forkalt.

Apache-2.0audioRuns locally0.6B and 1.7B cards (family)

Qwen3-TTS: Specs, License & How to Run

Qwen Apache-2.0 TTS family — CustomVoice, VoiceDesign, Base 0.6B/1.7B. Not Audio8.

Official Qwen3-TTS introduction figure from the Qwen3-TTS repository assets
Official project image. Source · License: Qwen product image.
License
Apache-2.0
Parameters
0.6B and 1.7B cards (family)
HF likes
2,191 (2026-11-02)
First seen
2026-01-21

TL;DR. Qwen3-TTS is a audio model from Qwen. Parameters: 0.6B and 1.7B cards (family).License: Apache-2.0. It is listed under ElevenLabs.

DeveloperQwen
Typeaudio
ArchitectureDiscrete multi-codebook LM TTS with Qwen3-TTS-Tokenizer-12Hz (card). Family splits: CustomVoice (9 named speakers; 1.7B also takes instruct), VoiceDesign (1.7B description-driven), Base (3-second voice clone; 0.6B and 1.7B). pipeline_tag text-to-speech on the CustomVoice / VoiceDesign / 0.6B-Base cards we opened. Distinct from Audio8-TTS-Realistic-0.6B (Edge0 DualAR) and from Qwen3.8-27B (native VLM, not TTS).
Parameters0.6B and 1.7B cards (family)
Modalitiesaudio, speech, text
LanguagesChinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian
LicenseApache-2.0 — Hugging Face cardData.license is apache-2.0 on the 12Hz CustomVoice, VoiceDesign, Base, and Tokenizer cards we opened (2026-11-02). There is no LICENSE blob on Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice (resolve/main/LICENSE 404). GitHub QwenLM/Qwen3-TTS is Apache License 2.0 on the LICENSE file. Weights follow the card tag plus the inference-repo grant. Not legal advice.
HardwareNot yet verified — Official path is pip install -U qwen-tts (Python 3.12 recommended on the CustomVoice README). FlashAttention 2 is named as optional. Peak VRAM is not independently measured. “97ms” first-packet latency on the VoiceDesign / CustomVoice overview is vendor. 1.7B CustomVoice usedStorage is 4,617,449,338 bytes (~4.30 GiB). (not independently tested)
Hugging Face likes2,191 (as of 2026-11-02)
Model cardhttps://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
First seen2026-01-21

Qwen3-TTS is Qwen’s 12Hz text-to-speech family, not a single checkpoint. Official hosts we opened on 2026-11-02: the Qwen3-TTS collection, lead card Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice (created 2026-01-21T08:56:49Z; last modified 2026-01-29T08:00:54Z; 2,191 likes / 1,749,479 downloads), siblings listed below, github.com/QwenLM/Qwen3-TTS (14,154 stars / 1,831 forks; Apache-2.0 LICENSE), the launch note qwen.ai/blog?id=qwen3tts-0115, and paper arXiv:2601.15621. API gated was false on every card in that set we could open.

This page hangs off open ElevenLabs alternatives. That mapping is editorial. We have no sourced quality comparison against ElevenLabs.

It is not Audio8-TTS-Realistic-0.6B (Edge0 DualAR). Head-to-head facts: Audio8-TTS-Realistic-0.6B vs Qwen3-TTS. It is also not Qwen3.8-27B (native VLM) and not Qwen3.5-4B-3D-Coder (community Three.js).

The CustomVoice README’s “markedly improved robustness,” “97ms” first-packet latency, and “what you imagine is what you hear” lines stay vendor. This overview copies variant names and card stats only.

Family snapshot (2026-11-02)

Card Job (README) Likes / downloads Params (safetensors) usedStorage
12Hz-1.7B-CustomVoice 9 named speakers; instruct style control 2,191 / 1,749,479 1,916,676,352 4,617,449,338 (~4.30 GiB)
12Hz-1.7B-VoiceDesign Description-driven voice design 464 / 523,722 1,916,676,352 4,515,695,644 (~4.21 GiB)
12Hz-1.7B-Base 3-second clone / fine-tune base 585 / 3,437,605 1,928,677,440 4,539,846,851 (~4.23 GiB)
12Hz-0.6B-CustomVoice 9 named speakers 210 / 772,075 905,788,672 2,493,919,668 (~2.32 GiB)
12Hz-0.6B-Base 3-second clone / fine-tune base 360 / 744,498 914,643,008 2,511,637,364 (~2.34 GiB)
Tokenizer-12Hz Speech codec (encode/decode) 103 / 117,409 — —

License tag on each of those cards: apache-2.0. Languages copied from the family README: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian, plus dialectal speaker profiles on CustomVoice. A Qwen/Qwen3-TTS-12Hz-0.6B-VoiceDesign lookup returned 401 on 2026-11-02; it is not in the collection list we opened, so it is not claimed here.

CustomVoice speakers named on the 1.7B README: Vivian, Serena, Uncle_Fu, Dylan, Eric, Ryan, Aiden, Ono_Anna, Sohee. Those are product voice names, not a quality rank.

How to run (from the official README, not tested)

Copied names only. forkalt has not loaded a checkpoint.

pip install -U qwen-tts

The CustomVoice README recommends a fresh Python 3.12 env and optional flash-attn. Example entry: Qwen3TTSModel.from_pretrained("Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice", ...). Methods copied: generate_custom_voice, generate_voice_design, generate_voice_clone. Weights can also be pulled with huggingface-cli download for each repo id above.

Hosted Hugging Face Space Qwen/Qwen3-TTS is a demo, not the weight grant.

Vendor benches (self-reported)

The 1.7B VoiceDesign README prints Seed-TTS WER for Qwen3-TTS-12Hz-1.7B-Base: test-zh 0.77, test-en 1.24. Self-reported. The Audio8 Realistic card cites other Qwen Seed-TTS / CV3 / MiniMax figures as a reference comparison compiled from the Qwen3-TTS report and third-party tables — not a forkalt ranking. See Audio8 vs Qwen3-TTS.

License

Card tags are apache-2.0. LICENSE on QwenLM/Qwen3-TTS is Apache License 2.0. The 1.7B CustomVoice weight repo had no LICENSE file (404) on 2026-11-02. We record both. Not legal advice.

Related on this site: Audio8-TTS-Realistic-0.6B — Audio8-TTS-Realistic-0.6B vs Qwen3-TTS — ElevenLabs alternatives — Drama 3 — Phonon — LFM3-Audio-2B — LFM3-Omni-2B. Other Qwen cards on this site are different jobs: Qwen3.8-27B (VLM), Qwen-Image-2.1-Turbo. Still images beside a voiceover: Fooocus, Iris-3B vs Qwen-Image-2.1-Turbo. Catalog: home. Corrections: contact.

Benchmarks

Quoted from the cited source. forkalt has not independently reproduced these numbers.

NameScoreSource
Seed-TTS test-zh WER (1.7B-Base card)0.77 (self-reported)https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign
Seed-TTS test-en WER (1.7B-Base card)1.24 (self-reported)https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign

Deploy options

Comparisons

FAQ

Which Qwen3-TTS card should I download?

CustomVoice if you want nine named speakers (1.7B also takes style instructions). VoiceDesign if you want a natural-language voice description (1.7B only on the collection we opened). Base if you want 3-second voice cloning or a fine-tune starting point. Tokenizer-12Hz is the codec, not a TTS speaker model. This page is a family overview, not a substitute for each card.

Are the weights public?

Yes. Lead card Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice is ungated (created 2026-01-21T08:56:49Z; 2,191 likes / 1,749,479 downloads in the 2026-11-02 API check). Collection huggingface.co/collections/Qwen/qwen3-tts also lists 1.7B VoiceDesign, 1.7B Base, 0.6B CustomVoice, 0.6B Base, and Tokenizer-12Hz. Code is github.com/QwenLM/Qwen3-TTS (14,154 stars / 1,831 forks; Apache-2.0 LICENSE).

Can I use it commercially?

Card tags are apache-2.0. The CustomVoice weight repo had no LICENSE blob on 2026-11-02 (404). GitHub QwenLM/Qwen3-TTS LICENSE is Apache License 2.0. Not legal advice.

Does it beat ElevenLabs or Audio8-TTS-Realistic-0.6B?

No sourced bake-off. Seed-TTS rows on Qwen cards and the Audio8 reference table stay vendor-cited. See ElevenLabs alternatives and Audio8 vs Qwen3-TTS.

Related pages

Sources & verification

Last verified 2026-11-02 · Last updated 2026-11-02· Edited by forkalt editors · Editorial policy

License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.