Open weights aren't open for business: Qwen-Image-2.1-Turbo's licence

Kandinsky Lab open-sourced Kandinsky 6.0 Video on 6 October, and Qwen released Qwen-Image-2.1-Turbo three days later. Only one is licensed for paid work as released. Kandinsky 6.0 Video carries the MIT licence, which allows commercial use, though its Hugging Face files open only after you accept the repository's access conditions. Qwen-Image-2.1-Turbo carries the Qwen Research License, which allows research and evaluation and nothing else.

That is the whole lesson in one pair of releases. "Open weights" tells you where the file is, not what you may do with it. A checkpoint you can download and a licence to sell what it makes are two separate things.

Qwen-Image-2.1-Turbo: free to try, not to sell

Turbo is a faster checkpoint of Qwen-Image-2.1 that generates and edits images in eight sampling steps, on the same 7B-parameter image architecture as the full model. It loads in Diffusers with its sampling schedule built in, and uses the parent model's presets, from 2048 × 2048 square to 2752 × 1536 for 16:9.

The licence file grants the right to use, copy, modify and distribute the model for non-commercial purposes only. It defines non-commercial as "for research or evaluation purposes only." Commercial use needs a separate licence, which you request by email at model-business@notice.qwencloud.com.

The rest of the agreement matters for anyone who builds on the model:

• If you use the model or its outputs to train or improve an AI model that you distribute or make available, you must display "Built with Qwen" or "Improved using Qwen" in its product documentation.
• If you breach the terms, Qwen may end the licence, and you must delete the model and stop using it.
• Chinese law governs the agreement, disputes go to the People's Courts in Hangzhou, and you indemnify Qwen against third-party claims arising from your use.

The strongest objection is that the licence covers the model, and a picture is not the model. The document gives that reading little to stand on: generating an image is using the materials, and that use is what the grant limits. The licence also names outputs in several places, including the training rule and the warranty disclaimer, and no clause allows commercial use of them.

There is a second route. The repository says Pro and Turbo APIs went live the same day, and Alibaba Cloud Model Studio's release list logs qwen-image-2.1-turbo on 2026-10-09. The research licence treats arrangements it does not state as separate from its own terms. A studio that wants Turbo for client work can therefore request a commercial licence or read the API's own terms. It should not rely on the research licence for either.

Kandinsky 6.0 Video: MIT, with a gate to read

The family comes in two sizes, Lite at 3B parameters and Pro at 29B. Both render five-second clips with matching 44 kHz sound, lip-sync included, from a text prompt or a starting image. A separate super-resolution model raises the output to Full HD, 1920 × 1080.

The lab's own timing table shows the turnaround cost. A non-distilled Pro clip at Full HD takes about 400 seconds on an H100 and about 21 minutes on an RTX 4090, and the quick start asks for an NVIDIA GPU.

The MIT licence lets anyone use, copy, modify, publish, distribute, sublicense and sell copies, free of charge. It has one condition: the copyright and permission notice must go with all copies or substantial portions. It gives no warranty. For a studio, that means you can render a client's clip or ship a tool built on the model, as long as the notice travels with any copy of the model you pass on.

Two things still need checking. The Hugging Face repository is gated: you share your account details and accept conditions before the files download, and those conditions sit behind the login. The Pro pipeline also uses a Qwen2.5-VL text encoder and a CLIP text model. A licence tag describes a repository, so confirm the licence of every component you ship.

What to check in any open model's licence

• The grant and its definitions. Qwen's grant covers non-commercial use only, a separate clause bars commercial use, and a definition narrows non-commercial to research or evaluation.
• Outputs. Check whether the licence mentions what you make, and what it asks of you if you train on those outputs.
• Gates and bundled parts. Read the access conditions and the licence of each component.
• The exits. Check what ends the licence and which courts hear disputes.
• The route. Downloaded weights and a hosted API can sit under different terms.

What to watch

The Qwen licence file carries a release date, 20 September 2026, the day Qwen-Image-2.1 itself shipped. A new date on that file means new terms, so check it before each job. On Kandinsky, watch the gate conditions on the Pro repository, and check the licence tag on any other checkpoint you pull: the Pro page and the GitHub repository both list MIT.

Frequently asked questions

Can I use Qwen-Image-2.1-Turbo commercially?

Not under the Qwen Research License that comes with the weights, which allows research or evaluation only. For commercial use, the licence says to request a separate licence from Qwen at model-business@notice.qwencloud.com.

Is Kandinsky 6.0 Video free for commercial use?

It is released under MIT, which permits commercial use, including selling copies, as long as the copyright and permission notice stays with copies. The Hugging Face files sit behind an access gate whose conditions you accept first.

What does "open weights" mean?

It means the trained model files can be downloaded. What you may do with them is set by the licence, which can range from MIT to research-only.

How long are Kandinsky 6.0 Video clips?

Five seconds, with synchronized 44 kHz audio, and a separate model can upscale them to 1920 × 1080.