Breeze TTS 2
BreezeBlue's 3.5B real-time English and Chinese TTS with voice cloning, voice design and delivery directions. Non-commercial weights.
In short. Breeze TTS 2 reads text aloud in English or Chinese and starts speaking almost at once, which suits live conversation. It can copy a voice from a short clip, invent a new voice from a written description, or keep a voice while you tell it to speak slowly, sternly or cheerfully. The free weights are for research and non-commercial use only.
Summary
BreezeBlue, the trade name of Resonia, Inc., released Breeze TTS 2 on 25 August 2026. It is a text-to-speech model built for real-time conversation, and one model speaks English and Chinese. It has three ways to set the voice. Voice clone copies a speaker from a clean reference clip and its exact transcript. Voice design creates a voice from a written description with no audio. Voice direction keeps a reference speaker's identity while a written instruction sets tone, emotion and pace. Vocal events can be written into the text, such as (laugh), (sigh) and (clears throat) in English or [笑] and [叹气] in Chinese. The checkpoint holds about 3.5 billion parameters in 7.7 GB, and its codec is based on the Qwen3-TTS audio tokenizer.
BreezeBlue's own figures are under 40 milliseconds to the first audio and a real-time factor of 0.32, which is about 3.1 times faster than real time, on an NVIDIA H100 with the warmed-up fast path. The card says the default eager mode needs about 7.7 GiB of GPU memory and recommends a 12 GB GPU, while the fast path needs about 14.4 GiB and a 24 GB card. It requires Linux and a CUDA GPU. The code includes a command-line script and a single-concurrency streaming API that returns 24 kHz PCM audio. The card also says Breeze TTS 2 ranks first among open-weight models on the Artificial Analysis TTS leaderboard and ahead of proprietary systems, a claim shown as a chart that we could not verify.
The weights are free only for research and non-commercial use. The BreezeBlue license grants no commercial rights, has no revenue or creator exception and applies to audio you generate yourself with the open weights. A paid subscription to BreezeBlue's hosted platform allows commercial use of what you generate there, but not of the open weights. The source code is Apache 2.0. In the 30 days to 11 November 2026, Hugging Face counted about 26,000 downloads and the repository has nearly 1,000 likes.
Good for
- Real-time voice agents and assistants in English or Chinese, with streaming audio
- Designing a new voice from a text description, or steering the tone and pace of a cloned one
- Research and non-commercial projects, including voice-interface prototypes
- Expressive lines with laughs, sighs and coughs written into the text
Watch out for
- Non-commercial only, even for audio you generate yourself with the open weights, and commercial use needs a separate license or BreezeBlue's paid hosted service
- The open model covers English and Chinese only
- It needs Linux and a CUDA GPU, a 12 GB card for the default mode and a 24 GB card for the fast path, and the latency figures come from an H100
- Cloning a real person's voice needs their consent and the rights to the reference audio, and the license forbids removing watermarks or provenance signals
- The leaderboard rank and latency numbers are BreezeBlue's own
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Breeze TTS 2. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Breeze TTS 2 BF16 | 3.5B | 7.7 GB | 12 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more. The developer states a minimum of 12 GB of GPU memory.
The license in plain words
The license for the Breeze TTS 2 weights, version 1.1 of 1 September 2026, issued by Resonia, Inc., which does business as BreezeBlue. It is free for research and non-commercial use and grants no commercial rights at all, with no revenue-threshold or creator exception. Audio you make with the open weights carries the same limits.
- Attribution or model name must be shown
- Derived models must carry the original name
- Acceptable-use policy applies
More about BreezeBlue Non-Commercial
See also
Frequently asked questions
- What is Breeze TTS 2?
- Breeze TTS 2 reads text aloud in English or Chinese and starts speaking almost at once, which suits live conversation. It can copy a voice from a short clip, invent a new voice from a written description, or keep a voice while you tell it to speak slowly, sternly or cheerfully. The free weights are for research and non-commercial use only.
- Can I use Breeze TTS 2 commercially?
- Not without separate permission. Breeze TTS 2 is released under the BreezeBlue Non-Commercial, which allows research and personal use only.
- Is Breeze TTS 2 open source?
- Breeze TTS 2 is open-weight, not open source in the strict sense. You can download the model files, but the BreezeBlue Non-Commercial adds conditions that the Open Source Initiative definition does not allow.
- Can I run Breeze TTS 2 on my own computer?
- Yes. The smallest variant, Breeze TTS 2, needs about 12 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Breeze TTS 2 released?
- Breeze TTS 2 was released by BreezeBlue on August 25, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Breeze TTS 2 model card (Hugging Face) huggingface.co
- breeze-tts inference code and README (GitHub) github.com
- Breeze TTS 2 announcement (BreezeBlue) breezeblue.ai
- BreezeBlue Research and Non-Commercial License Agreement v1.1 (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| IndexTTS-2.5 Bilibili | 800M | – | Bilibili Model Use (Conditional) | 6 GB Laptop or phone | |
| Qwen3-TTS Alibaba Qwen | 910M – 1.92B | – | Apache 2.0 (Permissive) | 3.68 GB Laptop or phone | |
| VoxCPM2 OpenBMB | 2.29B | – | Apache 2.0 (Permissive) | 8 GB Laptop or phone | |
| OmniVoice Xiaomi | 613M | – | CC BY-NC (Non-commercial) | 3.75 GB Laptop or phone |