# CosyVoice 3 — card for AI agents

LLM-based zero-shot TTS with voice cloning in 9 languages and 18+ Chinese dialects, plus streaming and instruction control.

Category: Voiceover (text-to-speech) · maker FunAudioLLM (Alibaba) · checked 2026-10-25

## Price
- Price: Free
- Free tier: Open weights, free to download.

## License
Apache-2.0 — https://github.com/QwenAudio/CosyVoice/blob/main/LICENSE

- commercial: **yes** (Commercial use OK)
  > each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute
  (https://github.com/QwenAudio/CosyVoice/blob/main/LICENSE)
- attribution: **required** (Credit required)
  > You must give any other recipients of the Work or Derivative Works a copy of this License
  (https://github.com/QwenAudio/CosyVoice/blob/main/LICENSE)
- in app: **yes** (Can ship inside your app)
  > each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute
  (https://github.com/QwenAudio/CosyVoice/blob/main/LICENSE)

## Rules to follow
- Apache-2.0 weights (per the model card) and code; commercial use and products allowed.
- When redistributing the weights or code, include the license and keep notices.
- The GitHub README notes its demo content is for academic purposes only; this covers the samples, not the model.

## How to connect
- Download only
- Repository: https://github.com/QwenAudio/CosyVoice
- Website: https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512

## Notes
Hugging Face API, Oct 25, 2026: 258328 downloads, 673 likes; GitHub QwenAudio/CosyVoice: 24127 stars (FunAudioLLM/CosyVoice now redirects there).

## Numbers (AgentGid data)
- Gid Score: 43
- GitHub stars: 24,446

## Lines to paste into your prompt or AGENTS.md

```
- CosyVoice 3 (Voiceover (text-to-speech)). LLM-based zero-shot TTS with voice cloning in 9 languages and 18+ Chinese dialects, plus streaming and instruction control. Price: free; commercial use OK; credit required.
  Connect: download from https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512 (no API).
  Rule: Apache-2.0 weights (per the model card) and code; commercial use and products allowed.
  Rule: When redistributing the weights or code, include the license and keep notices.
  Rule: The GitHub README notes its demo content is for academic purposes only; this covers the samples, not the model.
  License: https://github.com/QwenAudio/CosyVoice/blob/main/LICENSE (checked 2026-10-25; details: https://agentgid.com/assets/cosyvoice-3/)
```

## Links
- AgentGid page: https://agentgid.com/assets/cosyvoice-3/
- Is it free for commercial use? https://agentgid.com/assets/cosyvoice-3/commercial-use/
- Source: https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512
- Source: https://github.com/QwenAudio/CosyVoice
- Source: https://github.com/QwenAudio/CosyVoice/blob/main/LICENSE


## CosyVoice 3: open speech model with voice cloning

CosyVoice 3 is a free AI model from Alibaba's FunAudioLLM team that reads text aloud in a natural voice. It can copy a voice from a short sample clip and speaks 9 languages plus many Chinese dialects. You run it on your own computer, and its Apache-2.0 license allows paid products.

### Where to use it
- A voice copied from a short clip: Give the model a short recording of a voice and its transcript, and it reads any new text in that voice. Use it for your own voice or one you have permission to use.
- Voiceovers in several languages: Make narration for videos, lessons or app tours in Chinese, English, Japanese, Korean, German, Spanish, French, Italian or Russian, even in the same cloned voice.
- Tone, speed and dialect on request: Ask in plain words for a dialect such as Cantonese, a faster pace, an emotion or a louder voice, written into the prompt.
- Spoken replies that start fast: For a voice assistant, feed text in pieces and play the sound while it is still being made, so the reply starts almost at once.
- Fixing hard-to-say words: When a name or term comes out wrong, spell its sound out in the text, with Chinese Pinyin or English phoneme codes.
- Your own speech server: Run the model as a small web service in a container on a machine with an NVIDIA card, and call it from your apps.

### How to start
- Get the code with its sub-modules and set up a Python 3.10 environment with conda, as the README shows.

```
git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git
cd CosyVoice
conda create -n cosyvoice -y python=3.10
conda activate cosyvoice
```

- Download the CosyVoice 3 model files from Hugging Face.

```
from huggingface_hub import snapshot_download
snapshot_download('FunAudioLLM/Fun-CosyVoice3-0.5B-2512', local_dir='pretrained_models/Fun-CosyVoice3-0.5B')
```

- Load the model with AutoModel and call inference_zero_shot (see the voice cloning recipe above), or run example.py.

```
python example.py
```


### License
Apache-2.0: free for personal and paid products, and you may change and share it. Keep the license and notices when you pass the code or model files on.

### Good to know
- The setup is for developers. It means getting the code with git, a conda environment with Python 3.10 and a separate model download.
- The optional text helper package is offered only as a Linux file, so on other systems the default helper is used.
- Japanese text has to be written in katakana first, as the example script notes.
- The sample audio on the project pages is for academic demonstration only; this note is about the demos, not the model license.

Checked 2026-11-10 against: https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512, https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512/raw/main/README.md, https://huggingface.co/api/models/FunAudioLLM/Fun-CosyVoice3-0.5B-2512, https://github.com/QwenAudio/CosyVoice, https://github.com/QwenAudio/CosyVoice/blob/main/README.md, https://github.com/QwenAudio/CosyVoice/blob/main/LICENSE, https://github.com/QwenAudio/CosyVoice/blob/main/example.py

### Mistakes to avoid
- Don't: git clone without --recursive, then importing cosyvoice → Do: git clone --recursive, or run git submodule update --init --recursive, and keep sys.path.append('third_party/Matcha-TTS') (The code relies on the Matcha-TTS sub-module, which a plain clone leaves empty.)
- Don't: pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host=mirrors.aliyun.com copied as is → Do: pip install -r requirements.txt from your normal package index, unless the person chose that mirror (The README command pulls packages from a third-party mirror and marks it as trusted.)
- Don't: Passing only the transcript of the voice sample as prompt_text to CosyVoice 3 → Do: Start the prompt text with 'You are a helpful assistant.<|endofprompt|>' and put the transcript after it (Every CosyVoice 3 example in the docs builds the prompt this way.)
- Don't: Feeding Japanese text in kanji or hiragana → Do: Convert Japanese text to katakana first (The example script says Japanese must be translated to katakana.)
- Don't: Installing vllm 0.10.x for faster CosyVoice 3 → Do: Use vLLM 0.11.x or newer, or 0.9.0, in a separate environment (Versions before 0.9.0 do not work and the ones in between are not tested.)

### Questions
- Can I use CosyVoice 3 in a paid app? Yes. The model card lists the Apache-2.0 license, which allows commercial use; keep the license with copies you pass on.
- Which languages does it speak? Chinese, English, Japanese, Korean, German, Spanish, French, Italian and Russian, plus 18 or more Chinese dialects and accents.
- Do I need an account to download it? No. The Hugging Face model is not gated, so the files download without signing in.
- Can it speak while it is still generating? Yes. It supports streaming text in and audio out, which the team reports at latency as low as 150 ms.

---
AgentGid's catalog is curated: every entry was checked against the maker's official pages (registry, repository, license and pricing pages) on the date shown; a license value shows the quote it stands on wherever AgentGid has one. Usage numbers come from AgentGid's own scan of public repositories set up for AI coding agents. Claims by makers are reported, not tested by AgentGid.
