Ming-Image-0.1-Design vs MAI Image 2.6 Preview: Open Weights Nobody Hosts, Against a Preview Microsoft Already Bills For

Two models, both described by their makers as not finished, and the word means opposite things on each side. MAI Image 2.6 Preview is unfinished in the sense that Microsoft has shipped it anyway: listed on the independent board as released 10 August 2026, tracked at $38.90 per 1,000 images, sitting fifth in the UI/UX Design category at 1,168.8 Elo on 818 blind comparisons — a preview with a price and a score. Ming-Image-0.1-Design is unfinished in the sense that inclusionAI has published the artefact and nothing around it: MIT-licensed 6B weights on Hugging Face since 17 September 2026, a validated single-GPU configuration, and no endpoint, no price, no provider, no independent score, and a Quick Start pointing at a GitHub repository that returns 404. One of these is a product in beta. The other is a weight file in a folder. This piece is about what separates them, which turns out to be almost entirely infrastructure rather than model quality.

Licence is the one dimension the open model wins

Start here, because it is the only place where the comparison is genuinely favourable to Ming-Image-0.1-Design and the margin is large.

• Ming-Image-0.1-Design — MIT licence on the weights. Download them, modify them, fine-tune them, ship them inside a commercial product, host them for customers, with no per-call fee to anyone and no usage terms to negotiate. The repository metadata confirms the licence and the safetensors/diffusers packaging.

• MAI Image 2.6 Preview — a commercial API on Microsoft's terms. There is a rate card and there are terms of service, and the model is accessed rather than possessed. Nothing in the public record suggests Microsoft intends to open these weights.

That is a real advantage and it should not be waved away. A team with a fixed GPU budget, a privacy constraint, or a need to fine-tune on proprietary design assets has no route to MAI Image 2.6 Preview at all — the API cannot be self-hosted, and a preview API carries no guarantee of long-term availability. Ming-Image-0.1-Design can be run on your own hardware forever, under a licence that has been understood by lawyers for decades.

The advantage only converts into value if you can run it, which brings us to the rest of the table.

Everything else, checked one line at a time

• Endpoint — MAI Image 2.6 Preview is served through Microsoft's own platform on a documented identifier vs Ming-Image-0.1-Design, which has none. Hugging Face reports the repository has inference disabled and is not deployed by any inference provider, and the model's own Quick Start tells you to clone a companion repository that does not exist.

• Price — $38.90 per 1,000 images, roughly 3.9 cents an image, which is second-cheapest in the top five of the UI/UX Design category behind Muse Image at $10.00 and comfortably under Grok Imagine Image 2.0 at $60.00 and GPT Image 2.5 at $210.72 vs no price, because there is nothing to buy. The weights are free; the 80 GiB CUDA GPU is not.

• Independent score — 1,168.8 Elo on 818 blind votes in the UI/UX Design category, fifth place, above Nano Banana 2 at 1,166.6 and below Grok Imagine Image 2.0 at 1,181.8. The model's sibling, MAI-Image-2.6-Flash, sits at 1,152.4 on the same board vs nothing. Ming-Image-0.1-Design's only number, 1,082 Elo, appears exclusively on a leaderboard graphic inside inclusionAI's own repository, on a board we could not find on the public web, with no vote count shown.

• Deployment support — a managed platform with uptime and support attached vs a model card recommending vLLM-Omni and linking an installation guide that resolves and a model-specific recipe that 404s. Half a deployment story.

• Resolution and configuration — not published in the figures we could verify for the preview vs 2048 x 2048 recommended or 1024 x 1024 for speed, 12 sampling steps, CFG 1.0, BF16, one 80 GiB CUDA GPU.

What "preview" costs you on the Microsoft side

An honest comparison has to note that the Microsoft column is not without risk of its own. A preview is a preview: identifiers can change, rate cards can move, quotas can be tightened, and a preview endpoint can be withdrawn or superseded — and Microsoft has already shipped a Flash sibling in the same family, which tells you the line is moving quickly. Building a production pipeline against a preview identifier means accepting that the thing you integrated may be renamed under you.

The difference is that the Microsoft risk is a business risk with a counterparty, and the Ming-Image-0.1-Design risk is an engineering risk with no counterparty. If a preview API changes, there is a support channel, a deprecation window, and a documented successor. If the missing repository never appears, there is nobody to ask — the model's documented run path is a 404, and the only alternative is to reverse-engineer the pipeline from the file tree yourself.

Which is not impossible. The repository is well structured for it: a transformer folder with five shards, a VAE, a scheduler, and separate mllm, connector and mlp components, which is the shape of a multi-stage pipeline with a multimodal language model doing the text encoding. The card notes prompt enhancement can be handled by Ling-3.0-flash-VL or qwen3.8-27B, both siblings from the same organisation. A capable team could stand this up. It is just not the same thing as a product.

Where a routing layer fits, and where it does not

Neither model in this comparison is routable through OrcaRouter today, and it is worth stating that without decoration: our catalogue contains no Microsoft MAI entry and no inclusionAI entry as of 23 September 2026, so neither the $38.90 rate nor anything else above is a price you can reach through us. Both figures are the vendors' own published or board-tracked rates.

What the routing layer is for in a matchup like this is the gap between now and later. The economics of image generation have come apart from quality in the last quarter — the category leader costs $210.72 per 1,000 images, the fifth-place model costs $38.90, and a model thirteen Elo below that costs $10.00 — which means the correct model for a given workload is a moving target rather than a fixed choice. Keeping the generation call behind one endpoint across 200-plus models, with provider list prices passed through at 0% markup and automatic failover when a provider wobbles, turns that moving target into a model-ID change instead of an integration project. That is worth more than any single model on this list, and it is the argument for not welding your pipeline to a preview identifier in the first place.

Which one to pick, and what would change the answer

For anyone shipping work this quarter, the answer is MAI Image 2.6 Preview and it is not close. It is callable, it is cheap for its quality band, it has 818 independent blind votes behind a fifth-place ranking, and the preview risk is a manageable commercial risk. The one thing to do before committing is to treat the preview identifier as volatile — keep it in configuration, not in code, and have a second image model one line away in case the identifier moves.

Ming-Image-0.1-Design is worth downloading and not worth planning around. Pull the weights if you have an 80 GiB card and a reason to test text-rich design generation under a permissive licence, because that is the one thing it offers that Microsoft's API structurally cannot. Do not put the 1,082 Elo in a slide — it is a vendor graphic on a board nobody else can inspect, and the model's own repository shows zero downloads, which is what a claim with no outside witness looks like at this stage. The card's recommended settings are at least concrete and checkable: 2048 x 2048 at 12 sampling steps, CFG 1.0, BF16.

The event that would rewrite this piece is the same one that would rewrite every comparison in this series: an independent Ming-Image-0.1-Design score, with a sample count next to it, on a board that inclusionAI does not operate. At that point the question becomes genuinely interesting — a self-hosted 6B model at zero marginal cost against a 3.9-cent preview API at 1,169 Elo — and for the first time both sides of the comparison would have a number that means the same thing.