Qwen-Music Report: 5M+ Training Hours, SOTA on 13 of 16 Metrics

A technical report posted to arXiv on July 13 introduces Qwen-Music, a text-to-music and cover-song system whose authors say it achieves state-of-the-art results on 13 of 16 objective musicality and audio-quality metrics, evaluated across 600 Chinese and English prompts.

The architecture arrives in three pieces. A tokenizer compresses audio into what the paper calls a 25 Hz single-codebook stream of Music Semantic Tokens. An autoregressive LLM uses a melody-token-based chain-of-thought mechanism to plan melodies before generating full songs. A separate renderer then produces high-fidelity stereo waveforms. Training is reported at more than 5 million hours of multilingual music covering hundreds of languages, with a recipe that combines a quality-aware pre-training curriculum, supervised initialization, offline DPO and online GSPO.

Why the scale figure matters is straightforward. Music generation has so far been dominated by proprietary systems with opaque training data and no published methodology. A technical report attached to a well-known model family, with a concrete architecture and a specific training-hours claim, moves the argument out of demo reels and into something other labs can reason about.

The honest caveat is that the abstract does not name the leading proprietary systems the paper's professional evaluators reportedly preferred it over, does not enumerate which of the sixteen metrics were topped and which were not, and does not give evaluator counts or methodology detail for the human preference test. Take the specifics as reported by the authors, not as settled by independent replication.

The piece worth watching is what happens if weights, the tokenizer, or the 600-prompt evaluation set ever ship publicly. That would give researchers and smaller music-tech teams a serious baseline to build against, and would raise the bar on incumbents whose moat has been closed data and closed models.