MuLaCover is a controllable AI cover-song and music-remix model from the HeartMuLa team, the same group behind the heartlib music-generation stack. It transforms a reference song using new lyrics and a text style description, or generates directly from melody and chord MIDI, letting a user keep the original lyrics or replace them while remixing genre, instrumentation, topic, and mood. Rather than copying a reference recording frame by frame, MuLaCover converts it into symbolic melody, harmony, and optional drum conditions, and the model injects that symbolic lead sheet into a pretrained text-to-song backbone through gated adaptive cross-attention. The README's news log dates the standalone mulacover inference package and reference-audio/MIDI workflows to 11 Sep 2026, and the public code and model weights release to 16 Sep 2026. Licensing is split by material: the source code and documentation are Apache-2.0, but the official model weights and any outputs generated with them are restricted to non-commercial use under CC BY-NC 4.0 plus the repository's MODEL_LICENSE, and the Apache-2.0 code license grants no commercial rights to either. Running it locally requires Linux with an NVIDIA GPU, Python 3.10, and ffmpeg/libsndfile1; the maintainers verified Ubuntu 24.04 with PyTorch 2.10.0, CUDA 13.0, and an NVIDIA B300 GPU. A full setup needs the MuLaCover weights, the HeartCodec-oss codec, and the Qwen3-Embedding-0.6B style encoder; reference-audio input also needs a YourMT3 checkpoint and five ChordNet checkpoints. These are all separate downloads: the repo vendors the transcription code but not its checkpoints. Lazy loading is on by default so those models don't all need GPU memory at once. The README reports evaluation only as a figure, not a numeric table: full-song cover generation was scored via an objective evaluation across six systems and a human evaluation across four systems, with 22 participants. A paper is listed as "coming soon." Community support runs through Discord and GitHub Issues; the authors credit the YourMT3 project and Music X Lab's chord-recognition work used in its symbolic-audio frontend.