Breeze TTS 2 is a text-to-speech model with open weights released on August 25, 2026 by BreezeBlue, a Shanghai-based team (legal entity: Shanghai Weifenglan Technology Co., Ltd.). The official PyTorch inference code and weights live in the breezeblue-ai/breeze-tts repo, with weights mirrored on HuggingFace at BreezeBlue/breeze-tts-2. It is a single model with roughly 3.48 billion total parameters (checkpoint metadata reads 3,466,363,713; the HuggingFace page labels it 3B), generating natural English and Chinese from one set of weights. On Artificial Analysis' Provider Voice expressiveness leaderboard -- a blind-listening arena -- it scored 1215 Elo, first among open-weight models. At the same time, its weight license says in plain text: research and non-commercial use only. The code is free to modify; the weights are not free to sell. That contradiction is the core of this deep dive.
Let me draw the boundary first: this article only dissects Breeze TTS 2 as a single model -- how to get it (plain repo address: https://github.com/breezeblue-ai/breeze-tts ), how to run it, and how to avoid the license traps. How F5-TTS, CosyVoice, Fish Audio and other rivals compare is covered in the sibling open-source TTS models comparison; the only overlap between the two pieces is background numbers.
1. Read two leaderboards, not one headline
The Artificial Analysis Provider Voice leaderboard is a blind arena: listeners do not know which model produced the audio, and votes on expressiveness convert into Elo ratings. In the snapshot from late August 2026, Breeze TTS 2 scored 1215, first among open-weight models, leading the previous open-weight champion Fish Audio S2 Pro (1125) by 90 points. In the all-model reference including closed systems, it sits sixth overall (Elo 1215, snapshot late August 2026), between ElevenLabs v3 Conversational (1219) and Gemini 3.1 Flash TTS (1210). The "open weights No.1" claim in the official blog post comes from this expressiveness board.
But there is a second report card from the same period. The Controlled Voices board puts every model under identical conditions: eight fixed cloned voices (four American, four British) reading the same scripts. On that board, Breeze TTS 2 ranks only third among open weights: 1002 Elo, tied with Fish Audio S2 Pro and eight points behind Mistral Voxtral TTS (1010). Together the two boards give the full picture: free-form expressiveness it genuinely leads, but stability on fixed, pre-cloned voices is a separate question. If your use case is "one locked voice reading scripts every day," do not let the headline number decide.
One more qualifier: these scores, like the latency figures below, come from BreezeBlue's own reporting or from leaderboard snapshots, with no independent third-party replication yet, and leaderboards move daily -- always cite the snapshot date.
Some background on the team. BreezeBlue has 15 people, founded in June 2025, with members from Google, ByteDance, Alibaba, StepFun, and Huawei. Founder Yang Bin holds a PhD from the University of Toronto; during his doctorate he co-founded Uber ATG Research with his advisor, later co-founded Waabi, and then served as a technical partner at MiniMax. On August 25, 2026, the team announced a 6 million USD seed round (co-led by Yuanjing Capital and Redpoint China) and opened the model weights the same day.
2. The dual-track license: the most important section
Breeze TTS 2 runs two sets of rules on two different things, and mixing them up is how teams get burned.
Track one: the code. The GitHub repo's license field verifies as Apache-2.0. The inference code and benchmark scripts can be used, modified, and redistributed freely -- reading the implementation, changing the sampling logic, wiring it into your own pipeline all go through unblocked.
Track two: the weights. The model weights, any derivative models, and the output produced by self-hosting fall under the BreezeBlue Research and Non-Commercial License, restricted to research and non-commercial scenarios. To go commercial there are exactly two roads: written authorization from RESONIA, INC. (BreezeBlue's corporate entity), or the official hosted platform -- priced at 34 USD per million characters, against ElevenLabs Eleven v3 at 100 USD per million characters on the same leaderboard. The pricing message is hard to miss.
The traps hide in edge cases, so let me name the three most commonly misjudged ones. First, a free app monetized with ads counts as commercial use -- your users pay nothing, but the moment ads appear, the weight license forbids self-hosting it. Second, internal tools that support a revenue-generating business also sit in a commercial context; do not prop up a production workflow for paying customers on non-commercial weights. Third, university evaluation reproductions and paper experiments are research scenarios and are fine. There is no small-business exemption and no revenue threshold.
One-sentence summary: the Apache-2.0 code gives you the freedom to understand and modify; the non-commercial weights mean making money with them requires a separate piece of paper. Evaluation, research, prototyping -- go ahead. Commercialization -- decide first between a license deal and the hosted platform.
3. Architecture: four stages, 3.48 billion parameters
There is no arXiv paper yet; the technical details come from the repo and from developers who dissected the checkpoint. The pipeline has four stages.
Stage one, the text encoder: a T5Gemma2 structure with 26 layers, encoding the target text, instructions, and other conditions into vectors.
Stage two, the backbone: a Qwen3-style structure with 28 layers and hidden size 2048, using grouped-query attention (16 attention heads to 8 KV heads). This is the frame-level autoregressive engine, predicting codebook index 0 for every audio frame.
Stage three, the depth decoder: 12 layers with hidden size 1024, conditioned on the backbone's hidden states and the first codebook, it autoregressively expands the remaining 15 codebooks, completing 16-way residual vector quantization (RVQ).
Stage four, the audio tokenizer: a Mimi-structure codec with 24 kHz output and a frame rate of about 12.5 Hz -- one second of audio equals roughly 12.5 tokens, each carrying 16 codebooks. It loads through qwen_tts.Qwen3TTSTokenizer, which the docs say is based on Qwen3-TTS (itself Apache-2.0).
All four stages total about 3.48B parameters; the checkpoint directory is about 7.2 GB (including the audio tokenizer) with weight tensors around 6.5 GiB. The design logic is easy to read: the backbone only draws the coarse skeleton (the first codebook), while a much smaller depth decoder fills in the fine detail -- a two-stage autoregression traded for inference efficiency.
4. Four capability routes: clone, design, direct, and paralinguistics
The model routes inputs through four templates, one per use case.
Plain TTS is the baseline: text in, speech out. Voice Design creates a voice with no reference audio -- you describe the target voice in natural language (age, accent, pace, temperament), and the official guidance is to keep the instruction in the same language as the target text with a CFG scale of 4. Voice Clone copies a voice from a reference clip plus a verbatim transcript: 3 seconds is enough to start, the community norm is 3 to 10 seconds of clean speech, and the official analysis takes the first 60 seconds and keeps the best 30; references longer than 30 seconds push the output away from the original voice. One hard requirement: the ref-text transcript must match the reference audio word for word, repeats and slips of the tongue included, and the audio should be clean, non-looping, low in background noise. Voice Direction hands a cloned voice to a "director": you append natural-language instructions about emotion and delivery on top of a clone; the official August 2026 benchmark reports VDS 4.25 with speaker similarity 0.67, the best among participating models -- vendor-built, vendor-reported numbers, as is Role Fit 78.02 (first, 5.24 points clear) for Voice Design.
Paralinguistic control lives in inline tags. English uses round brackets -- (laugh), (cough), (clears throat), (sigh); Chinese uses the same four cues in square brackets, written in pinyin as [xiao], [kesou], [qing sangzi], [tanqi] -- and the model generates the corresponding laughs and coughs in place. Default sampling parameters are temperature 0.9, top_p 1.0, top_k 50, and repetition penalty 1.1; most users will not need to touch them.
On language support, precision matters: the open weights cover English and Chinese only, bilingual from a single model. The "50+ languages" figure describes the official hosted platform's capability, not what the self-hosted weights cover. Never blend the two.
5. Getting started: environment, VRAM, and real costs
The environment bar is not low. Official requirements: Python 3.10+, Linux, CUDA, with dependencies pinned to torch 2.9.1, transformers 4.57.3, and qwen-tts 0.1.1:
pip install torch==2.9.1 transformers==4.57.3 qwen-tts==0.1.1The Docker image compiles by default for H100 (Hopper, sm90); A100 users rebuild with FLASH_ATTN_CUDA_ARCHS=80. If you would rather stay inside ComfyUI, the community node pack Saganaki22/ComfyUI-Breeze-TTS-2 (Apache-2.0, 123 stars at press time) supports INT8 quantization (5 to 7.5 GiB VRAM), dialog orchestration for up to 8 speakers, and a Whisper auto-transcription node.
Two VRAM tiers -- do not flip the direction: eager mode uses about 7.7 GiB and the official recommendation is a 12 GB card at minimum; the --fast-all fast path uses about 14.4 GiB and calls for roughly 24 GB. Low latency costs extra memory: the fast tier is the hungrier one, not the leaner one. For other hardware, look to the community: one developer's real-world test on an RTX 4060 8GB with INT8 reported 4.98 GiB usage, a 7.4 GB peak, a first audio chunk at 0.39 seconds, and 2-3x real-time speed; on Mac there is a community MLX 4-bit port of about 2.3 GB that runs on a 16 GB MacBook -- unofficial, without official backing.
Latency figures also need their caliber. For the open weights, TTFA (time to first audio) is under 40 milliseconds on a warmed-up H100 fast path; RTF is 0.32, roughly 3.1x real time. The official hosted platform claims "first audio byte within 25 milliseconds" -- a hosted-service number, a different caliber from self-hosting, never to corroborate the other. On throughput, Artificial Analysis measured about 45 characters per second, slower than several competitors; budget expectations for long texts. Output is streaming 24 kHz mono 16-bit PCM.
6. Three roads, pick one
Researchers and evaluators: take the weights. The code is Apache-2.0 and the three benchmark suites (voice design, voice direction, latency) are all open, so reproduction is as easy as the official team could make it.
Building a commercial product: self-hosting is closed off by the weight license, no gray zone. Either negotiate written authorization with RESONIA, INC. or use the hosted API; at 34 USD per million characters the cost is manageable for most mid-scale applications, and the official voice library plus Voice Galaxy (15,000+ community and studio voices) saves you building a voice lineup from scratch.
Just need output fast: use the hosted platform and skip the environment, the GPU, and the license questions.
One line of positioning inside the open TTS field: like F5-TTS it belongs to the alignment-free cloning school, with expressiveness and paralinguistic control as strengths, and a 45-characters-per-second generation speed plus commercially restricted weights as costs. Whether to adopt depends on whether you want "most human-like" or "fastest and cheapest" -- the multi-model comparison piece has the full answer.
FAQ
Q1: Can I use Breeze TTS 2 commercially? A1: Split it in two. The code is Apache-2.0 -- modify and use freely. The weights, derivative models, and self-hosted output are research/non-commercial only. Two roads to commercial use: written authorization from RESONIA, INC., or the official hosted platform (34 USD per million characters). A free app monetized with ads counts as commercial, and there is no small-business exemption.
Q2: What GPU do I need?
A2: Eager mode runs on about 7.7 GiB of VRAM, so a 12 GB card minimum; the --fast-all fast path needs about 14.4 GiB, calling for around 24 GB -- the fast tier uses more VRAM, not less. The community ComfyUI node pack supports INT8 (5 to 7.5 GiB), with a developer's test running on an RTX 4060 8GB; on Mac, try the community MLX 4-bit build (about 2.3 GB, unofficial).
Q3: How do I clone a voice? A3: Prepare a reference clip plus a verbatim transcript that matches the audio exactly, repeats and all. 3 seconds is the minimum; the community norm is 3 to 10 seconds of clean, low-noise, non-looping speech. Beyond 30 seconds the output drifts from the original voice -- the system analyzes the first 60 seconds and keeps the best 30.
Q4: How good is the Chinese? A4: The open weights are a single bilingual English-Chinese model, so Chinese is a first-class citizen, with the full paralinguistic tag set for laughter, coughing, throat-clearing, and sighing. The 1215 Elo on the leaderboard comes from an English-heavy blind-listening caliber and is vendor-reported; test the Chinese with your own copy. Remember: "50+ languages" is the hosted platform's capability, not the open weights'.
Q5: How does it compare to ElevenLabs? A5: On the same Provider Voice expressiveness snapshot (late August 2026), Breeze TTS 2 sits at 1215 and ElevenLabs v3 Conversational at 1219 -- nearly level. The real difference is form: ElevenLabs is a closed hosted service, while Breeze offers both open weights and hosting, at 34 USD per million characters versus Eleven v3's 100. Leaderboards move daily; cite the snapshot date.
Related reading
- Open-Source TTS Models Comparison: the sibling piece -- several open TTS models scored side by side, covering the "which one to pick" half.
- Voice Cloning Toolkit Comparison: for cloning voices without deploying a model yourself.
- Cloud vs Local Voice AI Review: hosted API or local inference -- the cost and privacy math lives here.
- VoiceStudio Resource: voice creation workspaces for users who never want to touch a terminal.
- Local Speech-to-Text SOP with Faster-Whisper: preparing transcripts for cloning and post-processing voice data both start with local transcription.