Qwen-Image-2.1 vs MAI-Image-2.6 Preview: One Is Ranked, One Is Downloadable

MAI-Image-2.6 Preview sits at the top of the independent image-editing leaderboard and you cannot buy it. Qwen-Image-2.1 is a 33 GB download anyone can run today and does not appear on that leaderboard at all. Between those two facts is the whole comparison, and it is not a comparison of quality — it is a comparison of evidence. Microsoft's model has votes from strangers and no checkout page. The other model has a checkout-free download and no votes from strangers. Deciding between them means deciding which of those gaps you can live with. Qwen-Image-2.1 was open-sourced on 20 September 2026; MAI-Image-2.6 Preview has been in private preview since 19 August 2026.

What the ranking actually measures

Artificial Analysis runs blind pairwise comparisons and turns them into an Elo. On the September 2026 snapshot of its image-editing board, MAI-Image-2.6 leads at roughly 1,122 — ahead of GPT Image 2 (high) at about 1,117. On text-to-image it sits second, around 1,149, behind GPT Image 2 (high) at about 1,178. Those are real votes from real people comparing outputs they did not know the provenance of, which is a categorically better evidence class than a vendor's own benchmark.

The caveat is built into the word "Preview". A preview build accumulates votes from a self-selecting population — the people with invite access — and its scores can move materially when the general-availability build replaces it. Treat the Elo as a strong signal about the model's direction and a weak signal about the model you will eventually be able to call.

Qwen-Image-2.1's evidence is the mirror image. The vendor's own Qwen-Image-Bench puts it at 60.28, ahead of Nano Banana 2.0 at 59.82 and GPT Image 1.5 at 59.65, first among open models. Those are vendor-run numbers with no third-party reproduction, and the margins are under a point — a gap that would not survive a change of prompt set. The one independent datapoint is a GenAI Showdown human-scored run at 7/15, up from 4/15 for the original Qwen-Image and behind Ideogram 4's 8/15. On the Artificial Analysis image boards, Qwen-Image-2.1 has no entry. That is not a bad score. It is the absence of one, because independent Elo requires enough people running the model publicly to generate votes, and a model that is one day old and self-hosted only has not had the chance.

Where the two models genuinely differ on the sheet

• Access — Qwen-Image-2.1 is open weights, roughly 33 GB, downloadable today; MAI-Image-2.6 Preview is invite-gated on MAI Playground and Microsoft Foundry with no public API and no rate card.
• Resolution ceiling — Qwen-Image-2.1 outputs natively up to 2048×2048 at 40 inference steps with seven aspect-ratio presets; the MAI-Image-2.5 line capped output at roughly 1,048,576 pixels, and Microsoft has not published a higher figure for the 2.6 Preview.
• Transparency — Qwen-Image-2.1 emits alpha from a 64-channel RGBA latent space, and can extract a subject from a photograph into a transparent layer; no transparent-background output has been documented for the MAI-Image line.
• Reference images — Qwen-Image-2.1 accepts up to 10 for multi-subject composition and virtual try-on; Microsoft has not published a reference-image count for the preview build.
• Editing controls — Qwen-Image-2.1 takes region edits through coloured circles annotating multiple areas in one instruction, free-hand painted marks, or a separate original-plus-mask pair that preserves the full source image; MAI-Image-2.5 was positioned as an editing specialist with local editing and face-identity preservation, and the 2.6 Preview inherits that framing.
• Licence — Qwen-Image-2.1 ships under the Qwen Research License Agreement dated 20 September 2026, non-commercial use only; MAI-Image-2.6 Preview is a commercial preview with no published commercial terms.

The row that should give you pause is not any of the capability lines. It is the last one: neither model has a published commercial term sheet. One forbids commercial use outright until you obtain a separate licence. The other has no price at all, because there is no way to buy it yet.

Editing fidelity is the axis both of them claim

Both models are being sold on the same promise: you can edit an image repeatedly without the subject drifting. Microsoft's framing for the MAI-Image line is face-identity preservation and local editing with visual reasoning behind it — the model is meant to understand what you are pointing at, not just where. Qwen's framing for Qwen-Image-2.1 is explicit about the failure mode it targets: facial-identity drift and the warping of product text, texture and shape across repeated edits. Both are vendor claims, and both describe the same problem.

The practical difference is which one you can test this afternoon. Qwen-Image-2.1 is a download and a diffusers pipeline — QwenImage21Pipeline merged into Diffusers, with day-zero support in ComfyUI, SGLang, vLLM-Omni and LightX2V. You can run twenty edit iterations on your own subject and judge the drift yourself. MAI-Image-2.6 Preview requires an invitation, so for most people the fidelity question stays a claim until access opens.

There is one more edit-specific capability worth flagging because it is unusual. Qwen-Image-2.1 can chain local edits into a simple frame-by-frame animation — the same transparency mechanism that produces a cutout also lets a sequence of region edits build motion. It is not a video model and should not be compared to one, but it is a cheap way to produce a short loop without touching a video endpoint.

Cost, and why neither number is on a price page

Qwen-Image-2.1 has no hosted price because there is no hosted endpoint. The cost is your own GPU time, and the serving numbers are published: SGLang-Diffusion measures 2.748 s for a 1024×1024, 40-step generation on a single B200, 8.23 s on an RTX PRO 6000 Blackwell 96 GB, and 4.74 s on a 24 GB RTX 4090 using Cache-DiT and INT8 kernels, denoise only. Those are single-request latencies, not throughput, but they are enough to sanity-check whether one card covers your volume.

MAI-Image-2.6 Preview has no published price of its own — Microsoft sells it through Foundry and the Playground, not on a public rate card. The independent board does list one, and it is the only price signal attached to this model: $38.9 per 1,000 images, recorded against the preview build by Artificial Analysis. That sits between the standard MAI-Image-2.5 at $48.1 per 1,000 and MAI-Image-2.6-Flash at $19.5, which is roughly where a mid-tier model should land. Treat it as the board's figure rather than a quote you can buy at.

Where OrcaRouter fits in a decision like this is on the hosted side, and being precise about it matters. We route neither model. MAI-Image-2.6 Preview is reachable only through Microsoft's own surfaces while it stays invite-gated, and we carry no Microsoft image route at all. Qwen-Image-2.1 is self-hosted only, and we route no Qwen-Image model of any version. What we do carry is the image line you would otherwise pair these with: OpenAI's GPT-Image-2, GPT-Image-1.5 and GPT-Image-1-mini, Google's Imagen 4 tiers and Gemini image previews, and xAI's Grok Imagine image endpoint — 200+ models behind one OpenAI-compatible endpoint, provider list price passed through at zero markup, automatic failover across providers, a routing DSL for composing fallbacks, and model fusion for panel-style calls. The pass-through matters for a comparison like this one: when a vendor in that set moves a rate, our number moves the same day, which is the only way to keep a routing layer honest against direct vendor pricing.

Which gap you can live with

• You need a number you can defend — MAI-Image-2.6 Preview, if you can get access. Independent blind votes at the top of the editing board beat a vendor benchmark, and if the decision has to be justified to someone else, that is the evidence that survives scrutiny.
• You need to ship something today — Qwen-Image-2.1. The weights are public, the frameworks support it, and the licence permits research and evaluation immediately. No invitation, no waitlist.
• You need transparency or 2K — Qwen-Image-2.1, and this is the one row where the sheet is not close. Alpha comes out of the sampler, and the MAI-Image line has not documented transparent output at all.
• You need to sell the output — neither, as things stand. Qwen-Image-2.1 forbids it without a separate licence and MAI-Image-2.6 Preview cannot be bought. For a commercial pipeline, reach for something with a published rate card and a commercial licence.

The two things worth watching are both about the same gap closing. If Microsoft takes MAI-Image-2.6 out of preview with a public rate card, it stops being a leaderboard curiosity and becomes a purchasable option with the best independent editing evidence on the board. If Qwen publishes commercial terms for Qwen-Image-2.1 and the model starts appearing on the independent boards, the licence row stops doing all the work. Until then, the honest read is that this is not a matchup between two products — it is a matchup between two kinds of missing information.