Qwen-Image-2.1 vs LLaDA-Image: The Licence Claim Changed, and Only One of Them Says So

When LLaDA-Image shipped from inclusionAI, Ant Group's research lab, on 4 September 2026, the reading of its repositories was straightforward: no licence declared, therefore all rights reserved. That reading is now out of date. All four of its Hugging Face model cards — LLaDA-Image, LLaDA-Image-Turbo, LLaDA-Image-FP8 and LLaDA-Image-Turbo-FP8 — declare apache-2.0 in their frontmatter today, and it is the first line of the card. Not one of those four repositories contains a LICENSE file, and the project's GitHub repository created on 31 August 2026 has no licence file either. Set that against Qwen-Image-2.1, released by Aliba​ba's Qwe​n team with open weights on 20 September 2026 under the Qwe​n Research License Agreement — a real document, dated, non-commercial only, and the reason its own repositories tag the licence as other. Two open-weights image models, sixteen days apart, and the one with the weaker-sounding paperwork is the one that declared a permissive licence. The catch is entirely in how it was declared.

This is a comparison of paper, not pixels. Neither model has an independent image score — LLaDA-Image in any of its four variants appears on neither Artificial Analysis image board, and Qwen-Image-2.1's two rows there carry a No API available note. What you can actually decide today is what you are allowed to do, and that answer moved for one of these two models without an announcement.

What the two licence surfaces actually say

Read the primary sources rather than the summaries, because the gap between them is the whole story.

• Qwen-Image-2.1 — released under the Qwe​n Research License Agreement dated 20 September 2026, granted "FOR NON-COMMERCIAL PURPOSES ONLY", with commercial use requiring a separate licence requested from the vendor. Aliba​ba's Hugging Face repositories for the model report the licence field as other, which is the accurate tag: it is not an OSI-approved licence and the model is not open source in that sense. It remains downloadable, runnable and benchmarkable.

• LLaDA-Image — the Hugging Face model card declares license: apache-2.0 in its YAML frontmatter. Apache 2.0 is an OSI licence that permits commercial use. If the frontmatter is the licence, this is the more permissive of the two by a wide margin.

• LLaDA-Image, the rest of the paper trail — there is no LICENSE file in the repository tree. The card lists a README, a model_index.json, a .gitattributes and the weight directories, and nothing else. The project's GitHub repository has no licence file either; GitHub's own repository API reports no licence for it. Apache 2.0's own text requires that a copy of the licence accompany the distribution, and there is no copy in either place.

So the honest description of LLaDA-Image's licence position is not "no licence" and not "Apache 2.0". It is: the card asserts Apache 2.0 and the repository does not ship the licence text. Those two facts pull in the same direction but they are not the same fact, and only one of them is what a procurement reviewer will look for.

Why the card reads the way it does

The card's own frontmatter is the giveaway. The model family's headline is LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes, and two of its five stated highlights are about openness: image-only pre-training to establish the visual prior, and a unified diffusion model where both the backbone and the DiT are diffusion models trained in one framework. The word "recipe" is doing deliberate work. The card is pitching a reproducible pipeline, not just a checkpoint.

Then read the open-source plan at the bottom of the same card:

• Inference code and model weights — ticked.

• Training code — not ticked, marked coming soon.

• A separate arXiv report, numbered 2609.03796 on the card, carries the method.

That is a coherent position and not a dishonest one: weights and inference code now, training code later, method in a paper. It is also the context in which an Apache 2.0 frontmatter line makes sense as intent — this is a project that wants to be read as fully open. The gap between the intent and the missing licence file is exactly the kind of thing a card's frontmatter cannot express.

What the numbers say, and who produced them

The one quantitative claim attached to LLaDA-Image is a benchmark score, and it needs its provenance stated plainly. The card reports "SOTA on Qwen-Image-Bench" with overall scores of 53.53 in English and 53.38 in Chinese. Those are inclusionAI's own figures, measured on a benchmark named after another lab's model, and they have not been reproduced by anyone outside the lab. The Base model reaches them at 50 sampling steps; the Turbo variant is distilled with Twin-DMD to run in 2 to 4 steps. Neither figure has a vote count behind it and neither appears on an independent board, because LLaDA-Image is not on one.

Qwen-Image-2.1 is on both Artificial Analysis boards — 1,034 Elo at rank 18 of 166 on text-to-image and 1,074 Elo at rank 18 of 97 on editing, from 5,286 and 5,536 blind comparisons respectively. Its rows on both boards carry the Open Weights marker next to the No API available note. That is independent measurement of a self-hosted model, which is a harder thing to get than it sounds, and it is the one axis where these two models are not in the same category at all.

The decision this actually changes

If you were relying on the earlier reading — that LLaDA-Image declares no licence, so treat it as all rights reserved — the position has improved, and the improvement is real enough to act on for evaluation and internal work. Keep the qualification attached: a licence named only in YAML, with no licence text in the repository, is weaker evidence than a licence file sitting next to the weights. If LLaDA-Image is going into a product, that is a question for Ant Group's research lab, not something to resolve by pointing at a metadata tag.

Qwen-Image-2.1 is the reverse trade: the paperwork is unambiguous and the answer is no. A dated research licence that forbids commercial use is easier to plan around than an unclear permissive one, because there is nothing to interpret. It is also a change from the earlier Qwen-Image line, which shipped under Apache 2.0, so teams that built on the previous generation cannot carry the assumption forward.

Neither of these models is served on OrcaRouter, and neither will be until the licence question resolves — a research-only licence and a licence file that is not in the repository are both reasons to keep the hosted path separate from the experiment. The image models we do front are Google's Imagen tiers and Gemini image previews, xAI's Grok Imagine image endpoint and OpenAI's GPT-Image family, all behind one OpenAI-compatible endpoint with provider list price passed through at zero markup, automatic failover across providers, and a routing DSL for composing several models into a single call. Run the downloadable model for the work the licence covers; call the hosted one for anything that reaches a customer.

What to watch

Two things turn this from a paperwork comparison into a model comparison, and both are checkable without waiting for anyone's announcement.

The first is whether a LICENSE file appears in the LLaDA-Image repositories. A licence named in frontmatter and shipped as text is what Apache 2.0 actually requires, and until that file exists the declaration is the only evidence. The second is whether the training code ships. It is the last unticked box on the card, and it is the difference between weights that happen to be downloadable and the "fully open training recipes" the title promises.

Until then, the fair summary is the one the sources support: LLaDA-Image now claims the more permissive licence and has not yet documented it, Qwen-Image-2.1 documents a licence that forbids what most teams want to do, and only one of the two has been measured by anyone other than its own lab.