AIToolScan

HeartMuLa Overview

HeartMuLa is the official open-source repository for a family of music foundation models — described as the most powerful open-source music generation model of 2026. It generates music conditioned on lyrics and tags with multilingual support, released under Apache 2.0.

• HeartMuLa Music LM: Generates music from lyrics and style tags with multilingual coverage across almost all languages.
• HeartCodec: A 12.5 Hz music codec delivering high-fidelity audio reconstruction for efficient generation pipelines.
• HeartTranscriptor: A Whisper-based model fine-tuned for accurate lyrics transcription from generated or input audio.
• HeartCLAP: An audio–text alignment model that unifies embedding space for music descriptions and cross-modal retrieval.
• Pretrained Checkpoints: Download HeartMuLa-oss-3B-happy-new-year and HeartCodec-oss-20260123 from Hugging Face or ModelScope.
• Lyrics Controllability: The oss-3B release leads open-source models in lyrics controllability and music quality; internal 7B rivals Suno in musicality.
• Local Deployment: Clone the repo, pip install -e ., download checkpoints, and run inference locally with Python 3.10.
• Online Demos: Try generation on Hugging Face Spaces and ModelScope Studios, or explore samples at heartmula.github.io.