Skip to content

Fish Agent V0.1 3B

A speech-to-speech model without semantic tokens

Visit official site

What Fish Agent V0.1 3B is

The model is described as a speech-to-speech and text-to-speech model that captures environmental audio detail without semantic tokens, continued from Qwen-2.5-3B-Instruct.

What you can do with it

  • Capture and reproduce environmental audio information
  • Work across 8 languages in speech tasks
  • Use a semantic-token-free architecture
  • Draw on 700,000 hours of multilingual training audio

Getting started

  • Find the model on Hugging Face
  • Read the model detail page
  • Set up the environment from the GitHub guide
  • Download the model files and configure them

Pros & cons

✓ What we like

  • No traditional semantic encoder or decoder is needed
  • Other languages carry about 20,000 hours each

! What to watch out for

  • Released under BY-CC-NC-SA-4.0 for non-commercial use
  • The listed language hours differ by language

FAQ

How many languages are supported?

The material states 8 languages including English and Chinese.

What license applies?

The material states the model and code use BY-CC-NC-SA-4.0.

Last reviewed: 2026-11-15

More AI audio tools tools

View all →

How we review