Accept the BreezeBlue Research and Non-Commercial License to access these weights

These weights are a derivative of Breeze TTS 2 and remain governed by the BreezeBlue Research and Non-Commercial License Agreement. The complete, unmodified Agreement is the LICENSE file in this repository. Section 4(f) requires each recipient to accept it independently before accessing or using the weights, which is why this repository is gated. Approval is automatic once you accept. The full Agreement is reproduced at the bottom of the model card so that you can read it before deciding.

Read the Agreement before you accept. Gating means the LICENSE file in this repository is not readable until after you accept, so the complete text is reproduced in full at the bottom of the model card below, and the identical authoritative copy is publicly readable at https://huggingface.co/BreezeBlue/Breeze-TTS-2/blob/main/LICENSE (SHA-256 8b826c0957648ea1df2b870ba2627a93b502930c3ba40fa1af696cdfc2d2f0a3, byte-identical to the LICENSE shipped here). This is not an open source licence: it permits research and non-commercial use only, and any commercial use requires a separate written licence from BreezeBlue. Accepting shares your Hugging Face username and email address with the repository owner.

Vireo TTS 3B — MLX, mixed 4-bit, activation-pruned

An unofficial, unaffiliated Apple-silicon build derived from BreezeBlue/Breeze-TTS-2.

Derived from Breeze TTS 2 by BreezeBlue and licensed for research and non-commercial use only.

This repository is not produced, endorsed, sponsored or reviewed by BreezeBlue or RESONIA, INC. "Vireo" is this build's own name and carries no affiliation. The reference to Breeze TTS 2 is descriptive attribution required by Section 4 of the licence.

A 2.1 GB build of Breeze TTS 2 for Apple silicon: mixed low-bit quantization plus activation-calibrated width pruning, running on a native MLX inference path with a streaming MLX codec.

On the one machine it was measured on — an M2 MacBook Pro with 16 GB — it generates faster than real time in voice-clone mode (1.12–1.21× real time) with 285-322 ms to first audio once warm, in 3.0 GB of peak GPU memory. Those figures are that machine's, not a general claim; see Benchmarks and What these numbers do not establish.

	This build	Companion full-quality build
Weights on disk	2.1 GB	5.9 GB
Peak GPU memory, single stream	3.0 GB	7.0 GB
Median real-time factor	0.870 (1.15× real time)	2.436 (0.41× real time)
Warm time to first audio, clone mode	285-322 ms	415-481 ms
Word error rate	0.0093	see that card
Precision	mixed int4/int8, pruned	bfloat16, unpruned

Which build should I use?

Use this one if you want interactive or real-time speech on Apple silicon, if you are on a 16 GB machine, or if you are running many requests. It is the build that was optimized and measured end to end.

Use the bfloat16 build if you want the unmodified numeric behaviour of the released checkpoint in MLX form — as a quality reference, to measure what this build's compression costs, or because you have memory to spare and do not need speed. It is roughly 2.8× slower and needs 2.3× more memory on the same machine.

Neither build is a replacement for upstream's PyTorch/CUDA release on non-Apple hardware.

Usage

The MLX runtime ships in this repository, as the breeze_mlx/ package. It is the same code every number on this page was measured with. You do not need a separate checkout, and you do not need PyTorch to generate speech.

Install

pip install "mlx>=0.32" numpy soundfile transformers
hf download mchen04/Vireo-TTS-3B-MLX-mixed4bit --local-dir vireo        # accepts the licence gate first
cd vireo && python -m breeze_mlx.cli . --text "Hello there." --output out.wav

That is the whole dependency list for text-to-speech. transformers is used only to build the tokenizer from tokenizer.json; no model is loaded through it.

Verified environment

Everything on this page was produced on the machine described under Benchmarks. The clean-room check that runs the downloaded repository with no local checkout used:

python       3.12.13
mlx          0.32.2
numpy        2.5.2
soundfile    0.14.0
transformers 5.16.1     # tokenizer only; 4.57.3 also verified

The optimization work itself was done with transformers==4.57.3. Both major versions are tested: the runtime builds the tokenizer directly from tokenizer.json rather than through AutoTokenizer, because the automatic route inspects config.json and transformers 5 rejects this model's config for reasons unrelated to tokenizing. Both routes produce identical token ids.

Optional extras

For	Install
Voice cloning from a reference clip	`pip install torch torchaudio qwen-tts`
Streaming HTTP server (`breeze_mlx.server`)	`pip install fastapi uvicorn python-multipart`
Long-form narration (`breeze_mlx.audiobook`)	`ffmpeg` and `ffprobe` on `PATH`

Voice cloning also needs upstream's codec encoder weights, which a converted bundle does not carry — decoding is native MLX and never uses them, and they are 682 MB:

hf download BreezeBlue/Breeze-TTS-2 --include 'audio_tokenizer/*' --local-dir breeze-tts-2
export BREEZE_AUDIO_TOKENIZER=$PWD/breeze-tts-2/audio_tokenizer

Python

from breeze_mlx import BreezeMLXRuntime, GenerationConfig, MLXCodec

bundle  = "vireo"                      # this repository, downloaded
codec   = MLXCodec(bundle)
runtime = BreezeMLXRuntime(bundle, codec=codec,
                           generation=GenerationConfig(max_new_tokens=750))

request = {
    "text": "The morning fog had not yet lifted from the harbour.",
    "instruction": "A warm, thoughtful young woman with a clear voice.",
    "speaker": "S0",
}

chunks = []
for chunk in runtime.stream(request, template="tts_instruction",
                            cfg_scale=1.0, seed=42):
    if chunk.audio.size:
        chunks.append(chunk.audio)     # float32, 24 kHz mono, streamed

Voice cloning takes a reference clip and its exact transcript:

from breeze_mlx import encode_reference

codes = encode_reference("reference.wav")          # needs the extras above
request |= {"ref_text": "exact transcript of reference.wav"}
for chunk in runtime.stream(request, template="ref_edit_tata",
                            cfg_scale=1.0, seed=42, audio_codes=codes):
    ...

Voice design without a reference clip uses template="tts_instruction" and benefits from classifier-free guidance (cfg_scale=4.0), which runs a two-branch batch and is measurably slower per second of audio.

Other entry points

python -m breeze_mlx.bench .                 # the benchmark on this page
python -m breeze_mlx.server .                # streaming HTTP server
python -m breeze_mlx.audiobook --help        # long-form narration
python -m breeze_mlx.build_final --help      # rebuild the weights from source

File layout

File	What it is
`weights.safetensors`	1745 tensors: quantized language stack (`text_encoder.`, `backbone.`, `depth.`, `lm_head.`, `text_encoder_proj.*`). Quantized modules are stored as `.weight` / `.scales` / `.biases` triplets.
`codec.safetensors`	253 tensors: the audio codec decoder in MLX layout, float16
`mlx_meta.json`	source repo and revision, per-stage bit allocation, the exact `{module: [group, bits, mode]}` map, and the pruning record
`codec_meta.json`	codec geometry: 16 codebooks, 24 kHz, 1920 samples per frame, sliding window 72
`config.json`, `generation_config.json`	byte-identical to upstream
`tokenizer.json`, `tokenizer_config.json`, `special_tokens_map.json`	byte-identical to upstream
`audio_tokenizer/*.json`	byte-identical to upstream
`LICENSE`	the complete BreezeBlue Agreement, byte-identical to upstream's
`NOTICE`	required attribution notice
`MODIFICATIONS.md`	every change made to the Model Materials
`breeze_mlx/`	the MLX runtime: the package that loads and runs these weights (Apache-2.0)
`LICENSE.apache-2.0`	the licence covering `breeze_mlx/`, which is not the licence covering the weights
`SHA256SUMS`	digest of every file in this repository

Loading tokenizer.json with transformers prints a warning about a Mistral regex pattern. It comes from the upstream tokenizer file, which is unmodified here, and is cosmetic.

Model

Breeze TTS 2 is a three-stage speech model followed by a residual-vector-quantized codec. Shapes below are this build's, after pruning.

Stage	Type	Layers	Hidden	MLP width	Note
Text encoder	`t5gemma2_text`	26	1152	6912	not pruned
Backbone	`qwen3`	28	2048	4608	pruned from 6144
Depth decoder	16-codebook RVQ head	12	1024	6144	pruned from 8192
Codec	Mimi-style, 24 kHz	8	512	—	float16

The source checkpoint holds 2.92 B parameters in the language stack and 114 M in the codec.

The depth decoder runs once per codebook — fifteen times per audio frame — so its width dominates steady-state cost, which is why it is pruned hardest and quantized most carefully.

Benchmarks

One machine. One voice. Read the caveats.


Machine	MacBook Pro `Mac14,7`, Apple M2, 8 cores (4P/4E), 16 GB unified memory
OS	macOS 26.5.2 (build 25F84)
Stack	`mlx==0.32.2`, Python 3.12.13
Method	fastest of 3 repeats per timing probe; interference from other processes is one-sided, so the minimum is the least biased estimate of what the code can do — and therefore a best case, not an expectation
Reference clip	one 7.5 s English clip of a single adult female speaker, not distributed

Single stream

Clone mode (ref_edit_tata, CFG 1.0) except the last row.

Case	Audio	Generated in	Real-time factor	× real time	Time to first audio
Short (1 sentence)	2.80 s	2.50 s	0.892	1.12×	285 ms
Medium (2 sentences)	6.80 s	5.76 s	0.847	1.18×	292 ms
Long (4 sentences)	22.24 s	18.41 s	0.828	1.21×	322 ms
Voice design, CFG 4.0	6.00 s	6.62 s	1.103	0.91×	230 ms

Median real-time factor 0.870. Voice design at CFG 4 runs a two-branch batch and is slower than real time; it is reported separately rather than folded into the headline.

Verified from a clean machine

Both repositories are checked end to end after every upload, from a process with no access to any source checkout: an empty Hugging Face cache, a fresh virtual environment holding only mlx, numpy, soundfile and transformers, and no PyTorch installed at all. The runtime imported is the breeze_mlx/ in the downloaded repository — the check asserts that and fails if any other copy is importable.

	this build	the other build
Files hash-verified	40 (24 of them runtime)	40 (24 of them runtime)
Hash mismatches	0	0
Anonymous download	refused (`GatedRepoError`)	refused (`GatedRepoError`)
Load	2.16 s	3.08 s
Audio produced	7.76 s	7.84 s
Generated in	6.59 s	20.36 s
Real-time factor	0.849	2.597
× real time	1.177×	0.385×
Time to first audio	389 ms	1114 ms
Peak GPU memory	2622 MB	6515 MB
Peak process RSS	2640 MB	3212 MB
Output	186,240 samples, all finite	188,160 samples, all finite
Output peak / RMS	0.3127 / 0.0423	0.4324 / 0.0653

The generated WAV is byte-identical to the one produced from the local bundle before upload:

331e73b10f74e9e3c24ccd236dfb7ff77cee388614e7081a3d5cf107c9b66909  7.76 s of 24 kHz mono PCM-16

This is a voice-design prompt at CFG 1.0 with seed=42 and max_new_tokens=750, so it is reproducible, but it is a single sample and not a benchmark — the numbers under Benchmarks are the measured ones.

Startup and memory


Load (weights + codec, cold process)	4.25 s
Cold time to first audio, first generation of the process	529 ms
Warm time to first audio, clone mode	285-322 ms
Warm time to first audio, best probe overall (voice design)	230 ms
Peak GPU memory, single stream	3070 MB

Batch

Offline batched synthesis, one shared voice, no classifier-free guidance. Wall time includes codec decode. Same machine, fastest of 2 repeats.

Batch	Audio produced	Wall time	× real time	Peak GPU memory
1	4.80 s	3.97 s	1.21×	4871 MB
2	9.44 s	6.93 s	1.36×	5302 MB
4	20.80 s	12.53 s	1.66×	6385 MB
8	38.40 s	20.13 s	1.91×	8180 MB

Batching buys throughput, not latency: decode is bandwidth-bound and the bytes pulled per frame are the same whether one sequence or eight are in flight. Batch 8 needs 8.2 GB, so on a 16 GB machine it is comfortable and on an 8 GB machine it is not. All 8 sequences produced finite, non-empty audio and none hit the token limit.

Quality

Measured by the same sealed harness that drove the optimization: 24 probes, three seeds each, transcribed with Whisper small.en, speaker similarity by microsoft/wavlm-base-plus-sv x-vector cosine against the held-out reference.


Word error rate (mean)	0.0093
Word error rate (median)	0.0000
Speaker similarity (mean)	0.958
Speaker similarity (min over probes)	0.856
Stability (probes that produced audio)	1.00
Peak clipped fraction	0.0
Non-finite samples	0

Against a held-out probe set rotated in once, at the end: word error rate 0.0368, speaker similarity 0.955. That number is higher than the development figure and is the more honest estimate of unseen-text behaviour.

Do not read 0.0093 as "better than uncompressed"

The bfloat16 companion was run through the same harness at the same settings, and scored word error rate 0.0235 — worse than this build's 0.0093. That comparison flatters this build for a reason that should be stated rather than enjoyed.

	This build (mixed 4-bit)	bfloat16 companion
Word error rate, development probes	0.0093	0.0235
Word error rate, held-out probes	0.0368	not measured
Speaker similarity, mean	0.9580	0.9598
Speaker similarity, worst probe	0.8559	0.9031
Peak GPU memory (harness)	2807 MB	6609 MB

This configuration is the survivor of more than sixty candidates, every one of them scored on this probe set with this ASR, with word error rate acting as a gate. Selecting that hard against a metric buys performance on it — which is exactly what the held-out figure of 0.0368 shows. The bfloat16 build was never tuned against these probes.

And the floor moved the other way: this build's worst-case speaker similarity is 0.8559 against the companion's 0.9031. Compression did cost something; it cost consistency, in the place the mean does not show.

The defensible claim is narrow: at a third of the memory and 2.8× the speed, this build stays close to the uncompressed one on measured quality, and is not uniformly better than it.

Calibration scope

The pruning step ranks feed-forward channels by their mean absolute gated activation on real generation traces. Those traces were produced from six fixed English prompts, 60 frames each, conditioned on one 7.5-second English reference clip of a single adult female speaker.

What this means, stated precisely:

What the weights contain: a choice of which channels to keep. For each layer the calibration produces a ranking, and the ranking is reduced to a set of surviving channel indices. That set is the only thing that reaches the weights.
What the weights do not contain: no reference audio, no transcript, no speaker embedding, no codec codes, no cached voice. This was verified by enumerating every tensor name in both files and inspecting the safetensors headers; there are no voice, reference, speaker or cache tensors, and __metadata__ is empty in both files.
The reference clip, its transcript and its encoded codes are not distributed — not in this repository and not anywhere else. The speaker is not identified here.
What the calibration biases: the surviving channels are the ones that mattered for one voice on English prompts. Speaker similarity was the constraint that bound hardest during optimization, and several candidates were rejected on it, so this build is tuned to preserve that voice. Whether it preserves other timbres equally well is untested. The model supports Chinese; no Chinese-language evaluation was run.

This also sets a reproducibility boundary: because the calibration reference is not distributed, a third party cannot rebuild these exact weights. See Reproducing this build.

What these numbers do not establish

Speed does not transfer. Every figure comes from one M2 with 16 GB. The decode loop is memory-bandwidth-bound, so an M2 Pro, M3 or M4 will differ, and not always in the direction you expect. No claim is made about any machine other than the one named above.
The headline is a best case. Speed is the fastest of three repeats. During this work the machine's disk was above 90% full with several gigabytes of swap in use, and individual runs stalled for seconds when memory-mapped weights were paged back. The two-pass method is what makes the numbers stable, not an absence of interference.
Quality was measured, not judged. Word error rate, x-vector similarity, F0 statistics and artifact counters catch intelligibility loss, timbre drift, prosodic flattening and signal defects. A candidate could still sound worse in a way none of them registers. There is no listening test behind these numbers.
Word error rate is instrument-relative. Whisper small.en. A stronger ASR would give different absolute numbers; only the deltas between candidates drove decisions.
Speaker similarity is one clip of one voice. It does not discriminate finely between similar-sounding speakers of the same demographic — an unrelated voice-design output of the same demographic scored 0.856 against the same reference, the same value as the weakest retained clip.
One language was evaluated. English only.

Known costs of the compression

These are real and were accepted deliberately:

A 6-bit depth decoder measured word error rate 0.0018 — an order of magnitude better than what shipped — at 19% less speed. If your operating point is quality-first rather than real-time-first, this build is the wrong choice and the bfloat16 companion or a higher-bit rebuild is the right one.
The optimization rule was willing to spend up to 2% speed for a measured quality gain, and did so four times. If part of that quality signal was noise, the rule spent real speed for imagined quality. It is bounded but it is a real asymmetry, and it was frozen before any measurement was taken.
The sealed holdout pool was honour-based: it was readable by the optimizer and discipline, not mechanism, kept it sealed. It was used once, at the end.
Sampling is reproducible for a given seed within this implementation but does not reproduce upstream's PyTorch token sequences — MLX and PyTorch use different random number generators. Correctness was established by stagewise numeric parity and greedy-decode agreement, not by matching sampled output.

Reproducing this build

The conversion, quantization and pruning pipeline is deterministic and rebuilds these weights bit for bit — 1745 of 1745 tensors identical — from the pinned source revision, given the same calibration reference clip:

python -m breeze_mlx.build_final \
  --src   <upstream checkpoint @ c1c8ca18b70b30822735633991d9ebf4898e47d4> \
  --dst   <output bundle> \
  --ref-audio <reference.wav> \
  --ref-text  "<exact transcript of reference.wav>"

Three things have to be pinned or the rebuild diverges, and all three are part of the recipe rather than incidental:

The calibration substrate. Scores are measured on a uniform int4 group-64 copy of the model, not on the final per-module bit allocation. Measuring on the final bundle produces a different ranking and a measurably worse model.
The trace length. 60 frames per calibration prompt. 40 frames ranks the backbone's channels differently.
The calibration voice. Calibrating on the voice you intend to clone rather than on voice-design prompts. Building without --ref-audio scored word error rate 0.0514 against this build's 0.0093, at identical speed.

Honest boundary: you cannot reproduce this exactly. The reference clip is private and is not distributed, so item 3 cannot be satisfied by a third party. A rebuild with your own reference clip will follow the same recipe and land at a similar operating point, but it will not be bit-identical to these files and its quality will differ. Two such rebuilds were measured at word error rate 0.0305 and 0.0514 against this build's 0.0093. That gap is disclosed rather than smoothed over.

The pipeline also had a real reproducibility bug: a capability added for a rejected experiment (quantizing the per-codebook output heads) kept firing for any preset that set embeddings, silently changing the substrate later calibrations ran on. It was found because a fresh build of the documented recipe scored 0.048 where the shipped bundle scored 0.009, at identical speed and similarity. Head quantization is now an explicit per-preset field. This is recorded because a reproducibility claim is worth less without the story of when it was false.

Checksums

SHA256SUMS in this repository covers every file. The weights:

6595b8e0bae4c87e00f7306ee7fd9efd1c7eaf3a5bf63cfc27d6bccf4cf2830e  weights.safetensors
cea6c821f4a3a69fdddd6a0500ce1ace293041107219d3fa173badce91730889  codec.safetensors

Source checkpoint, verified against the Hub's published LFS digests before conversion:

abf813781256e10cbe81f2dbb415f897556225d4dfa0282d67aa8ea164e114a9  model-00001-of-00002.safetensors
36aa73b1a11361e1774db90d9c63c63303b294c022de112aa51904d940edcef1  model-00002-of-00002.safetensors

Related work

Other Apple-silicon ports of the same upstream model exist and may suit you better. This build is distinguished by the activation-calibrated pruning and the per-stage mixed bit allocation, not by being the only MLX conversion:

Licence, attribution and responsible use

Derived from Breeze TTS 2 by BreezeBlue and licensed for research and non-commercial use only.

Breeze TTS 2 is licensed under the BreezeBlue Research and Non-Commercial
License Agreement. Copyright (c) 2026 RESONIA, INC. All Rights Reserved.

The complete, unmodified Agreement is LICENSE in this repository, byte-identical to the one upstream ships. It is not an open source licence.

Two licences, covering different things

	Licence	Applies to
The weights	BreezeBlue Research and Non-Commercial (`LICENSE`)	`weights.safetensors`, `codec.safetensors`, the tokenizer and the configs
The runtime	Apache-2.0 (`LICENSE.apache-2.0`)	the `breeze_mlx/` package

This split is not an interpretation. Section 1.2 of the BreezeBlue Agreement says Model Materials "do not include source code expressly licensed under the Apache License, Version 2.0", and Section 14.2 says such source is governed by its own licence. The code in breeze_mlx/ is therefore Apache-2.0 and carries none of the non-commercial restriction.

That does not make the pair commercially usable. The runtime is a loader for weights that remain restricted, and running these weights for a Commercial Purpose is forbidden whatever the code licence says.

Research and non-commercial use only. Commercial use of these weights, of anything derived from them, or of their output requires a separate written licence from BreezeBlue. There is no creator, monetization, small-business or revenue-threshold exception.
Each recipient must accept the Agreement independently (Section 4(f)). That is why this repository is gated. If you redistribute these weights you must impose the same requirement, ship the complete Agreement and the NOTICE file, keep every attribution and copyright notice, state your own modifications, and licence your distribution under this same Agreement without adding conflicting terms.
Voice cloning consent. Section 5(c) forbids using this model to clone, simulate or imitate a real person's voice, identity or likeness without that person's explicit and legally sufficient consent, and Section 5(d) requires you to hold all rights to any reference audio you submit. Section 5(i) forbids distributing synthetic audio where a reasonable person would be materially misled about its origin.
No endorsement. BreezeBlue and RESONIA, INC. have not reviewed, approved or endorsed this build. No trademark licence is granted beyond the descriptive attribution Section 4 requires.
Warranty and liability. The Agreement disclaims all warranties and caps BreezeBlue's aggregate liability at US$100. Sections 9, 10 and 11 apply to you.

Commercial licensing: https://breezeblue.ai · contact@breeze.blue

Appendix: the complete licence text

Gating this repository has one awkward consequence — the LICENSE file inside it is not readable until after you have accepted. Asking someone to agree to a document they cannot open is not informed consent, so the complete Agreement is reproduced here, in the model card, which renders publicly.

This text is byte-identical to the LICENSE file in this repository and to upstream's own copy at BreezeBlue/Breeze-TTS-2/LICENSE, which is not gated and can be read there instead. All three have SHA-256 8b826c0957648ea1df2b870ba2627a93b502930c3ba40fa1af696cdfc2d2f0a3.

BreezeBlue Research and Non-Commercial License Agreement, Version 1.0 — full text

BREEZEBLUE RESEARCH AND NON-COMMERCIAL LICENSE AGREEMENT

Version 1.0
Last Updated: August 25, 2026

This BreezeBlue Research and Non-Commercial License Agreement (the
"Agreement") is entered into between RESONIA, INC., doing business as
BreezeBlue ("BreezeBlue," "Licensor," "we," or "us"), and the individual or
entity accessing, using, modifying, or distributing any Model Materials
("You," "Your," or "Licensee").

By accessing, downloading, using, modifying, or distributing any portion of
the Model Materials or a Derivative Model, You acknowledge that You have read,
understood, and agreed to this Agreement. If You act on behalf of an entity,
You represent that You have authority to bind that entity, and "You" includes
both You and that entity.

This Agreement permits research and non-commercial use of the Model Materials
free of charge. It does not grant any commercial rights and is not an open
source license. Commercial use requires a separate written license from
BreezeBlue.


1. DEFINITIONS

1.1 "Model" means the Breeze TTS 2 model.

1.2 "Model Materials" means the Model weights, checkpoints, parameters,
adapters, model-specific tokenizer or codec weights, embeddings, and other
machine-learning artifacts that BreezeBlue expressly makes available under
this Agreement, together with any accompanying documentation expressly
designated as part of the Model Materials.

Model Materials do not include source code expressly licensed under the
Apache License, Version 2.0, or any third-party component made available under
its own license.

1.3 "Derivative Model" means any modification, adaptation, fine-tune, LoRA,
merge, quantization, distillation, or other model that is based on or derived
from the Model Materials, including a model created using the weights,
parameters, activations, embeddings, or Outputs of the Model Materials to
replicate or materially transfer their speech or audio generation
capabilities. Outputs themselves are not Derivative Models.

1.4 "Output" means audio, speech, text, metadata, or other content generated
by the Model Materials or a Derivative Model in response to an input.

1.5 "Research Purpose" means academic or scientific research that is not
primarily intended for commercial advantage or monetary compensation,
including peer-reviewed research, reproducibility studies, benchmarking,
evaluation, security testing, and red-teaming.

1.6 "Non-Commercial Purpose" means personal, educational, hobbyist, or other
use that is not primarily intended for commercial advantage or monetary
compensation, including limited evaluation and testing.

1.7 "Commercial Purpose" means any use primarily intended for or directed
toward commercial advantage, business benefit, or monetary compensation,
whether direct or indirect. Commercial Purpose includes, without limitation:

(a) production use or use in a product, service, or business operation;

(b) hosting or making the Model Materials, a Derivative Model, or their
functionality available through an API, SaaS offering, application, plug-in,
website, or other service;

(c) creating or providing content, products, or services for a fee, for a
client, in connection with advertising, sponsorship, subscription revenue,
or any other monetized activity;

(d) using the Model Materials, a Derivative Model, or Outputs for an
organization's internal operations beyond limited evaluation; or

(e) distributing, licensing, selling, or otherwise commercializing the Model
Materials or a Derivative Model.

The status of a Licensee as a for-profit entity does not, by itself, make
limited internal research, benchmarking, evaluation, security testing, or
red-teaming a Commercial Purpose, provided that the use is not deployed in
production, made available to customers or other external end users, used to
generate revenue, or used to train or improve a non-BreezeBlue generative
model.

1.8 "Affiliate" means an entity that directly or indirectly controls, is
controlled by, or is under common control with a party, where "control" means
ownership or control of more than fifty percent (50%) of the voting interests
of that entity.


2. RESEARCH AND NON-COMMERCIAL LICENSE

Subject to Your continued compliance with this Agreement, BreezeBlue grants
You a limited, non-exclusive, worldwide, non-transferable, non-sublicensable,
royalty-free license under BreezeBlue's intellectual property rights in the
Model Materials to:

(a) access, use, reproduce, and modify the Model Materials;

(b) create Derivative Models; and

(c) distribute the Model Materials and Derivative Models as expressly
permitted by Section 4,

in each case solely for a Research Purpose or Non-Commercial Purpose.

No rights are granted except as expressly stated in this Agreement.


3. COMMERCIAL USE

No Commercial Purpose is permitted under this Agreement. Any Commercial
Purpose involving the Model Materials, a Derivative Model, or an Output
requires a separate written commercial license from BreezeBlue.

For clarity, this Agreement does not include a creator, content monetization,
small-business, revenue-threshold, or other implied commercial-use exception.

This Agreement governs only the open-weight Model Materials and their
self-hosted use. Outputs generated through BreezeBlue's hosted products or
APIs are governed by the applicable Terms of Service, plan terms, and any
separate written agreement, rather than this Agreement.

For commercial licensing inquiries, contact:

Website: https://breezeblue.ai
Email: contact@breeze.blue


4. DISTRIBUTION AND ATTRIBUTION

You may distribute the Model Materials or a Derivative Model solely for a
Research Purpose or Non-Commercial Purpose and only if You:

(a) provide every recipient with a complete copy of this Agreement and any
NOTICE file accompanying the Model Materials;

(b) retain all copyright, attribution, provenance, and proprietary notices;

(c) include the following notice in a NOTICE file distributed with the Model
Materials or Derivative Model:

"Breeze TTS 2 is licensed under the BreezeBlue Research and Non-Commercial
License Agreement. Copyright (c) 2026 RESONIA, INC. All Rights Reserved.";

(d) clearly identify any modifications You made and, for a Derivative Model,
state prominently in its model card, repository, and documentation:

"Derived from Breeze TTS 2 by BreezeBlue and licensed for research and
non-commercial use only.";

(e) license the distributed Model Materials or Derivative Model under this
Agreement without adding terms that conflict with or reduce the protections
of this Agreement; and

(f) ensure that each recipient independently accepts this Agreement before
accessing or using the Model Materials or Derivative Model.

You may not use "BreezeBlue," "Breeze TTS 2," or any confusingly similar name
as the primary name of a Derivative Model, product, or service, or in any way
that suggests endorsement, affiliation, official status, or successor status.
Reasonable descriptive attribution required by this Section is permitted.

Publication of non-commercial research results, benchmark results, and
illustrative samples in papers, reports, conference presentations, model
cards, repositories, or non-monetized demonstrations is permitted, provided
that You reasonably acknowledge Breeze TTS 2 and comply with this Agreement.


5. USE RESTRICTIONS

You will not, and will not permit any third party to:

(a) use the Model Materials, a Derivative Model, or any Output for a
Commercial Purpose without a separate written commercial license;

(b) use the Model Materials, a Derivative Model, or any Output to create,
train, fine-tune, distill, evaluate for training selection, or otherwise
improve any non-BreezeBlue foundational or generative speech, audio,
language, or multimodal model, except for creating a Derivative Model solely
for a Research Purpose or Non-Commercial Purpose under this Agreement;

(c) use the Model Materials, a Derivative Model, or any Output to clone,
simulate, imitate, or impersonate a real person's voice, identity, likeness,
or persona without that person's explicit and legally sufficient consent;

(d) submit or use reference audio, speaker embeddings, voice prompts,
recordings, performances, or other materials unless You have all rights,
licenses, permissions, and consents necessary for that use;

(e) use the Model Materials, a Derivative Model, or any Output to defraud,
deceive, harass, stalk, threaten, defame, exploit, or cause harm to any person;

(f) generate or distribute sexual, exploitative, or otherwise harmful content
involving a minor;

(g) engage in deceptive political persuasion, election interference, voter
suppression, synthetic robocalls, or false impersonation of a candidate,
public official, or other person in a political context;

(h) identify, authenticate, track, surveil, or profile a person using voice or
other biometric characteristics without a lawful basis and all consent
required by applicable law;

(i) distribute synthetic audio in circumstances where a reasonable person
would be materially misled about its origin or authenticity, or fail to make
any disclosure required by applicable law;

(j) remove, obscure, alter, or circumvent any copyright notice, attribution,
watermark, provenance signal, safety mechanism, or other proprietary notice
included in the Model Materials or Outputs;

(k) attempt to extract or reconstruct non-public training data, source data,
trade secrets, or proprietary components from the Model Materials, except to
the extent such a restriction is prohibited by applicable law; or

(l) use, export, re-export, transfer, or make available the Model Materials or
a Derivative Model in violation of any applicable law, regulation, export
control, or trade sanction.

Nothing in Section 5(k) prohibits ordinary inspection, loading, conversion,
quantization, modification, or fine-tuning of the released Model Materials as
expressly permitted by this Agreement.


6. LEGAL COMPLIANCE AND THIRD-PARTY RIGHTS

You are solely responsible for ensuring that Your use of the Model Materials,
Derivative Models, inputs, and Outputs complies with all applicable laws and
regulations and does not infringe or violate any third-party right, including
intellectual property, privacy, publicity, biometric privacy,
recording-consent, consumer-protection, election, and other applicable rights.

This Agreement grants no right to use, reproduce, clone, simulate, imitate,
or otherwise exploit any third-party voice, identity, likeness, persona,
performance, recording, copyrighted work, trademark, or other protected
material. You are solely responsible for obtaining all necessary rights,
licenses, permissions, notices, and consents.

BreezeBlue does not represent or warrant that any Model Material or Output is
cleared for Your intended use, unique, accurate, non-infringing, or suitable
for any particular purpose.


7. INTELLECTUAL PROPERTY

7.1 Model Materials. BreezeBlue and its licensors retain all right, title, and
interest in and to the Model Materials. Except for the limited license granted
in Section 2, no right or interest in the Model Materials is transferred to
You.

7.2 Derivative Models. As between You and BreezeBlue, You own the original
modifications You contribute to a Derivative Model, to the extent permitted by
applicable law. That ownership is subject to BreezeBlue's ownership of the
underlying Model Materials and to every restriction in this Agreement. A
Derivative Model remains governed by this Agreement.

7.3 Outputs. As between You and BreezeBlue, and to the extent permitted by
applicable law, You own Outputs You lawfully generate. Output ownership does
not remove or limit the non-commercial restrictions in this Agreement, grant
rights in third-party material, or authorize any use prohibited by this
Agreement.

7.4 Trademarks. No trademark license is granted except the limited right to
make accurate descriptive references required for attribution under Section
4. All goodwill arising from permitted use of BreezeBlue's marks belongs to
BreezeBlue.

7.5 Feedback. If You voluntarily provide suggestions, comments, or other
feedback about the Model Materials, You grant BreezeBlue a perpetual,
irrevocable, worldwide, royalty-free, fully paid, transferable, sublicensable,
non-exclusive license to use and exploit that feedback without restriction or
compensation.

7.6 Intellectual Property Claims. If You or Your Affiliate brings a claim or
legal proceeding alleging that the Model Materials, a Derivative Model, or an
Output infringes intellectual property rights owned or licensable by You or
Your Affiliate, the licenses granted to You under this Agreement terminate
automatically when the claim or proceeding is filed.


8. REPRESENTATIONS

You represent and warrant that:

(a) You have legal capacity and authority to enter into this Agreement;

(b) You will use and distribute the Model Materials, Derivative Models, and
Outputs only as permitted by this Agreement;

(c) You have all necessary rights and consents for Your inputs and intended
uses; and

(d) You are not prohibited from receiving or using the Model Materials under
applicable export-control, sanctions, or other laws.


9. DISCLAIMER OF WARRANTIES

TO THE MAXIMUM EXTENT PERMITTED BY APPLICABLE LAW, THE MODEL MATERIALS,
DERIVATIVE MODELS, OUTPUTS, AND ALL RELATED MATERIALS ARE PROVIDED "AS IS" AND
"AS AVAILABLE," WITH ALL FAULTS AND WITHOUT WARRANTIES OF ANY KIND, WHETHER
EXPRESS, IMPLIED, STATUTORY, OR OTHERWISE.

BREEZEBLUE DISCLAIMS ALL WARRANTIES, INCLUDING WARRANTIES OF TITLE,
NON-INFRINGEMENT, MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE,
ACCURACY, RELIABILITY, SECURITY, AND ANY WARRANTY THAT THE MODEL MATERIALS OR
OUTPUTS WILL BE ERROR-FREE, UNINTERRUPTED, UNIQUE, OR SUITABLE FOR YOUR USE.

YOU ARE SOLELY RESPONSIBLE FOR DETERMINING THE APPROPRIATENESS AND LAWFULNESS
OF YOUR USE AND ASSUME ALL RISKS ARISING FROM THAT USE.


10. LIMITATION OF LIABILITY

TO THE MAXIMUM EXTENT PERMITTED BY APPLICABLE LAW, BREEZEBLUE AND ITS
AFFILIATES, LICENSORS, OFFICERS, DIRECTORS, EMPLOYEES, CONTRACTORS, AND AGENTS
WILL NOT BE LIABLE UNDER ANY THEORY OF LIABILITY FOR ANY INDIRECT, INCIDENTAL,
SPECIAL, CONSEQUENTIAL, EXEMPLARY, OR PUNITIVE DAMAGES, OR FOR ANY LOSS OF
PROFITS, REVENUE, BUSINESS, GOODWILL, DATA, OR USE, ARISING OUT OF OR RELATING
TO THIS AGREEMENT, THE MODEL MATERIALS, ANY DERIVATIVE MODEL, OR ANY OUTPUT,
EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.

TO THE MAXIMUM EXTENT PERMITTED BY APPLICABLE LAW, BREEZEBLUE'S TOTAL
AGGREGATE LIABILITY ARISING OUT OF OR RELATING TO THIS AGREEMENT WILL NOT
EXCEED ONE HUNDRED U.S. DOLLARS (US$100).


11. INDEMNIFICATION

You agree to defend, indemnify, and hold harmless BreezeBlue and its
Affiliates, licensors, officers, directors, employees, contractors, and agents
from and against any third-party claim, demand, action, proceeding, damage,
loss, liability, judgment, settlement, cost, or expense, including reasonable
attorneys' fees, arising out of or relating to:

(a) Your access to, use, modification, or distribution of the Model Materials
or a Derivative Model;

(b) Your inputs or Outputs;

(c) Your violation of this Agreement or applicable law; or

(d) Your infringement, misappropriation, or violation of any third-party
right.


12. TERM AND TERMINATION

This Agreement begins when You first access or use the Model Materials and
continues until terminated.

BreezeBlue may terminate this Agreement if You materially breach it. For a
material breach that is reasonably capable of cure, BreezeBlue will provide
written notice and at least thirty (30) days to cure before termination.
BreezeBlue may terminate immediately for a breach of Section 3, Section 5, or
applicable law, or where delay would reasonably create legal, security, fraud,
or safety risk.

Upon termination, You must immediately cease using and distributing the Model
Materials and Derivative Models, delete all copies under Your control, and,
upon reasonable request, confirm deletion in writing. Termination does not
require deletion of an Output lawfully created before termination, but no
Output may be used in violation of this Agreement or applicable law.

BreezeBlue may stop distributing the Model Materials prospectively. Such
discontinuation alone does not terminate rights already granted to a Licensee
that remains in compliance with this Agreement.

Sections 3, 5, 6, 7, 9, 10, 11, 12, and 13 survive termination to the extent
necessary to give them effect.


13. GOVERNING LAW AND JURISDICTION

This Agreement is governed by and construed in accordance with the laws of the
State of Delaware, without regard to its conflict-of-laws principles. The
United Nations Convention on Contracts for the International Sale of Goods
does not apply.

To the extent permitted by applicable law, any dispute arising out of or
relating to this Agreement shall be brought exclusively in the state or
federal courts located in the State of Delaware, and each party consents to
the personal jurisdiction and venue of those courts.


14. MISCELLANEOUS

14.1 Entire Agreement. This Agreement and any NOTICE file accompanying the
Model Materials constitute the entire agreement concerning the Model
Materials and supersede prior or contemporaneous understandings concerning
the same subject matter.

14.2 Separate Software Licenses. Source code and third-party components may be
made available under separate licenses. Those separate licenses govern only
the materials expressly identified as subject to them and do not grant any
right to use the Model Materials.

14.3 Assignment. You may not assign or transfer this Agreement or any right
under it without BreezeBlue's prior written consent. BreezeBlue may assign
this Agreement in connection with a merger, acquisition, reorganization, sale
of assets, or operation of law.

14.4 Severability. If any provision is held invalid or unenforceable, it will
be enforced to the maximum extent permitted, and the remaining provisions
will remain in effect.

14.5 No Waiver. Failure to enforce a provision is not a waiver of the right to
enforce it later.

14.6 No Endorsement. Nothing in this Agreement creates an agency,
partnership, joint venture, employment, fiduciary, sponsorship, or endorsement
relationship between the parties.

14.7 Updates. BreezeBlue may publish new versions of this Agreement. A new
version applies only if You expressly accept it or access Model Materials
released under that new version. Your continued use of a previously obtained
version of the Model Materials remains governed by the version You accepted,
unless applicable law requires otherwise.


15. CONTACT

Questions and commercial licensing requests may be sent to:

RESONIA, INC., doing business as BreezeBlue
Website: https://breezeblue.ai
Email: contact@breeze.blue

Downloads last month: -

MLX

Hardware compatibility

Quantized

Model tree for mchen04/Vireo-TTS-3B-MLX-mixed4bit

Base model

BreezeBlue/Breeze-TTS-2

Quantized

(5)

this model