What is Fun-CosyVoice3?
Fun-CosyVoice3-0.5B-2512 is the open release of CosyVoice 3, the text-to-speech line from Alibaba's Tongyi speech team, published in December 2025 alongside a reinforcement-learning (RL) tuned variant. At 0.5 billion parameters it speaks Chinese, English, Japanese, Korean, German, Spanish, French, Italian and Russian, plus more than 18 Chinese dialects and accents such as Cantonese, Minnan and Sichuanese.
It clones a voice zero-shot, within a language or across languages, takes spoken instructions on language, dialect, emotion, speed and volume, accepts tags such as [breath], and lets you correct pronunciation with Pinyin or CMU phonemes. It streams text in and audio out with latency the team puts as low as 150 ms. The weights are under Apache 2.0; the Hugging Face page records about 260,000 downloads in the past month.
How much does Fun-CosyVoice3 cost?
| Plan | Price | What you get |
|---|---|---|
| Self-hosted | Free | Apache 2.0 |
| cosyvoice-v3-flash (Model Studio) | $0.13 per 10,000 characters | Singapore region; 10,000 free characters for 90 days |
| cosyvoice-v3-plus (Model Studio) | $0.26 per 10,000 characters | Singapore region; 10,000 free characters for 90 days |
Free under Apache 2.0. Alibaba Cloud's hosted cosyvoice-v3-flash costs $0.13 and cosyvoice-v3-plus $0.26 per 10,000 characters internationally.
Is Fun-CosyVoice3 worth it?
Fun-CosyVoice3 is the strongest open choice for Chinese dialects, and at 0.5B parameters it is small enough for real-time voice agents. Few open models handle Cantonese, Minnan and a dozen-plus regional accents from one checkpoint, and fewer do it under Apache 2.0 with streaming that the team puts as low as 150 ms.
The accuracy numbers back it up on Chinese. On the team's test sets the RL variant records a 0.81% character error rate in Chinese and 5.44% on the hard set, the lowest in its table and ahead of ByteDance's Seed-TTS at 1.12% and 7.59%. English is solid rather than leading at 1.68%. The trade-off is voice likeness: the RL model's speaker similarity in English is 69.5 against 76.2 for Seed-TTS, and the base model, at 71.8, keeps the voice closer while making more errors.
For serving, the repository says CosyVoice 2 and 3 run on vLLM 0.9.0 and 0.11 or later. A TensorRT-LLM runtime with roughly fourfold speed-up is documented only for CosyVoice 2. No VRAM figure is published. Alibaba also sells hosted CosyVoice v3 models on Model Studio at $0.13 and $0.26 per 10,000 characters, or $13 and $26 per million, though Alibaba does not say whether they match these weights.
The catch: there are no misuse safeguards at all. The repository calls its content 'for academic purposes only', but the weights are Apache 2.0, there is no watermark, no consent wording and no guidance on reference clips. The project's roadmap has no entries for 2026, and the open line still stops at 0.5B while Alibaba's hosted CosyVoice has moved to v3.5.
Our take: for a Chinese-market voice agent, Fun-CosyVoice3 is the default to beat; outside Chinese, the comparison is with VoxCPM2 and Qwen3-TTS, and each team should listen before committing.
Where it earns its place
A fintech customer-service team in Guangdong
Answer phone calls in Cantonese and Mandarin with one brand voice
Clone the brand voice once and instruct the model to speak Cantonese or Mandarin per call; stream through vLLM for first audio around 150 ms. Apache 2.0 means no per-minute fee.
A regional media company producing short videos
Narrate the same script in Sichuanese, Shanghainese and Minnan
Pass dialect instructions with a single narrator clip and fix names with Pinyin tags such as [j][ǐ]. The hosted cosyvoice-v3-flash at $0.13 per 10,000 characters is an option if no GPU is free.
An ML engineer building a multilingual voice agent on a budget
Choose between accuracy and voice likeness
Run the RL variant where words must be right (0.81% Chinese error rate) and the base model where the cloned voice matters more (78.0 Chinese similarity against 77.4).
Before you commit
- Setup
- Clone the repository with --recursive, create a Python 3.10 conda environment, install requirements and download the weights with snapshot_download. The optional ttsfrd package improves text normalisation.
- Integrations
- Python API, vLLM 0.9.0 and 0.11 or later, ModelScope Gradio demo. Alibaba Cloud Model Studio sells hosted cosyvoice-v3-flash and cosyvoice-v3-plus.
- Data and privacy
- Self-hosted, so text and reference audio stay local. On Model Studio, data handling follows Alibaba Cloud's terms. The open release carries no consent or watermarking measures.
- Cost at scale
- Free weights; hosted cosyvoice-v3-flash at $0.13 per 10,000 characters makes 50 million characters a month about $650, and v3-plus about $1,300.
- Limits
- Nine languages; no published VRAM or maximum text length; TensorRT-LLM acceleration documented only for CosyVoice 2.
What’s good
- 0.81% Chinese character error rate for the RL variant, ahead of Seed-TTS
- 18-plus Chinese dialects and accents from one checkpoint
- Streaming with latency as low as 150 ms, vLLM support
- Apache 2.0 at 0.5B parameters
What’s not
- No watermark, consent check or misuse guidance
- RL variant trades speaker similarity for accuracy
- No published VRAM figure or 2026 roadmap updates
- Nine languages, fewer than VoxCPM2 or OmniVoice
Best for
Teams building Chinese-language voice agents or dialect content who want a small, commercially usable model they can run themselves.
Fun-CosyVoice3 alternatives
Fun-CosyVoice3: common questions
- Is Fun-CosyVoice3 worth it?
- Fun-CosyVoice3 is the small, fast, Apache-licensed cloner to pick for Chinese and its dialects, and a sound one for streaming agents in eight other languages. We rate it 3.9 out of 5.
- How much does Fun-CosyVoice3 cost?
- Free under Apache 2.0. Alibaba Cloud's hosted cosyvoice-v3-flash costs $0.13 and cosyvoice-v3-plus $0.26 per 10,000 characters internationally.
- What are the downsides of Fun-CosyVoice3?
- No watermark, consent check or misuse guidance; rL variant trades speaker similarity for accuracy; no published VRAM figure or 2026 roadmap updates; nine languages, fewer than VoxCPM2 or OmniVoice.
- Who is Fun-CosyVoice3 best for?
- Teams building Chinese-language voice agents or dialect content who want a small, commercially usable model they can run themselves.
- How much does Fun-CosyVoice3 cost for a team or in production?
- Free weights; hosted cosyvoice-v3-flash at $0.13 per 10,000 characters makes 50 million characters a month about $650, and v3-plus about $1,300.
- Is Fun-CosyVoice3 safe for company data?
- Self-hosted, so text and reference audio stay local. On Model Studio, data handling follows Alibaba Cloud's terms. The open release carries no consent or watermarking measures.
- What are the best Fun-CosyVoice3 alternatives?
- The closest alternatives we have reviewed are Qwen3-TTS, VoxCPM2, IndexTTS-2.5 and OmniVoice.
Answers checked Nov 7, 2026.