Request: More permissive license for Qwen-Image-2.1
Hi! Thanks for releasing Qwen-Image-2.1! ^^/
I noticed that the new Qwen Research License restricts the model to non-commercial research/evaluation use. I think this change is, unfortunately, a step backwards for the community.
Would you consider switching back to a more permissive license, such as Apache-2.0, like the previous Qwen-Image & Qwen-Image-Edit releases?
If a fully permissive license is not possible, a common alternative in AI is a commercial-friendly "community" license. For example, Meta's Llama Community License allows commercial use for most users, while very large-scale entities must request a separate license.
I think allowing commercial use is quite important for adoption. Many developers, artists, open-source projects, and small companies may want to experiment with the model today while still keeping the possibility of using their work commercially in the future.
With the current research-only license, Qwen-Image-2.1 unfortunately becomes much less useful for that kind of work, and personally I would rather continue using the previous Apache-2.0 Qwen-Image models.
If possible, an open license like Apache 2.0 or MIT could make adoption, collaboration, and contributions much easier.
In any case, thank you for continuing to release the models and weights publicly. I really like the Qwen models, and I hope this is something that could be reconsidered :)
I agree about the Apache license, but not about the Meta license.
For example, I am doing ebook to audiobook conversions with Qwen TTS, which is under the Apache license. I am using books that are in the public domain (Standard Ebooks, Project Gutenberg) and books under CC by SA (e.g. wikipedia, wikisource). I would like to release the audiobooks in the public domain or under the CC by SA licenses. But the Meta license would kill that, because I would have to encumber the audiobooks under that license. See what I'm talking about here: https://recito.org
I think the situation may be a bit less restrictive than it first appears.
The Meta Llama Community License does not require generated outputs to be licensed under the Llama license. The redistribution requirements apply to the Llama model/materials and their derivatives, while outputs are treated separately.
There are still some restrictions on outputs (for example, Llama 3 outputs cannot be used to improve another LLM), but the license does not appear to require the output itself to inherit the Llama license.
So in your audiobook example, a Llama-style model license would not by itself prevent you from releasing the generated audiobook under CC0 or CC BY-SA, to the extent that you have the necessary rights to do so. The license of the source book would still need to be respected separately, of course.
That is one of the reasons I think community licenses can be an interesting middle ground: they can restrict certain uses of the model without necessarily imposing those same licensing terms on everything generated with it.
There are a few different approaches already in use:
- Meta Llama 3 Community License: commercial use for most users, with a separate license required above 700M monthly active users.
- FLUX.1 [dev] Non-Commercial License: the model is non-commercial, but generated outputs may be used commercially.
- Stability AI Community License: free commercial use for individuals and organizations under USD $1M in annual revenue.
That said, I'd absolutely prefer Apache 2.0 ^_^
The Meta Llama Community License does not require generated outputs to be licensed under the Llama license. The redistribution requirements apply to the Llama model/materials and their derivatives, while outputs are treated separately.
There are still some restrictions on outputs (for example, Llama 3 outputs cannot be used to improve another LLM), but the license does not appear to require the output itself to inherit the Llama license.
So in your audiobook example, a Llama-style model license would not by itself prevent you from releasing the generated audiobook under CC0 or CC BY-SA, to the extent that you have the necessary rights to do so. The license of the source book would still need to be respected separately, of course.
Yes, it WOULD prevent me from releasing under CC0 or CC by SA. Because if I release under CC0/or CC by SA, it means someone could use the output to improve another LLM, violating the Llama license. I would have to encumber my files with the rule "cannot be used to improve another LLM", which would mean the files could NOT be CC0 nor CC by SA, because those licenses are incompatible with these restrictions.
Again, yet another reason why these ad-hoc licenses are bad.
I think Flux or revenue cap to request a new license is ideal for them. I think they stopped mostly because they didn’t make any money off it to justify to execs.
I think Flux or revenue cap to request a new license is ideal for them. I think they stopped mostly because they didn’t make any money off it to justify to execs.
They (Alibaba) would make more money if their API access wasn't such a labyrinth of pages just to use it. They should make it simple like Openrouter, Vercel, Requesty, etc. I stopped using it because it started blocking me because my "free usage" was used up. But I didn't even want to use it for free! Somehow they autoamatically signed me up for that. I want to pay per API hit. But trying to get it going again wasn't clear (!). I had my paid usage stop during production for just dumb reasons, so I just use Vercel 90%+ of the time now...