Downloads · 30 days
66
12% of all-time downloads
jiangzhuo9357/qwen3-tts-12hz-0-6b-base-gguf
qwen3-tts-12hz-0-6b-base-gguf is a text-to-speech model from jiangzhuo9357. Use it when you need text read aloud. It is set up for synthesize.cpp. The card lists the license as apache-2.0.
GGUF conversions of the official Qwen3-TTS-12Hz-0.6B-Base checkpoint for synthesize.cpp.
Downloads · 30 days
66
12% of all-time downloads
All-time downloads
556
Public
Repo size
6.7 GB
Likes
0
Public
Click a slice to open those files.
.gguf6.7 GB · 100%
From the Hugging Face model README
GGUF conversions of the official Qwen3-TTS-12Hz-0.6B-Base checkpoint for synthesize.cpp.
Ported from QwenLM/Qwen3-TTS revision
022e286b98fbec7e1e916cb940cdf532cd9f488e and
validated on 2026-08-17 against the pinned upstream PyTorch implementation at bfloat16 on CUDA.
Voice cloning from a reference recording. Hand it a few seconds of speech and it synthesizes new text in that voice, either from the audio alone or -- when you also supply the recording's transcript -- through the transcript-assisted path upstream calls ICL.
This package catalogues NO preset voices. A caller must prepare a Voice Profile; there is no voice id to name. That is the structural difference from the CustomVoice package, which catalogues nine and cannot clone.
| Profile | Download | Size | Tensor storage | SHA-256 | What to know before choosing it |
|---|---|---|---|---|---|
| BF16 | qwen3-tts-12hz-0-6b-base-BF16.gguf | 2399.9 MB (2,516,522,464 bytes) | 478 BF16 + 416 F32 | 993f4cd164df2b2f83e72ec51d79586f5721732d52ee24c5b5462858c610999f | — |
| F16 | qwen3-tts-12hz-0-6b-base-F16.gguf | 2400.1 MB (2,516,706,912 bytes) | 304 F16 + 590 F32 | e74841947b62f9ac63f09bd7443959cee325e66ae46cedfb346a48553ff0a046 | LARGER than the BF16 source by 184,448 bytes, and that is not a defect. Both are two-byte types, so the matrix weights do not shrink while the sensitive tensors widen BF16 to F32. F16 is a speed profile for this family, not a size one. It clears every gate and does not pay; it is carried for completeness rather than recommended. |
| Q8_MIXED | qwen3-tts-12hz-0-6b-base-Q8_MIXED.gguf | 1590.4 MB (1,667,606,112 bytes) | 266 Q8_0 + 38 F16 + 590 F32 | 808667ae6b14e77a4379f2be3b43af3a961fb7864dd1b44eb94c061e62a789a9 | THE DEFAULT RECOMMENDATION, and the only profile that pays. 33.7% smaller than the source, RTF 0.863 against BF16's 3.15 -- faster than real time -- and 1.06 GiB less peak RSS, with no measured cost on any gated quantity and no audible degradation in the 2026-08-17 audit. |
| The 38 F16 tensors are the speaker encoder's convolution weights. They are never block-quantized: a block runs along the kernel extent, which is 1, 3 or 5 here, against Q8_0's block of 32. The codec half stays F32 under every profile, so the reference codes this package produces are byte-identical across all three. |
All profiles use the same Qwen3-TTS architecture and public synthesize.cpp API. The profile name describes a versioned storage policy, not the language or Execution Backend.
validation_level: port_validated
3 graph stages were replayed for 13 cases on DGX Spark CPU, NVIDIA GB10 CUDA 13.3. Duration structure was exact in every case. The codec decoder runs on the device; everything else runs on the CPU.
The autoregressive half is held there by the discrete-outputs rule — a sampled token index cannot absorb a tolerance — and it dominates the wall clock, so CUDA buys about 6% end to end (RTF 3.15 to 2.95) rather than the 35x the decoder stage itself sees.
The two graphs this variant adds are on the CPU by measurement, not by omission. A device twin would mirror 241 MB for roughly 3.7 s of extra load against those graphs' entire CPU cost of about 1.69 s on an 8-second reference, so the transfer is 2.2x the compute it would accelerate and no speedup makes it pay. They also run once per Voice Profile rather than once per synthesis, so a load-once process pays the mirror once and saves almost nothing — the opposite of what makes the decoder's twin worthwhile.
| Profile | Public-seam checks | Speaker x-vector cosine | CPU real-time factor |
|---|---|---|---|
| BF16 | 7 passed, 3 skipped as inapplicable | 0.99999467 | 3.15 |
| F16 | 7 passed, 3 skipped as inapplicable | 0.99999501 | not_measured |
| Q8_MIXED | 7 passed, 3 skipped as inapplicable | 0.99999501 | 0.863 |
THREE COLUMNS BECAUSE THIS VARIANT'S EVIDENCE IS NOT THE SAME KIND AS CustomVoice's. That package publishes waveform cosines from replayed codes. This one clones, so what matters is whether the Voice survives and whether the package is usable at all.
Public-seam checks counts the behavioural relations the C interface
promises: a named seed reproduces byte for byte, a different seed does not,
a random seed comes back concrete and replays, a different Voice changes the
audio, and the resolved language is reported. Three further checks in the
family-wide set are structurally inapplicable here and are counted as
skipped rather than passed -- a Voice Profile has no id for the seam to
echo, and two of them need a preset "dialect" Voice this package does not
catalogue.
Speaker x-vector cosine is the worst case of four, comparing the port's
ECAPA-TDNN embedding against the pinned PyTorch oracle's. Note where the
residual lives: at the x-vector's largest element the deviation is 0.68 of
one bfloat16 ulp, and recomputing the ORACLE's own final layer in float64
disagrees with the oracle by more than this port does. The comparison is
bounded by the reference, not by the port, which is why the halved profiles
score very slightly better than the source profile rather than worse.
CPU real-time factor is synthesis time over the audio produced, each row
measured on its OWN output length, on a Release-typed build. Below 1.0 is
faster than real time. Rows are never compared by dividing one RTF by
another's wall clock -- the profiles stop at different frame counts, which
is a property of their weights.
Quality evaluation has not been run. These results establish that the port, the unconditioned default-Voice request path, deterministic request path, and CPU/CUDA execution work. They do not claim perceptual equivalence, naturalness, intelligibility, or speaker similarity.
A listening audit found no obvious regression. One maintainer compared a small set against the reference and reported nothing audible. That is release evidence, not a measurement: no rated comparison, no panel, no score, and it does not change the validation level. It says a defect large enough to hear was not found in what was heard.
This package exposes no preset speaker catalog. Its package default has no fixed identity and no speaker conditioning: with no Voice Profile supplied, which speaker a request produces is emergent, and the synthesis seed alone does not pin it. Reproducing a speaker needs the same seed and the same Execution Backend, package, and decode settings; switching between the CPU and CUDA backends can produce a different speaker for an otherwise identical request. Callers who want a stable identity should instead build a Voice Profile from Reference Audio (voice cloning) and pass it to the request. It produces 24000 Hz mono F32 audio.
This package accepts raw UTF-8 text through the built-in
synthesize.qwen_bpe frontend, and also accepts exact token
IDs already produced by the same vocabulary. The frontend tokenizes text
directly: no grapheme-to-phoneme conversion happens or is needed, and the
runtime does not silently invoke eSpeak or download a frontend.
Build synthesize.cpp and synthesize a deterministic request:
git clone https://github.com/handy-computer/synthesize.cpp.git
cd synthesize.cpp
cmake -S . -B build -DSYNTH_BUILD_CLI=ON
cmake --build build -j
hf download jiangzhuo9357/qwen3-tts-12hz-0-6b-base-gguf qwen3-tts-12hz-0-6b-base-Q8_MIXED.gguf \
--local-dir models/qwen3-tts-12hz-0-6b-base
build/bin/synthesize-cli \
--model models/qwen3-tts-12hz-0-6b-base/qwen3-tts-12hz-0-6b-base-Q8_MIXED.gguf \
--output output.wav \
--text "Qwen3-TTS is awesome!" \
--language en \
--seed 0
The same local GGUF can be loaded through the public C ABI and wrapped by C++, Rust, or Python. Model loading never contacts Hugging Face.
Both the pinned source at 022e286b and the checkpoint at 5d839924 carry an
explicit Apache-2.0 grant, verified from the upstream model card rather than
from any port's README.
THESE THREE FILES NO LONGER LOAD, and that is a deliberate break rather than a defect in them. Read this before downloading.
They were converted on 2026-08-17, before the package format gained a key that
states positively which Voice Profile sources a package implements
(synthesize.voice.profile_sources). Until that key existed the runtime
inferred the answer from the Voice Mode, which worked only while
profile-sources meant exactly one model variant. It now means two, whose
source sets are disjoint, so the inference had to go and the loader requires
the declaration instead. No compatibility shim was written: the project is
pre-release and these packages have no dependent users, so a shim would buy
compatibility nobody needs at the cost of putting the inference back into a
loader that had just been made declarative.
What to do instead: convert from the upstream checkpoint with the current
scripts/convert-qwen3-tts.py. Everything else on this card -- the tolerances,
the validation counts, the x-vector cosines, the RTF figures -- was measured on
these exact files and still describes what the package computes. Only loading
changed.
Published 2026-08-17. Both the pinned source at 022e286b and the checkpoint
at 5d839924 carry an explicit Apache-2.0 grant, so no redistribution
assumption is required for this variant.
What it is, stated because the sibling package is not this: it clones a voice from a recording the caller supplies. The CustomVoice package cannot do that.
Reproduced from the pinned upstream repository card for offline provenance. The upstream repository remains authoritative.
Qwen3-TTS Technical Report | GitHub Repository | Hugging Face Demo
Qwen3-TTS is a family of advanced multilingual, controllable, robust, and streaming text-to-speech models. Trained on over 5 million hours of speech data spanning 10 languages, Qwen3-TTS supports state-of-the-art 3-second voice cloning and description-based control.
This specific checkpoint is the 0.6B Base model, which is capable of rapid voice cloning from a user-provided audio input.
pip install -U qwen-tts
# Optional: for optimized performance
pip install -U flash-attn --no-build-isolation
To clone a voice and synthesize new content using the Base model, you can use the following code snippet:
import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel
# Load the model
model = Qwen3TTSModel.from_pretrained(
"Qwen/Qwen3-TTS-12Hz-0.6B-Base",
device_map="cuda:0",
dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
)
# Reference audio for cloning
ref_audio = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-TTS-Repo/clone.wav"
ref_text = "Okay. Yeah. I resent you. I love you. I respect you. But you know what? You blew it! And thanks to you."
# Generate speech
wavs, sr = model.generate_voice_clone(
text="I am solving the equation: x = [-b ± √(b²-4ac)] / 2a? Nobody can — it's a disaster (◍•͈⌔•͈◍), very sad!",
language="English",
ref_audio=ref_audio,
ref_text=ref_text,
)
# Save the resulting audio
sf.write("output_voice_clone.wav", wavs[0], sr)
Qwen3-TTS covers 10 major languages (Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian) as well as multiple dialectal voice profiles to meet global application needs. Key features:
If you find this work useful, please consider citing the technical report:
@article{Qwen3-TTS,
title={Qwen3-TTS Technical Report},
author={Hangrui Hu and Xinfa Zhu and Ting He and Dake Guo and Bin Zhang and Xiong Wang and Zhifang Guo and Ziyue Jiang and Hongkun Hao and Zishan Guo and Xinyu Zhang and Pei Zhang and Baosong Yang and Jin Xu and Jingren Zhou and Junyang Lin},
journal={arXiv preprint arXiv:2601.15621},
year={2026}
}