Qwen-Image 2.1 vs Meta Muse Image: The Model You Can Call vs the Model You Can Keep

There is a version of this comparison that is about image quality, and it cannot be written yet — not honestly. Qwen-Image 2.1, which the Qwen-Image team published on 20 September 2026, has no independent score of any kind. Meta Muse Image, launched by Meta Superintelligence Labs on 7 July 2026 as Meta's first in-house image model, has been measured — it sits third on the Artificial Analysis image editing board at an Elo of about 1,105, inside the top five for text-to-image, and on the quality-versus-price Pareto frontier at $10 per thousand images. But the more useful question about these two is not which one is better. It is what you are allowed to do with each, and those answers are almost perfectly inverted.

Access and inspectability are two different things, and nobody offers both

Meta Muse Image is the one you can call. Since August 2026 it has been reachable through Meta's own Model API at a flat $0.01 per image — roughly $10 per thousand — over an OpenAI-compatible endpoint, which is a real change from how the model was described when it launched in July, when there was no developer route to it at all. That flat rate is unusual in a category where almost everything else is metered in tokens, and it makes cost forecasting trivial: a thousand images is ten dollars, whether they are simple or elaborate.

Qwen-Image 2.1 is the one you can keep. It is a weights download — roughly 33 GB across the diffusion transformer, the text encoder and the VAE — published under the Qwen Research License Agreement dated 20 September 2026, which grants rights "FOR NON-COMMERCIAL PURPOSES ONLY" and requires a separate licence for commercial use. There is no hosted endpoint. There is no flat rate. There is also nothing hidden: you can read the architecture, read the prompt-rewriting system prompt, diff it, override it, and run the whole thing on your own hardware.

So the trade is not quality against quality. It is a metered, measured, commercially licensed model you do not own against a free, unmeasured, non-commercial model you own completely. Very few teams can take either side of that without giving something up.

Why Muse is not quite a diffusion model in the usual sense

The thing that makes Meta Muse Image unusual is that it does not only generate. It runs a loop: it can search the web, write and run code, and review its own output before returning an image. That is Meta's own description, and it is the reason the model is interesting rather than merely competitive — the agentic steps are where prompt ambiguity gets resolved, and where a request like "make me a chart of this year's numbers" can be answered rather than approximated.

It is also the reason its behaviour is harder to reason about. A conventional image model is a function from prompt to pixels. An agentic one has a trajectory, and the trajectory is not something you can read from a model card. The vendor-reported figures — 94% multi-image character consistency, up to 10 reference images, output up to 1600 pixels on the long edge — describe the envelope, not the loop.

Qwen-Image 2.1 solves the same ambiguity problem in the opposite way: explicitly, and in the open. Alongside the image model, the release ships two prompt-rewriting checkpoints — Qwen/Qwen-Image-2.1-PE-T2I and Qwen/Qwen-Image-2.1-PE-I2I, each a fine-tuned Qwen3.5-VL 9B at about 18.8 GB — that take a vague instruction and rewrite it into a precise directive before the diffusion model sees it. The I2I variant always has an input image, so it is always an editing task. And critically, the rewriter's behaviour is specified in a system_prompt.txt that ships in the repository, which means the step that decides what your prompt means is something you can read and change.

Meta's agentic loop and Alibaba's prompt rewriter are both answers to "the model does not know what you meant." One is a service that decides for you. The other is a file you can edit.

What the numbers look like when only one side has them

Meta Muse Image is on both Artificial Analysis image boards. It is third in image editing at about Elo 1,105, behind MAI-Image-2.6 at 1,122 and GPT Image 2 (high) at 1,117, and it is inside the top five for text-to-image at around Elo 1,116. At $10 per thousand images it sits on the quality-versus-price Pareto frontier, which is the useful place to be: not the best model on the board, but one of the few where paying more buys you measurably more.

Qwen-Image 2.1 has none of that. It has a vendor blog post with a Qwen-Image-Bench chart that no third party has reproduced, and it has one hands-on report — an early-access tester in the Qwen Ambassador program who ran the final weights through a ModelScope Studio interface and reported roughly 10–15 seconds for text-to-image and 18–23 seconds for editing, with multi-reference consistency degrading from about three input images onward. One reviewer, no timer, no prompt set. It is a useful signal and it is not a benchmark, and the honest way to hold it is as a hint about where the model might land rather than evidence of where it is.

Where Qwen-Image 2.1 is concretely ahead is not quality but capability at the pixel level: native 2K output at 2048×2048 by default, 40 inference steps, seven aspect-ratio presets, up to 10 reference images, local edits by circle, painted annotation or a separate mask, and RGBA generation with transparent-layer editing and subject extraction from RGB photos. Meta's model caps at 1600 pixels and its transparency story is not published. If you need a real alpha channel out of the box, that is a decision made for you.

The flat rate is the part worth thinking about

A flat cent per image is a pricing decision, not a technical one, and it changes what the model is for. Token-metered image pricing punishes complexity: a long instruction with several reference images costs more than a short one, which means the cost of an image is coupled to how hard it was. At a flat rate that coupling disappears, and the rational behaviour changes with it — you stop optimising prompts for length and start using the model the way it was designed to be used, with the agentic loop and the reference images doing their job.

It is also the kind of pricing that only a company serving its own model can offer, which is worth naming when you are choosing a provider. OrcaRouter fronts 200+ models behind one OpenAI-compatible endpoint at provider list price with zero markup, with automatic failover across providers and a routing DSL that composes several models into a single call. The relevant property here is the pass-through: because we charge the provider's list price rather than a marked-up one, a vendor pricing change — a flat rate appearing, a token rate dropping — is live on our side the same day instead of waiting on a contract. Neither Meta Muse Image nor Qwen-Image 2.1 is among the models we route today and this article is not going to imply otherwise; the image models we do front are OpenAI's GPT-Image family, Google's Imagen 4 tiers and Gemini image previews, and xAI's Grok Imagine image endpoint.

The contrast that matters for this comparison is that one of these two models has a price at all. Meta Muse Image costs $0.01 per image, published, on an API you can call this afternoon. Qwen-Image 2.1 costs a 33 GB download, a GPU you already own, and a legal review of a licence that says non-commercial only.

Where they accidentally converge

Both models accept up to 10 reference images. That is not a coincidence so much as an industry settling on an answer: ten is enough for a character sheet, a product set, or a style reference pack, and past that the conditioning stops helping and starts fighting. Meta publishes a 94% multi-image character consistency figure for its model; Alibaba publishes no equivalent, and the one independent hands-on report suggests consistency begins to degrade around the third reference image. Two models claiming the same envelope with wildly different evidence behind it is the whole comparison in miniature.

They also converge on the problem neither has solved: neither gives you both. Meta Muse Image cannot be downloaded, inspected, or run on your own hardware, and its agentic loop is a black box by construction. Qwen-Image 2.1 cannot be sold, cannot be called, and cannot be measured against anything yet. If your constraint is a deadline, one of these is usable today. If your constraint is a compliance review or a need to understand exactly what your pipeline is doing, the other one is.

Pick by what you have to do next, not by which is better

If you need to ship image generation this quarter under a commercial licence with a measurable model behind it, Meta Muse Image is the answer and $0.01 per image is a number you can put in a budget. If you need to understand an image model end to end — the attention masking, the prompt rewriter, the VAE, the licence — Qwen-Image 2.1 is the only one of the two that lets you, and it is a research licence away from being unusable for anything else. Neither choice is a compromise on quality, because the quality question is still open on one side. It closes when enough people run Qwen-Image 2.1 in public to put it on a board, and the early-access testers were asked to publish by 29 September. That is the date this comparison becomes a real one.