HeartMuLa audio model

HeartMuLa is a model published in january 2026 that can make full songs with multilingual support including but not limited to English, Chinese, Japanese, Korean and Spanish.

Initially released under non-commercial research and educational use only licence, it then became available under the permissive Apache 2.0. Still,be aware that since HeartMuLa’s repo includes code taken from ConversationTTS, which uses a CC BY-NC 4.0 license, it  can’t fully relicense those sections under Apache 2.0 unless the original authors agree.

You can give it style tips, lyrics, or reference audio and it'll generate music.

You can change lots of stuff like the music style for different parts of a song, or ask it to make short tracks for videos. 

The paper says HeartMuLa can make clear, long-form music with control over stuff like lyrics and style.

It's built in layers and combines several tools to take on big music tasks.

When they scale it up to larger sizes like 7B parameters it gives results close to top commercial models.