MAI Image 2.7 Preview vs Qwen Image 3.0: The Open-Weight Exit Nobody Announced

The advice that used to settle this matchup no longer works. Until this summer, the answer to "Qwen image model or a closed American one" was easy: take Qwen, download the Apache-2.0 weights, run it yourself, and stop worrying about rate cards. Qwen Image 3.0 broke that. When it reached general availability on August 4, 2026, it arrived with two paid tiers, a hosted API, and no weights, no model card, no parameter count and no technical report — the first time Alibaba's image line has shipped closed. MAI Image 2.7 Preview, announced by Microsoft AI on October 27, 2026, is closed too, and it does not even have the API. So the honest comparison is between two models you cannot self-host, sold on disclosures that are nearly opposite.

That reframing is the article. Everything else — the Elo, the price per image, the resolution ceilings — follows from which of the two is actually purchasable and what each company was willing to publish about it.

What Qwen Stopped Shipping

Qwen-Image 1.0 and 2.0 were the open line. Apache-2.0 weights, the same day as the announcement, a technical report to go with them, and a community that fine-tuned them into everything from comic lettering to interior design. That is why "just self-host it" became the standing recommendation: the license permitted commercial use and the checkpoint was a download.

Qwen Image 3.0 ended that. The model launched invite-only on July 22, 2026 and reached GA on August 4 in two tiers, Standard and Pro, through Qwen Cloud and Alibaba Cloud's Model Studio. No weights were published with either tier. Alibaba has not published a parameter count, a model card or a technical report, and it has published no benchmark numbers at all — not on Qwen-Image-Bench, not anywhere else. The only independent signal is Artificial Analysis, where in its October 2026 snapshot Pro sat at roughly 1,094 Elo on text-to-image and 1,085 on image editing, with Standard at 1,062 on editing.

There is a wrinkle worth stating plainly because it will confuse anyone who searches: at least one comparison site lists Qwen Image 3.0 as Apache-2.0 with open weights. That is wrong, and it is the kind of wrong that costs a team a week. The open branch of the family is Qwen-Image-2.1, released September 20, 2026, under a non-commercial research license — open to inspect and fine-tune, not open to ship in a product. If a plan depends on self-hosting Qwen, it depends on 2.1 or earlier, not on 3.0.

What Microsoft Published, and What It Did Not

MAI Image 2.7 Preview's evidence is the mirror image. Microsoft has published a great deal about how it ranks and almost nothing about what it is.

The ranking is specific: second on Artificial Analysis's image-editing board at 1,193 Elo on AA-Image-Editing v2.1, evaluated 22 October 2026, six points behind GPT Image 2.5 Sunburst and eleven ahead of GPT Image 2.5 Flare. Microsoft also reports third on Arena Edit. Between MAI-Image-2.6 and 2.7 it says reasoning-based edits and scene-and-style edits both moved from seventh to first in Artificial Analysis's categories, with text and symbol edits from sixth to first, and that the largest lead is 24 Elo in scene and style. Those are the company's own characterisations of a third-party board — vendor-reported claims about independent data, which is a better class of claim than a home benchmark and still not the same thing as a measurement you can reproduce.

What is missing is everything a buyer needs. No output resolution. No reference-image maximum. No mask or bounding-box interface. No Foundry deployment. No rate card. No model card. The model is available in MAI Playground, Arena and Artificial Analysis — two evaluation surfaces and a demo — with text-to-image generation promised for later.

The Spec Lines, Side by Side

• Release — Qwen Image 3.0 invite-only July 22, 2026, GA August 4, 2026; MAI Image 2.7 Preview October 27, 2026.

• Weights — none published for Qwen Image 3.0, a reversal from 1.0 and 2.0; none for MAI Image 2.7 Preview, which is consistent with Microsoft's line. The open Qwen branch is Qwen-Image-2.1 under a non-commercial license.

• Scope — Qwen Image 3.0 generates and edits across Standard and Pro tiers; MAI Image 2.7 Preview edits only.

• Resolution — Qwen Image 3.0 tops out at 2048×2048 on the vendor's own figures; unpublished for 2.7, whose predecessor capped at 2,359,296 pixels.

• Price — Qwen Image 3.0 Standard is $0.03 per image at 1K or 2K with edit inputs at roughly $0.003 each; Pro is $0.04 at 1K and $0.075 at 2K. MAI Image 2.7 Preview has no published price.

• Access — Qwen Image 3.0 is on Qwen Cloud, Alibaba Cloud Model Studio and Qwen Chat; 2.7 is on MAI Playground, Arena and Artificial Analysis, with no API.

• Published benchmarks — Qwen Image 3.0 has no vendor benchmarks at all; 2.7 has a third-party board placement and category rankings, quoted by the vendor.

Two Different Kinds of Silence

It is tempting to file both of these under "undocumented", but they are not the same failure and they cost a team different things.

Alibaba's silence is about capability. We know the interface, the price, the two tiers and the maximum resolution; we do not know what the model is good at, because no benchmark was published and the only independent score is an October snapshot that puts Pro's editing Elo in the mid-1,080s, comfortably below the leaders on that board. A buyer can price the work — $0.03 an image for Standard undercuts Nano Banana 2.1's $0.0336 at 1K, though not Meta Muse Image's flat cent — but cannot predict quality per use case without running their own evaluation.

Microsoft's silence is about procurement. We know roughly how good it is, because a blind-voting board measured it, and we know which categories it claims. We do not know what it costs, what size it outputs, how many references it takes, or how to call it. A buyer can predict quality and cannot buy.

Rank the two on what a plan needs and the answer depends on the stage you are at. If you are choosing a model to build on, Alibaba's silence is survivable — run a paid evaluation, the interface is documented, the bill is known. If you are choosing whether to bet a roadmap on Microsoft's, its silence is not survivable, because there is nothing to integrate against.

What a Penny of Editing Looks Like at Qwen's Price

Qwen's pricing is the most concrete thing in this matchup, and it is worth working through because the tiers are not where you expect the difference to be.

Standard at $0.03 per image, 1K or 2K, with edit inputs around $0.003 each. A product-retouch batch of a thousand images with two reference inputs per edit costs $30 for outputs plus about $6 for inputs, so roughly $36. Pro at $0.075 for 2K output and two references lands near $81 for the same batch — about 2.2 times Standard for a tier whose only independent signal is an October snapshot.

For comparison, Meta Muse Image does the same thousand images for $10 flat, and Nano Banana 2.1 does them at 2K for about $50 before input tokens. Qwen Image 3.0 Standard is competitive but not the cheapest, and its input charges mean the gap widens on reference-heavy work.

Against MAI Image 2.7 Preview the arithmetic cannot be run at all, and that is the point. A preview is free in the sense that nobody is billing you, and expensive in the sense that every image costs a human round trip. At a thousand images, the labour is the entire budget.

The Category Claim Is the Thing to Test

Microsoft's most specific assertion about MAI Image 2.7 Preview is that it now leads Artificial Analysis's reasoning-based editing category, up from seventh. That is a narrow, testable claim, and it is the one worth designing an experiment around — because it is also the claim Qwen Image 3.0 cannot answer.

Alibaba published nothing about reasoning, tool use or multi-step edits. What its documentation does claim is a set of practical text capabilities: prompts up to 4,500 tokens, text legible down to about 10 pixels, twelve native languages, more than twenty fonts, up to six outputs per call, and server-side prompt rewriting. Hands-on accounts describe it as strong on UI mockups and diagrams with occasional garbled small text and slower latency than the Flash-tier models.

So the two models are aiming at different failures. Qwen Image 3.0 is optimised for text rendering and layout fidelity inside a single pass. MAI Image 2.7 Preview is optimised for edits that need a judgement about how the whole scene should change. If your work is a poster with a lot of type, you are testing Qwen's axis. If it is "put this product in that room and make the light agree", you are testing Microsoft's, and the leaderboard says Microsoft is ahead on it — on other people's images.

Where Neither One Answers the Phone

Neither model is on OrcaRouter, and it is worth being direct about that rather than implying otherwise. What is there is the set of image models that can actually be called — GPT Image 2.5 Sunburst and Flare, the GPT Image 2 family, Imagen 4 in three tiers, Google's Nano Banana line, Grok Imagine Image — behind one key at provider list prices with no markup.

That matters for this particular decision because of how you would settle it. Qwen Image 3.0's capability is unbenchmarked by its vendor, so the only way to learn whether it beats a model you already use is to run the same briefs through both. OrcaRouter's routing DSL lets one call fan a single brief across several models and collect the results in one shape, and model fusion goes a step further by having a panel of models answer together when a single edit keeps failing. Neither feature makes MAI Image 2.7 Preview callable — nothing does yet — but the former is how you would build the evidence for or against Qwen Image 3.0 without committing to its console, and the latter is what you reach for when a reasoning-heavy edit has to succeed on the first try.

What to Do, and What to Stop Assuming

Stop assuming Qwen's image line is open. It is not, at 3.0. If your architecture depends on downloadable weights, the model you want is Qwen-Image-2.1, and the license you will be reading is non-commercial — which means the self-hosting plan has a legal boundary to check, not just a hardware bill.

If you need a cheap editor with a documented interface this week, Qwen Image 3.0 Standard at $0.03 an image is a reasonable bet with an unusual caveat: nobody, including its vendor, has told you what it is good at. Budget an evaluation, not a decision.

If the work is whole-scene editing and you can afford a manual loop, spend the afternoon in Arena with MAI Image 2.7 Preview. The reasoning-based category leading is the most specific claim in the launch and the one most likely to hold, because it is the only claim backed by a board run rather than a press release.

And keep one date in view. The moment Microsoft puts MAI Image 2.7 Preview in Foundry with a per-image price, this comparison changes from a description of two purchase paths to a real contest, and the $0.03 Qwen edit becomes the number to beat. Until then, one of these models has a price and a problem, and the other has a score and a wall.