What is Breeze TTS 2?
Breeze TTS 2 is a text-to-speech model from BreezeBlue (the copyright line names Resonia, Inc.), released as open weights on 25 August 2026. The checkpoint holds about 3.47 billion parameters on a Qwen3 backbone with a Mimi-style audio codec, and the open release speaks English and Chinese. It clones a voice from a reference clip and its exact transcript, designs a new voice from a written description, and steers a cloned voice's tone and pace from an instruction. Inline tags such as (laugh) and (sigh) add vocal events.
The inference code on GitHub is Apache 2.0, but the weights carry the BreezeBlue Research and Non-Commercial License. Commercial use of self-hosted output needs a separate written licence; paid subscribers to BreezeBlue's hosted service can use that output commercially. The Hugging Face page records about 15,500 downloads in the past month and 831 likes.
How much does Breeze TTS 2 cost?
| Plan | Price | What you get |
|---|---|---|
| Self-hosted weights | Free | Research and non-commercial use only |
| Free (hosted) | Free | 400 daily credits; FAQ says no commercial use |
| Starter | $9/month | 30,000 credits a month |
| Creator | $19/month | 63,333 credits a month |
| Pro | $99/month | 330,000 credits a month |
| Scale | $299/month | 996,667 credits a month |
| Business | $999/month | 3,330,000 credits a month |
| Enterprise | By quote | On-premises and data residency options |
Weights free for research and non-commercial use only. Hosted plans: Free (400 daily credits, 10 characters each), Starter $9/month, Creator $19/month, Pro $99/month, Scale $299/month, Business $999/month; about $30 per million characters, or roughly $25 billed yearly. Enterprise by quote.
Is Breeze TTS 2 worth it?
Breeze TTS 2 is one of the two best-sounding open-weights voice models, and almost nobody can legally put it into a product. On Artificial Analysis's blind-listening arena it scores 1,233 Elo, 11th of 98 models and second of 18 open-weights ones, 16 points behind Fish Audio Drama 3 Preview and well short of Eleven v4 Turbo's 1,338. BreezeBlue's launch claim of first place among open models came from the 26 August snapshot, at 1,215 Elo; the leaderboard has since moved.
The 'free ElevenLabs alternative' label does not survive the licence. The weights permit research, personal and educational use, and limited evaluation inside a company. Production use, serving it through an API, client work, monetised content and internal operations beyond evaluation are all named as commercial, and the licence states there is no revenue-threshold or creator exception. Outputs are covered too, and may not be used to train other speech models. Commercial rights come only from a written deal or from BreezeBlue's own hosted plans, where the open-weights limits do not apply.
The engineering figures are BreezeBlue's own and were measured on an H100: under 40 ms to first audio and a real-time factor of 0.32 on the warmed-up fast path. Eager inference needs about 7.7 GiB of GPU memory, so a 12 GB card is the floor; the fast path needs 14.4 GiB. The bundled streaming server handles one request at a time and returns raw 24 kHz PCM, so a multi-user deployment means writing your own batching. The audio.cpp project ships BF16, Q8 and Q4_0 GGUF builds with streaming, which lowers the hardware bar for hobby use.
The language story also differs by channel. The open release supports English and Chinese only, while BreezeBlue's hosted service advertises 51 languages, and the company does not say whether the hosted model is the same checkpoint. On price, the hosted endpoint is not cheap: Artificial Analysis lists it at $34 per million characters against $15 for Fish Audio, and measures 45 characters per second of throughput against Fish Audio S2 Pro's 102.
Where it earns its place
A speech researcher benchmarking open TTS models for a paper
Compare voice cloning and instruction-following against Qwen3-TTS and Fish Audio
Download the checkpoint, run the voice-design and voice-direction examples with --cfg-scale 4, and score the output against BreezeBlue's published benchmark suites. Research use and publishing results are expressly permitted under the licence.
A hobbyist audiobook maker on a 12 GB consumer GPU
Narrate personal reading copies in a designed voice
Run eager inference at about 7.7 GiB, or a Q8 GGUF through audio.cpp. Fine for a private library; selling or posting the narrations for ad revenue counts as commercial use and is not allowed.
A product manager at a 20-person voice-app startup
Decide whether Breeze can replace an ElevenLabs bill
Self-hosting is ruled out by the licence, so the real comparison is BreezeBlue's hosted plans, about $30 per million characters, against your current provider. A 10M-character month is roughly $300 at list price; $999 Business covers about 33M characters.
Before you commit
- Setup
- Clone breezeblue-ai/breeze-tts, pip install the requirements on Linux with Python 3.10 or newer and an NVIDIA GPU, then point infer.py at the checkpoint. A Docker build targets H100 by default and A100 with a flag.
- Integrations
- Python CLI and a streaming HTTP endpoint at /v1/audio/speech that returns 24 kHz 16-bit PCM; GGUF builds run in audio.cpp; BreezeBlue's hosted API shares credits with its Studio app.
- Data and privacy
- Self-hosted, text and reference audio stay on your machine. The weights licence forbids commercial use of the model and its outputs, forbids training other speech models on outputs, and requires a NOTICE file and attribution on any redistribution.
- Cost at scale
- Self-hosting needs a 12 GB GPU for eager inference or 24 GB for the fast path, but production use requires a commercial licence on request. Hosted pricing is about $30 per million characters, roughly $25 on yearly billing; Artificial Analysis lists $34.
- Limits
- English and Chinese only in the open weights. The reference server is single-concurrency, the fast path adds cold-start time for CUDA graph capture, and cloning needs a clean reference clip with an exact transcript.
What's good
- 1,233 Elo on Artificial Analysis, second among 18 open-weights TTS models
- Voice cloning, voice design from text and directed cloning in one model
- Runs eagerly on a 12 GB GPU; GGUF builds exist in audio.cpp
- Apache 2.0 inference code with a streaming HTTP endpoint
What's not
- Weights and self-hosted outputs barred from any commercial use
- Open release covers English and Chinese only
- Latency figures are vendor-measured on an H100
- Reference streaming server serves one request at a time
Best for
Researchers, hobbyists and teams evaluating open voice quality who will either stay non-commercial or buy BreezeBlue's hosted plan for production.
Breeze TTS 2 alternatives
Breeze TTS 2: common questions
- Is Breeze TTS 2 worth it?
- Breeze TTS 2 sounds like a frontier voice model, but its licence makes it a demo you can run at home, not an ElevenLabs replacement you can ship. We rate it 3.8 out of 5.
- How much does Breeze TTS 2 cost?
- Weights free for research and non-commercial use only. Hosted plans: Free (400 daily credits, 10 characters each), Starter $9/month, Creator $19/month, Pro $99/month, Scale $299/month, Business $999/month; about $30 per million characters, or roughly $25 billed yearly. Enterprise by quote.
- What are the downsides of Breeze TTS 2?
- Weights and self-hosted outputs barred from any commercial use; open release covers English and Chinese only; latency figures are vendor-measured on an H100; reference streaming server serves one request at a time.
- Who is Breeze TTS 2 best for?
- Researchers, hobbyists and teams evaluating open voice quality who will either stay non-commercial or buy BreezeBlue's hosted plan for production.
- How much does Breeze TTS 2 cost for a team or in production?
- Self-hosting needs a 12 GB GPU for eager inference or 24 GB for the fast path, but production use requires a commercial licence on request. Hosted pricing is about $30 per million characters, roughly $25 on yearly billing; Artificial Analysis lists $34.
- Is Breeze TTS 2 safe for company data?
- Self-hosted, text and reference audio stay on your machine. The weights licence forbids commercial use of the model and its outputs, forbids training other speech models on outputs, and requires a NOTICE file and attribution on any redistribution.
- What are the best Breeze TTS 2 alternatives?
- The closest alternatives we have reviewed are Qwen3-TTS, OmniVoice, Chatterbox, Kokoro-82M and Eleven v4 and v4 Turbo.
Answers checked Oct 26, 2026.