Skip to main content
Model guide · checked November 2026

CosyVoice3 (CosyVoice 3): What It Is, Demo & How to Run It

CosyVoice 3, also written CosyVoice3, is the third generation of the open-source CosyVoice text-to-speech model. Its open release, Fun-CosyVoice 3.0 (December 2025), is a 0.5B-parameter model that clones a voice from a short sample in 9 languages and 18+ Chinese dialects. CosyVoice 3.5 is a newer model offered through Alibaba Cloud's API, with no open weights published.

CosyVoice versions: 1, 2, 3, Fun-CosyVoice 3.0 and 3.5

Which versions you can download, and which exist only as a paper or a hosted API.

CosyVoice versions: 1, 2, 3, Fun-CosyVoice 3.0 and 3.5
VersionModelOpen weightsDateSource
CosyVoice 1CosyVoice-300M (plus -SFT and -Instruct)YesJul 2024 (paper)arXiv
CosyVoice 2CosyVoice2-0.5BYesDec 2024README
CosyVoice 3 (paper)Research models up to 1.5B parametersNo — the 1.5B model is paper onlyMay 2025 (paper)arXiv
Fun-CosyVoice 3.0Fun-CosyVoice3-0.5B-2512 (base and RL)YesDec 2025README
CosyVoice 3.5cosyvoice-v3.5-flash, cosyvoice-v3.5-plusNo — Alibaba Cloud API onlyNot statedAlibaba Cloud

CosyVoice3 model sizes: 0.5B, 1.5B and 3.5

Which CosyVoice 3 models you can download, which exist only in the paper, and which are hosted-only. Each row cites where the size and availability come from.

CosyVoice3 model sizes: 0.5B, 1.5B and 3.5
ModelSizeHow you can use itSource
Fun-CosyVoice3-0.5B-2512 (base and RL)0.5B parametersOpen weights under Apache-2.0: GitHub, Hugging Face, ModelScopeModel card
CosyVoice 3 research model1.5B parametersPaper only — not released as open weightsarXiv
cosyvoice-v3-flash, cosyvoice-v3-plusNot listed (API model)Hosted on Alibaba Cloud Model Studio: Singapore and China (Beijing)Alibaba Cloud
cosyvoice-v3.5-flash, cosyvoice-v3.5-plusNot listed (API model)Hosted on Alibaba Cloud Model Studio: China (Beijing) onlyAlibaba Cloud

Want to hear the difference? The CosyVoice 3 demo page plays the official CosyVoice 2.0, 3.0-0.5B and 3.0-1.5B samples side by side.

What's new in CosyVoice 3 vs CosyVoice 2

The CosyVoice 3 paper aims at "in-the-wild" speech: natural, speaker-consistent audio across languages and recording conditions, not only clean studio data.

  • A multi-task speech tokenizer

    A new speech tokenizer trained on five supervised tasks — speech recognition, emotion recognition, language identification, audio event detection and speaker analysis — to make prosody more natural.

    CosyVoice 3 paper (arXiv 2505.17589)

  • Reward-model post-training

    A differentiable reward model is used for post-training; the paper notes it also applies to other LLM-based speech models.

    CosyVoice 3 paper (arXiv 2505.17589)

  • 100× more training data

    Training data grows from 10,000 to 1,000,000 hours, covering 9 languages and 18 Chinese dialects.

    CosyVoice 3 paper (arXiv 2505.17589)

  • A bigger model in the paper

    Parameters grow from 0.5B to 1.5B in the paper. The open Fun-CosyVoice 3.0 checkpoint is 0.5B.

    CosyVoice 3 paper (arXiv 2505.17589)

  • Pronunciation control and text normalization

    Correct a pronunciation with Chinese Pinyin or English CMU phonemes, and read numbers and symbols without a separate text front-end.

    QwenAudio/CosyVoice README

  • Bi-streaming and instructions

    Text-in and audio-out streaming with latency as low as 150 ms, plus instructions for language, dialect, emotion, speed and volume.

    QwenAudio/CosyVoice README

For a side-by-side table with benchmark numbers, see our CosyVoice 2 vs 3 comparison and the benchmarks write-up. Instructions are also how you steer emotion; see how to write them in our emotional text to speech guide.

Languages and Chinese dialects

The open Fun-CosyVoice 3.0 model covers 9 languages and 18+ Chinese dialects and accents, and clones voices across languages — for example, English speech in a voice recorded in Chinese.

9 languages
Chinese, English, Japanese, Korean, German, Spanish, French, Italian, Russian
18+ Chinese dialects and accents, including
Guangdong (Cantonese), Minnan, Sichuan, Dongbei, Shaanxi, Shanxi, Shanghai, Tianjin, Shandong, Ningxia, Gansu

Hugging Face: Fun-CosyVoice3-0.5B-2512

Hear dialect voices on our Cantonese & Dialects page, and mixed-language speech on the Multilingual TTS page.

Zero-shot voice cloning

Zero-shot means CosyVoice 3 copies a voice from a short reference recording at generation time, with no training per speaker. The official README lists multi-lingual and cross-lingual zero-shot cloning among its key features.

On CosyVoice.org, voice cloning runs on hosted CosyVoice 3-generation models: upload a sample you have permission to use and generate speech in your browser. Clone a voice

How zero-shot cloning works, and how CosyVoice compares with other open-source models: zero-shot voice cloning guide.

Open-source weights vs hosted models

There are two ways to use CosyVoice 3: download the open weights and run them yourself, or call Alibaba Cloud's hosted models through Model Studio (DashScope).

Open-source weights vs hosted models
CompareOpen weights (Fun-CosyVoice 3.0)Hosted (Alibaba Cloud Model Studio)
ModelsFun-CosyVoice3-0.5B-2512 and its RL variantcosyvoice-v3-flash, cosyvoice-v3-plus, cosyvoice-v3.5-flash, cosyvoice-v3.5-plus
WhereGitHub, Hugging Face, ModelScopeModel Studio WebSocket API and the DashScope Python/Java SDKs
RegionsWherever you run itv3: Singapore and China (Beijing). v3.5: China (Beijing) only
CostFree to download (Apache-2.0); you pay for your own GPUBilled per input character
SetupPython 3.10 environment, model download, your own GPU serverAn API key — no GPU

Download the open weights:QwenAudio/CosyVoice READMEHugging Face: Fun-CosyVoice3-0.5B-2512ModelScope: Fun-CosyVoice3-0.5B-2512

Model differences, official prices and a quick start for the hosted API are in our CosyVoice API guide.

Run CosyVoice3 locally: quickstart

Five commands take a fresh Linux machine from clone to the official example, following the QwenAudio/CosyVoice README. They download the open Fun-CosyVoice 3.0 model, Fun-CosyVoice3-0.5B-2512, from Hugging Face.

Linux, Conda and Python 3.10
git clone --recursive https://github.com/QwenAudio/CosyVoice.git && cd CosyVoice
conda create -n cosyvoice -y python=3.10 && conda activate cosyvoice
pip install -r requirements.txt
python -c "from huggingface_hub import snapshot_download; snapshot_download('FunAudioLLM/Fun-CosyVoice3-0.5B-2512', local_dir='pretrained_models/Fun-CosyVoice3-0.5B')"
python example.py

README: Install QwenAudio/CosyVoice example.py

That is the short version. The full steps (sox fixes, the optional ttsfrd package, the web UI, Docker, Windows notes and ComfyUI) are in our guide to how to install CosyVoice locally. If you would rather not set anything up, the free online tool on this site runs hosted CosyVoice 3-generation models.

CosyVoice3 on GitHub and Hugging Face

The official code is at github.com/QwenAudio/CosyVoice. The project moved there from FunAudioLLM/CosyVoice, and the old address redirects to the new one. The open Fun-CosyVoice 3.0 weights, Fun-CosyVoice3-0.5B-2512, are on Hugging Face (under FunAudioLLM) and on ModelScope.

QwenAudio/CosyVoice READMEHugging Face: Fun-CosyVoice3-0.5B-2512ModelScope: Fun-CosyVoice3-0.5B-2512

For the clone, submodule and model-download commands, see the CosyVoice GitHub install guide.

Deploy CosyVoice3: vLLM and OpenVINO

Besides the plain Python example, three upstream sources describe other ways to run CosyVoice3. We link each one and describe what it contains; we have not benchmarked any of them.

  • vLLM (official README)

    The README says CosyVoice2/3 support vLLM 0.11.x+ (V1 engine) and 0.9.0 (legacy), and lists the environment setup and the vllm_example.py script. The commands are copied on our install page.

    README: vLLM Usage

  • vLLM-Omni (vLLM project)

    The vllm-project/vllm-omni repository has CosyVoice3 examples for Fun-CosyVoice3-0.5B-2512: an offline-inference script, and an online-serving setup with a launch script and a client. It is maintained by the vLLM project, not by the CosyVoice team.

    vllm-project/vllm-omni: CosyVoice3 examples (GitHub)

  • OpenVINO (Intel notebook)

    The openvino_notebooks repository has a notebook, “Text-to-Speech (TTS) system with Fun-CosyVoice 3.0 and OpenVINO”. It converts the model to OpenVINO format, builds an inference pipeline and launches a demo. Its README labels it experimental and says the model is not fully validated with OpenVINO.

    openvino_notebooks: Fun-CosyVoice 3.0 notebook (GitHub)

We list only setups that appear in the sources above. Docker, Windows and ComfyUI are covered in the install guide.

Try CosyVoice 3 without installing anything

Type a line in the free online tool, or listen to the official demo sentences compared across CosyVoice 2.0, CosyVoice 3.0-0.5B and CosyVoice 3.0-1.5B.

Try CosyVoice free →Listen to the CosyVoice demo →

CosyVoice 3 FAQ

What is the difference between CosyVoice 3.0 and CosyVoice 3.5?

CosyVoice 3.0 is the open release, Fun-CosyVoice3-0.5B-2512, which you can download and run yourself. CosyVoice 3.5 is a hosted model family (cosyvoice-v3.5-flash and cosyvoice-v3.5-plus) on Alibaba Cloud Model Studio, listed for the China (Beijing) region; Alibaba's docs list voice cloning, voice design and instruction control for both. As of October 2026 there are no CosyVoice 3.5 weights in the official GitHub, Hugging Face or ModelScope repositories.

Is CosyVoice 3 open source?

Yes. Fun-CosyVoice 3.0 (Fun-CosyVoice3-0.5B-2512) is published with its training and inference scripts under the Apache-2.0 license. The 1.5B model described in the CosyVoice 3 paper has not been released as open weights.

Is cosyvoice3 the same as CosyVoice 3?

Yes. cosyvoice3, CosyVoice 3.0 and Fun-CosyVoice 3.0 all refer to the third generation; Fun-CosyVoice 3.0 is the name the official repository uses for the open release.

Can I use CosyVoice 3 commercially?

The open model is released under Apache-2.0, but the README also says its content is for academic and demonstration purposes — read both and get your own legal advice. On CosyVoice.org, commercial use is included in the Pro plan.

How much GPU memory does CosyVoice 3 need?

The official README and model card don't state a minimum GPU or VRAM requirement, so test on your own hardware. If you'd rather not run a GPU, use the hosted version on this site or Alibaba Cloud Model Studio.

How many languages does CosyVoice 3 support?

9 languages — Chinese, English, Japanese, Korean, German, Spanish, French, Italian and Russian — plus 18+ Chinese dialects and accents, according to the official model card.

What are the CosyVoice3 model sizes: 0.5B or 1.5B?

The open Fun-CosyVoice 3.0 release, Fun-CosyVoice3-0.5B-2512 (base and RL), has 0.5B parameters. The CosyVoice 3 paper also reports a 1.5B model, which has not been released as open weights. CosyVoice 3.5 is offered only through Alibaba Cloud's API, and no parameter count is listed for it.

Is there a free online CosyVoice 3 demo?

Yes. The free online tool on CosyVoice.org runs hosted CosyVoice 3-generation models, so you can type a line and hear it in your browser. Our demo page also plays the official CosyVoice 2.0, 3.0-0.5B and 3.0-1.5B samples side by side, with nothing to install.

Is CosyVoice 3.5 open source?

Not as far as we can find. The open release is Fun-CosyVoice 3.0 (Fun-CosyVoice3-0.5B-2512, Apache-2.0). CosyVoice 3.5 is offered as hosted models on Alibaba Cloud Model Studio, and as of November 2026 the official FunAudioLLM page on Hugging Face and the QwenAudio/CosyVoice repository list no CosyVoice 3.5 weights.

Where is the CosyVoice3 GitHub repository?

github.com/QwenAudio/CosyVoice. The project moved there from FunAudioLLM/CosyVoice, and the old address redirects. The weights are on Hugging Face and ModelScope. Our install guide has the clone command and the submodule step.

Can CosyVoice3 run with vLLM or OpenVINO?

The README says CosyVoice2/3 support vLLM (0.11.x+ on the V1 engine, and 0.9.0 legacy). The vLLM project’s vLLM-Omni repository also has CosyVoice3 examples, and an Intel notebook converts Fun-CosyVoice 3.0 to OpenVINO (labelled experimental). Links are in the deploy section above; we have not measured their speed.