Build1 publisher2 min readPublished
Reselling Qwen Image 2.1 now requires a separate licence from Alibaba
The 7B open-weight model fits on an RTX 3090 and scores 60.2 on the benchmark its own developers built, against 59.82 for Google's Nano Banana 2.0. The licence shipping with it forbids commercial resale.
The Engineer · Build desk

What happened
- Alibaba Cloud released Qwen Image 2.1, an open-weight image model whose generation component has 7 billion parameters and which Tom's Hardware reports runs on cards as old as the RTX 3090.
- Preliminary AI Arena testing put it at 1228, below Nano Banana 2 at 1260 and Muse Image at 1276, with the GPT Sunburst entry top of the list at 1423.
- Reddit user cgs019283 said the model generates 1MP images in around 25 seconds on RTX 5070 and 5080 cards, and that editing was the slowest function in their testing.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision Anyone planning to charge other people for access to these weights has to negotiate a licence with the developer before ordering GPUs.
- contradiction Which scoreboard you believe decides whether this is a like-for-like replacement for Nano Banana 2.0 or a cheaper approximation of it: the first-party board puts Qwen ahead, AI Arena puts it 32 points behind.
- capability Transparent-background output plus multi-reference compositing on one consumer card puts asset work inside a local pipeline, where the customer photos being composited never leave the building.
- constraint Teams that cleared the earlier Qwen image weights with counsel are now working under different terms, so they need a fresh legal review before they download this one.
The 7B figure describes the checkpoint, not the card. None of the user reports Tom's Hardware collected state a precision, a step count, or a batch size, so the timings are single runs on unknown settings [11][12]. One of them generates 2K images in about 50 seconds on an RTX 3060 with 64GB of memory [13]. That much host RAM behind a mid-range card usually means part of the model is moving over PCIe during the run, and an offloaded run tells you little about what the same card does when the weights fit in VRAM.
Pick a scoreboard and you get a different answer to the question a swap decision turns on. On the developers' own Qwen Image Benchmark the model finishes 0.38 points ahead of Nano Banana 2.0 [1] and 6.8 points behind GPT Image 2.5 Sunburst [2]. On AI Arena, which Tom's Hardware describes as less favourable conditions, it sits 32 points below Nano Banana 2 [3] and 195 below the GPT entry at the top of the list [4]. A 0.38-point win on a benchmark named after the winner is a tie. For either score to predict quality on your workload, your prompt mix and your grading would have to resemble whatever the benchmark used, and the article does not describe the prompt set.
Editing is where the release gets interesting. It takes up to 10 reference images, and the example the Qwen team cites is a photo of a person plus images of clothing items composited into one image of that person wearing them [6]. Transparency is native, so backgrounds come out empty for print-on-demand work like stickers [5]. AI Arena currently ranks it first among open-weight models for editing [9].
That is exactly the pipeline the licence clause lands on. Tom's Hardware reports that this version, unlike its predecessor, explicitly forbids commercial resale of the model and directs anyone who wants to do that to obtain a separate licence from the developer [4]. Resale of the model is a narrower thing than commercial use of its output. Whether a hosted endpoint counts as resale depends on wording the article describes but does not quote. The article does not report the terms of the separate licence.
Only one model in the first-party comparison has fewer parameters, Meituan's LongCat-Image at 6B [10], and Tom's Hardware notes that most of the alternatives are several times the size or closed-weight entirely [1]. If you generate assets for your own products, the clause as reported does not obviously reach you. If you plan to charge other people for access to the weights, you need that separate licence first.
What to watch
- Whether Alibaba Cloud publishes terms or a price for the separate commercial licence.
- An independent run of the Qwen Image Benchmark prompt set by someone other than the developers.
- Whether the AI Arena open-weight editing lead holds as more votes accumulate.