Impressive technology, but its licensing is extremely restrictive.
Breeze TTS 2 looks like impressive technology, but its licensing is extremely restrictive. It’s important to understand that the restrictions don’t apply only to the model itself as you can’t even use the generated audio commercially. The model and its output are limited to personal or research use.
That makes it difficult to evaluate for startups, creators, or businesses hoping to build real products around it. Hopefully, the team considers offering a practical commercial license in the future, because the technology has a lot of potential.
We offer generous free credits for evaluation, and our paid plans (with commercial license) are also very competitive compared to other commercial products.
Thank you for the response. However, this seems to reinforce the concern raised in this thread (on reddit https://www.reddit.com/r/LocalLLaMA/comments/1w1002h/breezetts2_initial_impressions_genuinely_frontier/ ). The Hugging Face release feels less like a model intended to support the open-source community and more like an advertisement for the BreezeBlue API.
There is little reason for developers to invest time and energy evaluating, integrating, or improving the model if neither the model nor its generated output can be used as a project grows beyond personal or research use.
As I said, Breeze TTS 2 sounds great, but practically speaking, it is only useful commercially if you use the paid API. A more permissive license, or at least a commercial license for self-hosted use, would give the community a genuine reason to test and build around it.
Also being that this model is based on Qwen3-style backbone with 28 layers and a 2,048-dimensional hidden state which is approximately the shape of Qwen3-1.7B --- seems like you'd want to consider the goodwill of the folks @Qwen https://huggingface.co/Qwen and keep sharing the love.
@a2industries
Thanks for your feedback, and we hear your use case.
We do not have terms to announce today, so I don’t want to overpromise, but we will keep you updated with any future licensing changes.
We are doing this open-weight research and non-commercial release because:
- we believe design & control are two under-explored directions with huge value for various TTS use cases;
- we put many efforts in training a good model along these 2 directions, and we want more people to see and use it;
- we want to foster reproducible evaluation, more research along this direction, community implementations of what's possible with this model, and of course, more feedback on the model;
- based on the community feedback, we keep improving the model, and are also working on something new.
Appreciate the straight answer, and I'm not asking anyone to overpromise.
I'd just point out that the non-commercial restriction works against all three of the goals you listed. Reproducible evaluation and serious community implementations aren't hobby-scale efforts, they're weeks of engineering, and people won't spend that on something they can't eventually ship. The same applies to feedback quality: the failure modes you actually want to hear about (long-form stability, rare phonemes, name and number handling, prosody drift, latency under load, cost per hour of audio) only surface in production traffic, not in a weekend demo. A non-commercial license filters out exactly the users who would generate that data.
The usual middle ground is a threshold license: free commercial use under some revenue or user cap, paid above it. That keeps the large-scale commercial value where you want it while letting builders integrate in good faith instead of waiting on the sidelines. Even a stated intent - "commercial terms are coming, and work built during the research period won't have to be thrown away", would change how much the community invests now.
Happy to be an early tester on whatever terms land, and to report back with real usage data on design and control specifically, since that's the part I care about, but only if the license allows.
Please change the license so it's a revenue cap. I don't want to use your API. I want to embed the model in my application. I don't want to pay a self-hosting deposit either. I want to create some apps with the model embedded, and if the app generates a certain amount of revenue (Probably no less than $100k), a licensing fee is required. Doing this will remove startup barriers to entry, and everybody wins.
As I mentioned above, giving credit where it's due, this is a great sounding model for the most part. However, until the license is changed I'd recommend using openbmb/VoxCPM2 and contributing back to it's improvement as it's not far off from Breeze.