---
license: apache-2.0
language:
- zh
- en
- ja
- ko
base_model:
- FunAudioLLM/Fun-CosyVoice3-0.5B-2512
---

Improvements comapred to v2:


 - BF16 instead of F16 for native precision
 - Includes full tokenization for speech tokens, theoretical support for text based instead of token based inference.