MuQ
Self-supervised music representation with Mel residual vector quantization; plus the MuQ-MuLan music-text model.
About
MuQ is a 2025 music foundation model that trains a transformer to predict tokens from a novel Mel residual vector quantization (Mel-RVQ) target instead of random projections or neural-codec tokens. The authors report that even the 0.9k-hour open-data version beats prior self-supervised music models across tagging, instrument classification, key detection, and more, with scaling to 160k+ hours improving things further. A sibling model, MuQ-MuLan, adds CLIP-style music-text contrastive learning in English and Chinese and is reported SOTA at zero-shot tagging on MagnaTagATune.
Tencent AI Lab released code, checkpoints, and a pip package (pip install muq) under MIT, with weights on Hugging Face under the OpenMuQ org. It is the newest model in this batch and the most actively developed. The caveat is the usual one for fresh releases: all headline numbers are provider-reported, independent replication has not had time to accumulate, and the scaled 160k-hour checkpoint's training data is proprietary even though the weights are open.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported: SOTA on the MARBLE music-understanding benchmark suite; MuQ-MuLan reported SOTA zero-shot tagging on MagnaTagATune. Numbers are provider-claimed; no independent replication yet.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified, metrics-v1). Evidence: 2025 paper (arXiv 2501.01108) from Tencent AI Lab with university collaborators; MIT-licensed repo tencent-ailab/MuQ (389 stars, recently active); HF OpenMuQ org with checkpoints; PyPI package. Strengths: novel Mel-RVQ objective, permissive license, active development, music-text variant included. Docked for: (1) all benchmarks provider-reported with no independent replication yet; (2) large-scale training data is proprietary. License: MIT.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 5.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · 2025 paper (arXiv 2501.01108), Tencent AI Lab with SJTU/Nanjing University: Mel residual vector quantization (Mel-RVQ) SSL objective; sibling MuQ-MuLan adds music-text contrastive learning (EN+ZH)
- · Repo tencent-ailab/MuQ (389 stars, active); PyPI package 'muq'; HF OpenMuQ org
- · All benchmarks (MARBLE SOTA; MuQ-MuLan zero-shot MTT tagging) provider-reported with no independent replication yet; 160k-hour training data proprietary