MuQ / MuQ-MuLan
Extract music representations or match music to English and Chinese text descriptions.
UnderstandingOpen-source code300M / 700M
- Start here
- Try text retrieval on a small library with your own relevance judgments; use MuQ features when you intend to train a downstream task.
- Training data
- The released MuQ encoder uses Million Song Dataset audio; MuQ-MuLan uses music-text pairs. The card does not itemize the full paired corpus.
- Training rights
- MSD metadata and access do not license all source recordings. The paired corpus is not a public track-level clearance ledger.
- License & output terms
- Code: MIT. Released MuQ and MuQ-MuLan weights: CC BY-NC 4.0.
- Limitations
- Non-commercial weights. The authors warn that the released MuQ has different data scale from the paper, so paper results should not be assumed for this checkpoint.
- Technique
- Self-supervised music features / contrastive music-text learning
- Runtime
- Local PyTorch inference, with about 300M parameters for MuQ and 700M for MuQ-MuLan; follow the selected model's preprocessing.
- Parameters
- 300M / 700M
Sources reviewed . Training and rights notes summarize provider disclosures. A model license does not clear every source recording or generated output. Parameter counts refer to the named variant or component, not total runtime memory.