Fun-CosyVoice 3
Alibaba Tongyi's Apache 2.0 0.5B zero-shot TTS: 9 languages, 18+ Chinese dialects, voice cloning and streaming from 150 ms.
In short. Fun-CosyVoice 3 reads text aloud in nine languages and more than 18 Chinese dialects, and it can clone a voice from a few seconds of sample audio. It streams, so speech can start while the text is still arriving, with latency as low as 150 milliseconds. At 0.5B parameters and with an Apache 2.0 license, it is easy to run and to use commercially.
Summary
The speech team of Alibaba's Tongyi Lab open-sourced Fun-CosyVoice3-0.5B-2512 on 15 December 2025 under the Apache 2.0 license. It is the third generation of CosyVoice and is described in the paper "CosyVoice 3: Towards In-the-wild Speech Generation via Scaling-up and Post-training" (arXiv 2505.17589). It is an LLM-based synthesizer in which a language model produces speech tokens and a flow-matching model and vocoder turn them into audio. The paper describes a new speech tokenizer trained on several tasks (speech recognition, emotion recognition, language identification, audio-event detection and speaker analysis), a differentiable reward method for post-training, and training data scaled to one million hours across 9 languages and 18 Chinese dialects. The paper also covers a 1.5B model, but Tongyi's announcement and the repository release only the 0.5B one.
The card lists 9 languages (Chinese, English, Japanese, Korean, German, Spanish, French, Italian and Russian), 18+ Chinese dialects and accents, and multilingual and cross-lingual zero-shot voice cloning. It supports pronunciation inpainting with Chinese Pinyin and English CMU phonemes, text normalization for numbers and symbols, and instructions for language, dialect, emotion, speed and volume. Its bi-streaming mode accepts streaming text and produces streaming audio with latency as low as 150 ms. The repository holds both a base and an RL post-trained checkpoint.
In the card's own comparison table, the RL version reaches a 0.81% character error rate on its Chinese test set and a 1.68% word error rate on its English test set, against 1.21% and 2.24% for the base version. These are developer-run results. The Hugging Face repository is about 9.8 GB because it also holds the RL checkpoint and ONNX copies; by our count the files for the base model add up to about 5.4 GB.
Good for
- Chinese and Chinese-English voice generation, including regional dialects
- Zero-shot voice cloning from a few seconds of reference audio
- Real-time voice agents that need streaming text in and streaming audio out
- Commercial products, since the license is Apache 2.0
Watch out for
- Only the 0.5B model is released, not the 1.5B model in the paper
- The repository is large, about 9.8 GB, because it contains several checkpoints and ONNX copies
- The benchmark numbers come from the developer's own table
- Cloning a real person's voice needs their consent
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Fun-CosyVoice 3. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Fun-CosyVoice3-0.5B-2512 FP32 | 500M | 5.4 GB | 6.21 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- Best open-source text-to-speech models in 2026
- Open speech models: transcription is open, text-to-speech is mixed
Frequently asked questions
- What is Fun-CosyVoice 3?
- Fun-CosyVoice 3 reads text aloud in nine languages and more than 18 Chinese dialects, and it can clone a voice from a few seconds of sample audio. It streams, so speech can start while the text is still arriving, with latency as low as 150 milliseconds. At 0.5B parameters and with an Apache 2.0 license, it is easy to run and to use commercially.
- Can I use Fun-CosyVoice 3 commercially?
- Yes. Fun-CosyVoice 3 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Fun-CosyVoice 3 open source?
- Fun-CosyVoice 3 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Fun-CosyVoice 3 on my own computer?
- Yes. The smallest variant, Fun-CosyVoice3-0.5B-2512, needs about 6.21 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Fun-CosyVoice 3 released?
- Fun-CosyVoice 3 was released by Tongyi Lab Speech Team on December 15, 2025.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Tongyi Lab: Fun-CosyVoice 3 major upgrade and open-source release (15 Dec 2025) tongyilab.substack.com
- Fun-CosyVoice3-0.5B-2512 model card (Hugging Face) huggingface.co
- CosyVoice repository (GitHub) github.com
- CosyVoice 3 paper (arXiv 2505.17589) arxiv.org
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3-TTS Alibaba Qwen | 910M – 1.92B | – | Apache 2.0 (Permissive) | 3.68 GB Laptop or phone | |
| VoxCPM2 OpenBMB | 2.29B | – | Apache 2.0 (Permissive) | 8 GB Laptop or phone | |
| OmniVoice Xiaomi | 613M | – | CC BY-NC (Non-commercial) | 3.75 GB Laptop or phone | |
| Chatterbox Resemble AI | 500M | – | MIT (Permissive) | 3.69 GB Laptop or phone |