Krea 3 vs Qwen Image 3.0: The Open Line Stopped Here

The most useful fact about Qwen Image 3.0 has nothing to do with its benchmarks. Alibaba's image line was open for two generations — Qwen-Image 1.0 and 2.0 shipped under Apache 2.0 with same-day technical reports, and Qwen-Image-2.1 followed in September under the more restrictive Qwen Research Licence. Qwen Image 3.0, generally available since 4 August 2026, has no weights, no technical report, no parameter count and no model card. The line that made Alibaba the open image lab simply stops, one model short of the flagship. Krea 3, released on 15 October 2026, is closed too — but Krea at least published an open image model earlier this year, and both companies have now made the same turn in the same direction. The matchup is really about what you get in exchange.

What Alibaba stopped publishing, and what it started charging

Qwen Image 3.0 exists only as a metered API product on Qwen Cloud and Alibaba Cloud's Bailian / Model Studio. That is a break in kind, not degree.

• Qwen-Image 2.1 — 7B parameters, Qwen Research Licence, non-commercial by default, with commercial use requiring a separately negotiated licence by email. Default resolution 2048×2048, seven presets, 40 steps, RGBA transparency in and out, up to ten reference images, local edits by drawn circles, painted annotations or a separate mask.

• Qwen Image 3.0 — no weights at all. No repository, no checkpoint, no technical report, no parameter count. Artificial Analysis lists it without an open-weights link, which is the leaderboard's way of saying the same thing.

• Two tiers, and both are on the board — at different heights. Qwen Image 3.0 Standard and Qwen Image 3.0 Pro. Artificial Analysis carries both: Pro at 1,085 Elo on image editing, Standard at 1,062. Alibaba itself publishes no benchmark for either tier, so the leaderboard placement is the only score either one has.

• Pricing runs $0.03 to $0.04 per image. Model Studio's documentation lists $0.035 for qwen-image-3.0; domestic consumer pricing is around ¥0.18 per image at Standard, rising to roughly ¥0.25 at 1K and ¥0.50 at 2K for Pro, with image input and editing input metered separately at ¥0.02 each — a reference image fed back for a second pass is billed again on input.

Krea made a comparable move with less distance travelled: Krea 2's weights are open under the Krea 2 Community Licence, Krea 3's are not, and the self-hosting answer is an enterprise conversation about licence terms and hardware sizing. The difference is that Krea's open release is five months old and Krea has never been the open-weights lab. Alibaba's was the whole pitch.

Two different editing problems, both priced per image

This is where the two models stop being interchangeable, and it is not about quality.

• Qwen Image 3.0 is built for density — text, layout, documents, structure. Vendor claims: 4,500-token prompt capacity, roughly 4.5 times the prior generation; legible text rendering down to about ten pixels including mathematics and LaTeX; native rendering in twelve languages across more than twenty fonts and a hundred artistic styles; generation and image-to-image editing in one model; and the ability to search the web while generating.

• Krea 3 is built for precision and repetition. Change only what you asked, keep the product's label and logo sharp, hold a character across a series, run it a hundred thousand times without the bill becoming the story. Bounding boxes place new elements; multi-turn edits hold across turns; up to six references per edit.

• Resolution — Krea 3 renders natively at 2K with 4K for print; Qwen Image 3.0 tops out at 2048×2048 and can return up to six images in a single call. The per-call multiplier is Qwen's advantage; the print ceiling is Krea's.

• Interface — Qwen's is a prompt and a reference set. Krea's is a prompt, a reference set, and a box you draw to say where the change goes.

If you are generating an infographic, a newspaper page, an exam paper or a UI mockup with real type in it, Qwen Image 3.0 was designed for that and Krea 3 was not. If you are swapping one product into forty scenes and the label must survive forty times, that is the reverse. The launch pages barely overlap in the examples they show, and that is not a coincidence.

What the independent boards say — about one of them

Qwen Image 3.0 has a measured record, and it is not a flattering one for a flagship.

• Image editing, read 16 October 2026 — Qwen Image 3.0 Standard 1,062 Elo from 7,473 samples at $30 per 1,000 images, sitting in the board's 21-26 rank band; Qwen Image 3.0 Pro 1,085 Elo from 12,851 samples at $40 per 1,000, in the 15-20 band. Twenty-three Elo is the whole distance between the cheap tier and the expensive one.

• Text-to-image on the same date — Standard 1,069 Elo from 5,619 samples at $30 per 1,000; Pro 1,094 Elo from 5,898 samples at $40 per 1,000. Same shape, same gap, slightly wider.

• The board's leaders are well clear. OpenAI's GPT Image 2.5 Sunburst sits at 1,198 on editing and 1,217 on text-to-image, with the Flare variant at 1,177 and 1,207 behind it. Qwen's Pro edition is mid-table with a twelve-thousand-vote sample behind it, which is at least a measurement rather than a claim.

• The tier that makes Qwen interesting is the weaker one, and now measurably so. The Standard edition is $10 per thousand images cheaper and 23 Elo worse on editing, and Alibaba publishes no benchmark that would let you argue the gap is smaller than that. That trade — a tenth off the price for a measurable step down in quality — is the actual decision, and it is the one the vendor's own launch page does not draw.

Krea 3 has none of this. No Artificial Analysis entry, no arena result, nothing third-party, as of 16 October 2026, one day after launch. What it has is a launch page with its methodology printed on it: the speed figures are median request-to-result times for each model's twelve-image batch including queue time, measured on Krea's platform from 24 September to 1 October 2026, and one comparison price on the same page is described in its own footnote as "a draft estimate pending validation." That level of disclosure is rare and it is still a vendor measuring itself.

So the honest shape of this comparison is: one model has a mediocre measured score and a hidden cheap tier, the other has no score and a revealed price. Neither position is comfortable, and choosing between them means choosing which kind of uncertainty you can live with.

The cost arithmetic, run honestly

At list prices the two are close enough that the difference is in the structure, not the number.

• Qwen Image 3.0 Standard — $0.035 per image at Model Studio list, roughly $30 per 1,000 on Artificial Analysis's derivation. Qwen Image 3.0 Pro is roughly $40 per 1,000.

• Krea 3 — $0.03 per image at 1K and $0.06 at 2K. At 1K the two are within half a cent; at 2K Krea 3 is nearly double Qwen Standard's per-image list.

• Qwen's metering has traps Krea's does not. Image input and editing input are billed separately at ¥0.02 each, so an iterative workflow that feeds a previous result back in pays again for it. A per-image price does not move when your prompt gets longer; a token-metered one can.

• Krea's metering has a floor Qwen's does not. Krea 3's per-image price is per edit, full stop, and there is no published cheaper tier for small outputs. Qwen Standard can undercut it, and does, on the rung where the two overlap most.

The practical read: if your workload is single-pass transformation of small-to-medium images, Qwen Image 3.0 Standard is plausibly cheaper and Krea 3 is plausibly faster. If it is a long multi-turn edit session on one asset, the input-metering on Qwen's side is the number to model before you commit, because that is where the bill compounds.

Availability, and where the routing layer fits

Qwen Image 3.0 is served on Qwen Cloud and Alibaba Cloud's Bailian / Model Studio, and Krea 3 is served on krea.ai, through a Krea API endpoint documented as beta, and through a set of third-party distribution partners. Neither is on OrcaRouter's catalogue — our image line runs to OpenAI's GPT-Image models, xAI's Grok Imagine image model and Google's Imagen and Gemini image families. We carry Alibaba's Qwen text models; we do not carry a Qwen image model.

Where that matters is the adapter problem. A team that wants Qwen Image 3.0 for dense text layouts and Krea 3 for catalogue repetition is running two closed vendors with two billing models, one of which meters by token and the other by image. That is exactly the shape of integration a single endpoint is built to collapse: one key across 200-plus models, provider list price passed through at 0% markup so a vendor's price change is live the same day, automatic failover across providers when one degrades, and a routing DSL that can express one workload's generation step and another's edit step without a second SDK.

The reason this comparison is on our pages at all is that both of these are models you would call, not models you would download. When neither vendor will ship you the weights, your leverage is entirely in how cheaply and reliably you can call them — and that is a routing question, not a model question.

The decision

• Dense text, documents, layouts, multi-language rendering — Qwen Image 3.0. The vendor claims are specific and unusual — legible type to about ten pixels, LaTeX and mathematics included, twelve languages across twenty-plus fonts — and no other model in this batch is positioned there.

• Product and character consistency at volume — Krea 3. Six references per edit, bounding boxes, precise mode, and LoRA training on the model itself, which Qwen Image 3.0 does not publish.

• You need the cheapest per-image call and can accept mid-table quality — Qwen Image 3.0 Standard at roughly $30 per 1,000. Its 1,062 Elo on editing is real, measured and 136 below the board leader; what does not exist is any vendor benchmark that argues the gap is narrower.

• You need a measured number before signing — Qwen Image 3.0, not Krea 3. Both Qwen tiers carry one, and the cheaper tier carries it too, which is more than Krea 3 can offer. Neither number is one you would put at the top of a procurement document; both are ones you can cite.

One thing both labs did this year is worth more attention than either launch. Alibaba spent two generations proving that a frontier image lab could ship open weights and a technical report on the same day, and stopped doing it for the model that matters. Krea did it once, in May, and has not repeated it. That is a direction of travel rather than a coincidence, and it is the reason the ability to call a model without integrating it has become the more portable skill.