Breeze TTS 2 is now the highest-rated open-weight model on the Artificial Analysis text-to-speech leaderboard at Elo 1215, ahead of ElevenLabs Eleven v3 at 1175, and its hosted price is $34 per 1M characters against ElevenLabs' $100 (Artificial Analysis Provider Voice Arena, verified 2026-09-11). The verdict: if you are a researcher, hobbyist or builder running your own GPU for non-commercial work, Breeze TTS 2 is the better buy. If you are a business shipping a paid product, choose ElevenLabs, because the Breeze weights carry a research and non-commercial licence and commercial use needs a separate written agreement from BreezeBlue (model licence, verified 2026-09-11).
That licence is the whole story: a model that wins on the scoreboard, then disqualifies most of the people reading it.
Pick Breeze TTS 2 if the output is for research, internal experiments, a paper, a personal project, or a demo that never bills anyone; and you have a GPU with at least 12 GB of memory. You get top-of-class open-weight quality with no per-character bill and no audio leaving your machine.
Pick ElevenLabs if you are putting voice into a product with customers. The licensing is unambiguous, the language coverage is far broader, the platform is mature, and Eleven v3 Conversational is priced at $50 per 1M characters for the low-latency agent case where Breeze's advantage narrows, per the Artificial Analysis leaderboard.
Pick a genuinely permissive open-weights alternative if you want commercial self-hosting without negotiation. That is where VoxCPM2 matters: it ships under Apache 2.0, which means the weights themselves are usable commercially. We compared it directly in VoxCPM2 vs ElevenLabs. For cloning specifically, Voicebox vs ElevenLabs covers the local route.
Two licences apply to two different artefacts.
The inference code on GitHub is Apache 2.0. You can use, modify and redistribute it commercially. The model weights on Hugging Face are covered by the BreezeBlue Research and Non-Commercial License: research and non-commercial use only, with commercial use requiring a separate written licence from BreezeBlue (licence text, verified 2026-09-11). There is no small-business carve-out and no revenue threshold below which commercial use becomes acceptable.
Practical consequences:
"Open weights" and "open source" are not synonyms. Read the LICENSE file before building a roadmap on a download.
On the Artificial Analysis Provider Voice Arena, all figures verified 2026-09-11:
| Model | Elo | Rank | Price per 1M chars |
|---|---|---|---|
| Cartesia Sonic 3.6 | 1282 | 1 | — |
| Breeze TTS 2 | 1215 | 7 | $34 |
| Eleven v3 Conversational | 1210 | 8 | $50 |
| Eleven v3 | 1175 | 14 | $100 |
Source: Artificial Analysis TTS leaderboard. Breeze TTS 2 beats Eleven v3 on narration quality at roughly a third of the hosted price, and it is the top open-weight entry of the 16 open-weight models among the 98 listed. But it does not beat everything. Eleven v3 Conversational is within 5 Elo points of it, and Cartesia leads the board outright. Elo drifts as votes accumulate, so treat single-digit gaps as ties.
The open-weight field trails back: Fish Audio S2 Pro 1128, Step Audio EditX 1104, Voxtral TTS 1078, Magpie-Multilingual 1065 (same leaderboard).
Three features are worth the download even for evaluation purposes, per the model card and repository README:
(laugh) or (sigh) are rendered as performance rather than read aloud as words.
It is bilingual English and Chinese — 2 languages against ElevenLabs' 74 (ElevenLabs), and Breeze's hosted API covers 50 (Artificial Analysis) — which is the sharpest limitation.
The published figures: under 40 ms time-to-first-audio on the H100 fast path, a real-time factor of 0.32 (about 3.1x faster than playback), and roughly 7.7 GiB of GPU memory for eager inference, with a 12 GB GPU given as the practical minimum (model card, verified 2026-09-11).
A 12 GB consumer card is enough to run it, though the 40 ms figure applies to datacentre hardware - expect higher latency locally. The audio tokenizer is built on Alibaba's Qwen3-TTS under Apache 2.0. For the wider local stack, see choosing a local LLM and free local dictation on a Mac.
Q: Is Breeze TTS 2 open source?
A: Partly. The code is Apache 2.0; the weights use the BreezeBlue non-commercial licence, so the model is open-weight, not open source.
Q: Can I use Breeze TTS 2 in a paid product?
A: Not under the published licence. Commercial use requires a separate written licence from BreezeBlue.
Q: Is Breeze TTS 2 better than ElevenLabs?
A: On the Artificial Analysis arena it scores 1215 vs Eleven v3's 1175, but Eleven v3 Conversational (1210) is level for real-time use and ElevenLabs covers far more languages.
Q: How much does Breeze TTS 2 cost?
A: $34 per 1M characters hosted. Self-hosting replaces that fee with your own GPU cost, subject to the licence.
Q: What GPU do I need to run it locally?
A: About 7.7 GiB eager, 12 GB GPU minimum, per the model card - a mid-range card suffices for testing.
Q: Which open-weights model can I use commercially instead?
A: VoxCPM2 (Apache 2.0) permits commercial self-hosting with no separate agreement. It scores lower but removes the licensing blocker.
No corrections yet. Elo ratings on the Artificial Analysis arena are live and shift with voting; figures here were captured on 2026-09-11 and will be re-checked.
Last verified: 2026-09-11.