MiniMax-H3-Image vs MAI-Image 2.7 Preview: One Is a File, One Is a Preview

Hand MAI-Image 2.7 Preview a text prompt and nothing comes back, because it does not generate from text. It edits. That single sentence explains most of the difference between it and MiniMax-H3-Image, and it is the part every side-by-side comparison of the two gets wrong, because the two models are not solving the same problem: one is a unified text-to-image and reference-editing checkpoint with downloadable weights, the other is Microsoft AI's editing-only preview, reachable through a playground and a leaderboard and not much else.

Both are recent. M​iniMax published MiniMax-H3-Image's weights on 10 November 2026 at 14:50 UTC; Microsoft's preview has been in the wild since late October, and the independent board it lives on shows it moving. What follows is not a feature list — it is an attempt to answer the only question that matters if you are choosing: which of these can you actually put into a pipeline, at what cost, and what do you give up by picking it.

The first difference is a capability, not a benchmark

MiniMax-H3-Image does two jobs from one checkpoint. It generates an image from a prompt, and it edits one or more reference images passed positionally as <Picture 1> through <Picture 16>. The same weights serve both; there is no separate editing endpoint to configure and no second model to pay for.

MAI-Image 2.7 Preview does the second job. It is an editing model — you give it an image and an instruction and it returns a revised image. There is no text-to-image mode to enable. For a team whose workflow is "generate a product shot, then iterate on it", that means Microsoft's model can cover the iteration half and cannot cover the first frame, so it is never a drop-in replacement for MiniMax-H3-Image. It is at best a second model in a chain.

That asymmetry is worth stating plainly because it changes what a comparison can even mean. You can compare them on editing quality. You cannot compare them on generation, because one of them does not do it.

One hands you weights, the other hands you a preview badge

MiniMax-H3-Image ships under the MiniMax-H3-Image Community License Agreement, which is not Apache 2.0 and deserves a lawyer's read before a commercial deployment, but it is a downloadable artefact: 68.65 GB of diffusion transformer, VAE and configuration, plus a 66.7 GB text encoder borrowed from the MiniMax-H3 video repository. You can run it on your own hardware through SGLang, vLLM-Omni, diffusers or ComfyUI, and vLLM's team shipped day-zero support for both NVIDIA and AMD on release day. You can also call it on M​iniMax's own API, where the model identifier is MiniMax-H3-Image and a Highspeed tier runs the same limits at a higher output rate.

MAI-Image 2.7 Preview is closed in the ordinary sense: no weights, no licence to read, no model card with a parameter count or an architecture. Microsoft has not published a per-image price, a maximum resolution, a reference-image ceiling, a mask interface, or a Foundry listing with an endpoint a developer could call today. What exists is the MAI Playground, an appearance on Arena, and an entry on Artificial Analysis. That is enough to form an opinion of the model and not enough to build a budget around it.

This is the difference that decides most real decisions. An open-weight model with no benchmark score is an engineering risk you can retire by running it. A closed model with no price is a procurement dead end, however good the samples look.

What the independent board actually says

As of 11 November 2026, MAI-Image 2.7 Preview sits at fourth on Artificial Analysis's AA-Image-Editing v2.1 board with a 1123 Elo, behind GPT Image 2.6 Sunburst at 1209, GPT Image 2.6 Flare at 1206 and GPT Image 2.5 Sunburst at 1126. Its sample count is 5,280, which is a real sample count rather than a placeholder. When Microsoft announced the preview on 27 October 2026 it described the model as second on that board at 1,193 Elo, six points behind GPT Image 2.5 Sunburst. The board has since filled with newer O​penAI entries and the position is now fourth at 1123 — which is what happens to a ranking in a fast month, and a reason to read any Elo with its capture date attached rather than as a permanent fact. The vendor's own account of the subscores (reasoning-based edits and scene-and-style edits moving from seventh to first, text and symbol edits from sixth to first, a 24-Elo lead in scene and style) remains a vendor characterisation of third-party data: a better class of claim than a home benchmark, and still not a measurement you can reproduce.

MiniMax-H3-Image has no row on that board, and no row on AA-Image-T2I v2.1 either. Not a low one — none. It was released on 10 November, sample collection for a pairwise-preference board takes time, and there is no number to quote. Anyone who tells you one model beats the other on an independent score is comparing a ranked model to an unranked one, which is not a comparison.

So the honest scoring position is: MAI-Image 2.7 Preview has a fourth-place editing score on a board whose top is crowded with O​penAI models, and MiniMax-H3-Image has no score at all. That is not an argument that M​iniMax's model is better or worse. It is an argument that the only people who know how good it is are the ones who have run it.

Dimension by dimension

• What it does — MiniMax-H3-Image: text-to-image and reference editing in one checkpoint. MAI-Image 2.7 Preview: editing only; no text-to-image mode.

• Output size — MiniMax-H3-Image: about sixteen megapixels at every aspect ratio, 4096×4096 at 1:1 up to 6272×2688 at 21:9, with a 20 MP API ceiling. MAI-Image 2.7 Preview: not published; its predecessor capped around 2.36 megapixels.

• References — MiniMax-H3-Image: up to 16, addressed by position in upload order. MAI-Image 2.7 Preview: no documented ceiling.

• Weights — MiniMax-H3-Image: 68.65 GB plus a 66.7 GB text encoder, under a community licence that is not Apache 2.0. MAI-Image 2.7 Preview: none, and no signal that any are coming.

• Price — MiniMax-H3-Image: $0.02179 for a 1K image, $0.02333 for 2K, $0.06879 for 4K list and $0.03439 under a 50% discount ending 17 November 2026 at 16:00 UTC. MAI-Image 2.7 Preview: no published price at all.

• Independent editing score — MAI-Image 2.7 Preview: 1123 Elo, fourth on AA-Image-Editing v2.1. MiniMax-H3-Image: no row on either board as of 11 November 2026.

• Sampling — MiniMax-H3-Image: fixed at six steps, no classifier-free guidance, no negative prompts; any other step count is rejected with HTTP 400. MAI-Image 2.7 Preview: not disclosed.

• Where you can call it — MiniMax-H3-Image: M​iniMax's own API plus self-hosting through four documented frameworks. MAI-Image 2.7 Preview: the MAI Playground, and wherever Microsoft puts it next.

An edit, priced on the one side that has a price

MiniMax-H3-Image bills tokens in and out rather than images, and the card's worked examples are the readable version: a 1,024×1,024 edit with one 1K reference comes to $0.02654, because you pay $1.50 per million input image tokens on top of the output. A 2K generation is $0.02333 and a 4K generation is $0.06879 list. Batch a thousand 2K edits with one reference each and you are in the region of $26 to $30, depending on how much the reference costs to encode — a number you can put in a spreadsheet today.

You cannot do that arithmetic for MAI-Image 2.7 Preview, and this is where the missing price stops being a documentation gap and becomes a planning problem. A preview with no published rate cannot be budgeted, cannot be compared, and cannot be committed to, regardless of where it sits on a leaderboard. If Microsoft's model is genuinely better at editing — and a 1123 Elo fourth place is a respectable place to be — the cost of finding out is not a per-image fee, it is a week of integration against an interface that may change before it ships.

OrcaRouter routes neither of these models, and this piece is not going to pretend otherwise. What we do route is image generation you can call on the same key as everything else — gpt-image-2.5-sunburst, gpt-image-2, the G​emini image previews, Imagen 4 across three tiers, G​rok Imagine Image — at provider list price with 0% markup passed through, so when a vendor moves a rate your bill moves the same day. For a team prototyping an image feature while waiting for M​iniMax's editing quality to be measured or Microsoft's preview to grow a price tag, that is a way to keep the pipeline moving without a second contract. It is not a substitute for either model, and the difference matters.

Sixteen references against an undocumented ceiling

The most underrated line in MiniMax-H3-Image's spec sheet is the reference count. Sixteen images, addressed positionally, is enough to condition an edit on a character sheet, a product, three angles of a room, a colour reference and a style board in one call. Very few image models document that capacity, and almost none of the closed ones document it at all — MAI-Image 2.7 Preview does not, so a workflow that depends on many references has to be validated empirically before it can be designed around.

One caveat on that number, because it will bite someone: the model card says 16 references, while vLLM-Omni's release recipe for the same weights documents up to 9 images per request. Both figures can be correct at once — a serving stack's batching limit and a model's trained capacity are different things — but if your workflow needs twelve references, the honest position is that you should test it on your own serving path rather than trust either document.

What to do with this, this week

If you need generation at high resolution with the option to self-host, the decision is not really between these two models; MAI-Image 2.7 Preview is not in that race. Take MiniMax-H3-Image, price it at $0.02333 for 2K and $0.06879 for 4K, decide before 17 November whether the 4K discount changes your plan, and read the licence before the download rather than after.

If you need the best editing available and you are willing to work inside a preview, MAI-Image 2.7 Preview's fourth place on an independent editing board is a real signal and its lack of a price is a real obstacle. Ask Microsoft for Foundry access and a rate card. Until one of those exists, an editing pipeline built on it is a prototype, not a product.

The thing to watch is not the next leaderboard refresh. It is whether MAI-Image 2.7 Preview gets a price and an endpoint, because that turns this from a comparison of an open model against a closed demo into a genuine choice between two purchasable things. And whether MiniMax-H3-Image gets a score, because right now the model with weights is the only one of the two whose quality you can measure yourself — which, for anyone who has been burned by a preview, may be the more valuable of the two unknowns.