Qwen-Image 2.1 vs Krea 2: Which Open Image Model You Can Actually Ship

Two open image models, both downloadable within the last four months, answering the same question in opposite ways. Qwen-Image 2.1 is the vendor's 7B text-to-image and image-editing model, released 20 September 2026 under a research-only licence — non-commercial only. Krea 2 is the first foundation image model Krea trained from scratch, published as two checkpoints in June 2026 under the Krea 2 Community License, which permits commercial use outright below a revenue line. On the Artificial Analysis text-to-image board today the two sit within eleven Elo points of each other. On the question a team actually has to answer before it builds anything, they are not close at all.

What Krea 2 actually is

Start here, because the naming is misleading and most comparisons get it wrong. Krea the company runs a hosted creative platform with a paid subscription; Krea 2 the model is a separate thing that happens to share the name. Krea released it as an open-weights Diffusion Transformer with two checkpoints, published on Hugging Face as krea/Krea-2-Raw and krea/Krea-2-Turbo, with inference code in a public repository.

The division of labour between them is deliberate and is stated plainly in Krea's own repository: RAW is the undistilled base, meant for fine-tuning, post-training and LoRA training; Turbo is an eight-step distilled checkpoint meant for fast inference. Krea's guidance is to train on RAW and run on Turbo, and its LoRA ecosystem is built that way — adapters trained against the base are applied to the distilled checkpoint at inference time. The two checkpoints share a licence and a family, but they are not interchangeable and they do not have the same settings.

One detail worth flagging, because it will confuse anyone reading the card: the Hugging Face metadata for both checkpoints lists 13B params in the sidebar while the card text says "Diffusion Transformer with 12 billion parameters." Krea has not reconciled the two figures publicly, so treat the parameter count as roughly twelve to thirteen billion and unconfirmed rather than as a settled number.

The numbers, from the board both vendors point at

Krea's repository states that Krea 2 is the top text-to-image model from an independent lab on Artificial Analysis; the technical report says it is "among the top 10" and "second place among models from independent labs." Those are Krea's own claims, made in June. What the live board shows on 13 October 2026 is slightly narrower, and worth reading directly:

• Qwen-Image 2.1 — 1,038 Elo, 95% interval 1,029 to 1,047, from 5,554 samples, rank 20 of 168. The board's price column reads "No API available."

• Krea 2 Large — 1,027 Elo, interval 1,020 to 1,034, from 17,514 samples, rank 25 of 168, at $60.00 per 1,000 images.

• Krea 2 Medium Turbo — 1,022 Elo, interval 1,014 to 1,030, from 6,454 samples, at $15.00 per 1,000 images.

• Krea 2 Medium — 1,014 Elo, interval 1,007 to 1,021, from 17,284 samples, at $30.00 per 1,000 images.

All four of those rows are from the text-to-image board. On the editing board, only Qwen-Image 2.1 appears at all — 1,072 Elo, interval 1,063 to 1,081, from 6,132 samples, rank 21 of 81. There is no Krea 2 entry on that board. All figures here are Artificial Analysis's blind-vote Elo, captured from the live English leaderboard on 13 October 2026; they are independent of both vendors, and the intervals are wide enough that the gap between Qwen-Image 2.1 and Krea 2 Large should not be read as a ranking.

The "Krea 2 Large" and "Krea 2 Medium" rows are also not the open checkpoints. They are Krea's hosted tiers, which is why they carry a per-image price and the open repositories do not. Comparing the downloadable Krea 2 RAW to a hosted Krea 2 Large is comparing a checkpoint to a service.

The real split is editing, not quality

Quality is close enough that the board does not settle it. Capability does. Qwen-Image 2.1 is one model doing generation and editing: it accepts a prompt alone, or up to ten reference images with an instruction, and it supports local edits by drawn circles, painted annotations, or a separate mask, with RGBA transparency in and out. Krea 2's open release is a text-to-image family. Its model card documents prompt-to-image generation; there is no editing checkpoint in the organisation and no masked-edit mode in the inference script, whose flags cover prompt, steps, guidance, resolution, seed and image count — nothing that takes an input image.

That asymmetry comes from the released material, not from what either lab can do. Krea's own platform sells editing, drawing on a roster of third-party editing models. The relevant fact for a reader downloading weights is narrower: if your workflow needs "take this photograph and change one thing about it," the open Krea 2 you can download does not do it, and Qwen-Image 2.1's whole architecture is built around doing it.

Two licences, two different walls

This is where the comparison stops being close, and it is the part most summaries flatten into "both are open." Neither is open source in the permissive sense and they are not open under the same terms at all.

• Qwen-Image 2.1 — the Qwen Research License Agreement, dated 20 September 2026. The grant is non-exclusive, worldwide, royalty-free, and limited to non-commercial purposes, with "Non-Commercial" defined as research or evaluation only. Anything else requires a separate commercial licence requested by email. Redistribution carries conditions: pass the agreement along, mark modified files, keep the attribution notice naming Hangzhou Tongyi Laboratory Technology Co., Ltd, display "Built with Qwen" if you use outputs to train another model you distribute, and do not use "Qwen" as a derivative's primary name. Governing law is Chinese, with jurisdiction in Hangzhou.

• Krea 2 — the Krea 2 Community License. Commercial use is permitted, with a revenue condition: you and your affiliates must stay under one million US dollars in total company-wide annual revenue, measured on a trailing twelve-month basis across all sources. Above that line, an enterprise licence is required from Krea before any further commercial use. Redistribution requires you to pass the agreement on, to prefix the model name — "Krea 2 [Your Model Name]" — and to keep the attribution notice. Krea states it claims no ownership of outputs, and you own your derivatives subject to Krea's underlying rights. Delaware law governs. Access is gated behind an agreement and an acceptable-use policy, and both checkpoints carry the licence tag krea-2-community-license on Hugging Face.

Read side by side, the shapes are different rather than one being stricter. Qwen-Image 2.1 is a hard no on commercial use with an escape hatch you have to negotiate. Krea 2 is a yes with a ceiling, and the ceiling is a specific, checkable dollar figure rather than a judgement call. For a small studio or an independent developer, that difference is the entire decision and the Elo gap is noise. For anyone above the threshold, both models end up in the same place: an email to the vendor before you ship.

What each one costs you to run

Neither number is published by the vendor, so the honest answer is that the comparison has to be built from settings rather than from a quote.

• Sampling — Krea 2 RAW runs at 52 steps with a guidance scale of 3.5, Krea 2 Turbo at 8 steps with guidance disabled; Qwen-Image 2.1 defaults to 40 steps.

• Resolution — Krea documents its base checkpoint up to 1K and its Turbo checkpoint across roughly 1K to 2K; Qwen-Image 2.1 defaults to 2048×2048 with seven published presets in the same range.

• What you download — the Qwen-Image 2.1 Diffusers repository measures 33.1 GB, most of it the Qwen3-VL 8B text encoder. Krea has not published a corresponding size for its checkpoints.

• What runs it — Qwen-Image 2.1 shipped with documented support in Diffusers, ComfyUI, vLLM-Omni, SGLang and LightX2V, plus AMD and FlagOS paths. Krea 2 documents ComfyUI and SGLang in its repository, with an inference script and a handful of training toolkits for LoRA work.

Krea published download counts of 84,354 for Krea-2-Raw and 87,771 for Krea-2-Turbo in the last month, as read on 13 October 2026 — figures from the hosting platform, not from Krea, and not a quality signal.

The one independent test that exists, and what it does not prove

There is a side-by-side worth knowing about, and it needs its limits stated. A Japanese-language write-up published 21 September 2026 ran 28 prompts through both models with identical prompts and seeds on an RTX 5090 with 32 GB, in ComfyUI v0.37.0, producing 76 comparison images. Its tally: Krea 2 won 10, Qwen-Image 2.1 won 7, 9 tied, and 2 failed for both. The author's summary of the pattern was that Qwen-Image 2.1 handles flat symbolic content well while Krea 2 handles the structure and physics of three-dimensional objects better, and noted Krea 2's result came at 8 steps against Qwen-Image 2.1's 25.

That is one operator, one seed per prompt, one visual judge, and one month before this article. It is not a benchmark and it is not a ranking. What it is good for is calibrating expectations: the two models are close enough that a person running real prompts will see wins on both sides, which is consistent with the eleven-point Elo gap on the live board.

Where OrcaRouter fits, and where it does not

Neither of these models is on OrcaRouter, and it is worth saying so before anything else: we checked our own live model list on 13 October 2026 and no Qwen-Image model and no Krea model is on it. Both comparisons above are about weights you run yourself or hosted endpoints you buy from each vendor directly. If either one gets routed through us, the routing comes after the licensing decision, not instead of it.

The reason that distinction matters here is the shape of the problem. Qwen-Image 2.1 and Krea 2 are not the last image models either lab will release, and the licence terms on this generation are not a guide to the next one — Krea has already moved from a subscription product to a community-licensed open release, and the company moved from Apache-2.0 on its earlier Qwen-Image checkpoints to a research-only licence on this one. A team that standardises its image pipeline on one vendor's SDK ends up renegotiating that pipeline every time the licence moves. 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 for composing several models into one call. The relevant property is not that we host these two — we do not — but that where a model is served through us, swapping the model behind a call is a configuration change rather than an integration project, which is the thing an open-weights licence change usually forces.

Which one to pick

The rule, rather than a verdict. If you need to edit existing images — feed in a photograph or a product shot and change part of it, work with masks, produce transparent assets — Qwen-Image 2.1 is the model with the capability and Krea 2's open release is not in the conversation. If you need commercial output and you are a small team under the revenue line, Krea 2 is licensed for exactly that and Qwen-Image 2.1 is not licensed for it at all. If you are fine-tuning, Krea 2's RAW checkpoint is explicitly built to be trained on and its LoRA ecosystem assumes you will; Qwen-Image 2.1 is fine-tunable under a research-only grant that constrains what you can do with the result. If you are above the revenue threshold and want either one in production, you are writing to a vendor either way — Krea at opensource@krea.ai, the vendor at the address in the licence agreement — and the choice reduces to which conversation you would rather have.

The fact that decides it most often: the licence, not the leaderboard. The two models are eleven Elo points apart on a board with intervals wider than that, and roughly a hundred dollars a month apart as hosted services, and the only one of those differences a reader cannot work around is the one printed in the licence file.