whatmodel.sh / tts / Fun-CosyVoice3-0.5B-2512
Fun-CosyVoice3-0.5B-2512
↗ share
+ contribute
Fun-CosyVoice3-0.5B-2512 is an open-weights text-to-speech model from Alibaba (FunAudioLLM), released December 2025. Its published size is 0.5B parameters. Licence: apache-2.0.
Catalogue identity notes
FunAudioLLM/Fun-CosyVoice3-0.5B-2512, the only released CosyVoice 3 weights (the 1.5B model is not open). Default llm.pt assumed where a source says 'default settings'; the repo also ships an optional RL-tuned llm.rl.pt. Distinct from CosyVoice2-0.5B and the hosted CosyVoice v3 Plus API.
Score within the Vapi Humanness · Bradley–Terry rating, 2026-09-25 snapshot comparison set only.
$ how it compares
The text-to-speech models ranked either side of Fun-CosyVoice3-0.5B-2512, out of 148 with a score, then more from Alibaba (FunAudioLLM).
| # | model | score | where | ||
|---|---|---|---|---|---|
| 26 |  VoxCPM 0.5B | 64±1 | local | ||
| 27 | VoxCPM2 | 64±1 | local | ||
| 28 | Fun-CosyVoice3-0.5B-2512 | 64±1 | local | ||
| 29 |  Wayu-Paxa-TTS-Edge | 64±4 | local | ||
| 30 |  Eleven Turbo v2.5 | 64±4 | api $50.00/M chars | ||
| more from Alibaba (FunAudioLLM) | |||||
| 49 | CosyVoice2-0.5B | 61±1 | local | ||
| 68 | CosyVoice-300M | 54±3 | local | ||
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
voice cloning▁·▁▁▁▁▁▁▇█▃▃ #20 of 37 · 28% of its score · 4 tests
Beats Gemini 3.1 Flash TTS Preview here, though they rank higher overall. Trails DMOSpeech 2, Qwen3-TTS 12Hz 0.6B Base, ZipVoice here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Seed-TTS test-en · speaker similarity (WavLM-large SV) +1 | 69.5% ✎ | █·█·██·▄·▄·▄ beats 4/10 | 13% |
| Seed-TTS test-zh · speaker similarity (WavLM-large SV) | 77.5% ✎ | ▃··▃█▅▅··▃·▃ beats 8/10 | 7.0% |
| Seed-TTS test-en · Whisper large-v3 WER +1 | 2.1% ✎ | ▃··▃▅··▅·█▃▅ beats 4/11 | 5.6% |
| Seed-TTS test-zh · Paraformer CER | 1.1% ✎ | ▂··▄··▃▂█·▄▂ beats 3/11 | 2.5% |
other tests10 not in a task · 71% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| CV3-Eval hard-en · WER +1 | 11.1% ✎ | █··██·██████ beats 1/8 | 14% |
| CV3-Eval hard-en · speaker similarity (ERes2Net SV) +1 | 74.3% ✎ | ▄█▄·▄··▄··▄█ beats 8/8 | 11% |
| CV3-Eval EN (zero-shot English) · WER +1 | 5.0% ✎ | ▃·······▃▅█▅ beats 6/8 | 11% |
| CV3-Eval ZH (zero-shot Chinese) · CER | 4.1% ✎ | █▄···█▄·█··▄ beats 0/8 | 7.6% |
| CV3-Eval hard-zh · CER | 7.4% ✎ | ▃▃·▃··▃▃█··▃ beats 8/8 | 6.1% |
| CV3-Eval multilingual voice clone · 9-language mean error lab | 7.8% ✎ | ▄···▄··▄··██ beats 1/6 | 5.2% |
| CV3-Eval hard-en · DNSMOS (predicted quality) | 3.98 / 5 ✎ | ▄·····█·▄▄·█ beats 6/6 | 5.0% |
| Raon-OpenTTS-Eval · overall speaker similarity (WavLM) | 64.7% ✎ | ▄·█·█▄·····▄ beats 5/6 | 4.8% |
2 more tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Raon-OpenTTS-Eval · human speaker-similarity MOS (overall) | 3.58 / 5 ✎ | █··██······▄ beats 2/6 | 4.4% |
| Raon-OpenTTS-Eval · overall WER (4 acoustic conditions) | 4.4% ✎ | ▄··▄█····▄·█ beats 4/6 | 2.4% |
$ where to run it
Other ways to run it
- Run it as published · needs ~5.42 GB · from huggingface.co/FunAudioLLM (official)
git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git && cd CosyVoice && git submodule update --init --recursive && conda create -n cosyvoice -y python=3.10 && conda activate cosyvoice && pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host=mirrors.aliyun.com && 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 -c "import sys; sys.path.append('third_party/Matcha-TTS'); from cosyvoice.cli.cosyvoice import AutoModel; import torchaudio; cosyvoice = AutoModel(model_dir='pretrained_models/Fun-CosyVoice3-0.5B'); [torchaudio.save('zero_shot_{}.wav'.format(i), j['tts_speech'], cosyvoice.sample_rate) for i, j in enumerate(cosyvoice.inference_zero_shot('CosyVoice is undergoing a comprehensive upgrade, providing more accurate, stable, faster, and better voice generation capabilities.', 'You are a helpful assistant.<|endofprompt|>希望你以后能够做的比我还好呦。', './asset/zero_shot_prompt.wav', stream=False))]"
Official weights and documentation: huggingface.co/FunAudioLLM ↗
$ how the score was worked out · UltraEval-Audio (OpenBMB) · T… 51%, Raon-OpenTTS: Open Models and… 43%, Qwen-Audio-3.0-TTS technical… 5%
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 UltraEval-Audio (OpenBMB) · TTS replication results | 9 | 51% |
 Raon-OpenTTS: Open Models and Data for Robust Text-to-Speech (arXiv:2605.20830v2) | 9 | 43% |
| Qwen-Audio-3.0-TTS technical report · lab-reported | 1 | 5% |
| prior · pull toward the middle | 1% | |
| weighted average | = 64.0 |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json