither candidate qualifies and the honest answer is a FLUX.2 tier — real, priced, callable, and independently scored — with FLUX 3 Image tracked as a future swap. If you build evaluation harnesses, Ming-Image-0.1-Design is the more interesting object in the pair: the weights are downloadable now, the licence permits whatever you want to do with them, and the first credible Elo anyone publishes for it will be a genuine piece of news. And if you are watching BFL, the signal to watch is not a blog post. It is a /flux-3 route appearing next to the /flux-2-* ones in the endpoint documentation, because that is the moment the image tier stops being a roadmap.