HeartMuLa-oss-3B

HeartMuLa-oss-3B is an open music model with 3 billion parameters from the HeartMuLa Team that creates a song with vocals from lyrics and style tags. Per the model card, it supports Chinese, English, Japanese, Korean and Spanish. It belongs to a family that also includes the HeartCodec audio codec (48 kHz stereo), the HeartTranscriptor lyrics recognizer and HeartCLAP. According to the repository, reference-audio input and finer control are still on the to-do list. The first release came out on 2026-01-14; the authors recommend the later build HeartMuLa-oss-3B-happy-new-year of 2026-02-13. The weights are public on Hugging Face without any acceptance step; the card and the repository state the Apache 2.0 license (the repository notes that the license of the code and all related weights was changed to Apache 2.0 on 2026-01-20). The authors do not state hardware requirements; the repository describes a lazy-loading mode to save memory on one card and advises splitting the models across two cards. The authors offer an official demo Space on Hugging Face; code is on GitHub and the paper is on arXiv. Date of check: 2026-10-18.