Qwen-Music Technical Report
Jin Xu, Kangdi Wang, Ruibin Yuan, Shun Lei, Xiong Wang, Xize Cheng, Xueyao Zhang, Yang Zhang, Yiheng Chen, Yongqi Wang, Yue Wang, Zhifang Guo, Zihan Liu, Zijian Lin, Dake Guo, Hangrui Hu, Lei Xie, Linhan Ma, Wei Xue, Wenxiang Guo, Xinfa Zhu, Xipin Wei, Yangze Li, Yuanjun Lv, Yuxuan Wang, Yunfei Chu, Zhiyong Wu
cs.SD
2026-07-13
Qwen-Music pairs a 33B LLM with Melody-CoT to plan the tune before a diffusion renderer builds the full sung song, beating Suno V5 and MiniMax Music in professional A/B.
AI music generation has been stuck in an awkward spot for a few years. It produces sound and follows a style, but song structure, vocal phrasing, and the coherence of melody and arrangement still trail professional production. Existing open and commercial systems (Suno, MiniMax) mostly push everything through one end-to-end model, coupling the semantic layer (what to write) with the acoustic layer (how it sounds), so neither can be tuned on its own.
Qwen-Music splits the two: a language model turns text, lyrics, and musical attributes into "music semantics," and a diffusion renderer turns that into audible stereo waveform.
The system is three stages. Qwen-Music-Tokenizer (0.6B, a 24-layer Conformer) compresses 48 kHz stereo audio into a 25 Hz single-codebook stream of semantic tokens with a 32,768-entry codebook at roughly 375 bit/s. Qwen-Music-LLM (33B, initialized from Qwen3.5-Omni) autoregressively generates those semantic tokens. Finally Qwen-Music-Render (a 1.3B diffusion Transformer) turns semantics back into waveform, going through a Spec-VAE spectral latent space and a Band-Mode Refiner that fixes frequency-dependent magnitude and phase.
The design worth pausing on is Melody-CoT. Before generating the full song, the LLM first "thinks" the melody skeleton with intermediate tokens at 6.25 Hz (melody extracted by RMVPE, encoded as relative MIDI), then expands into full tokens. The logic mirrors chain-of-thought in text: lock the high-level structure first, so detail generation has something to follow, and structure stays coherent while creativity stays controllable.
Training uses more than 5 million hours of multilingual music, quality-graded into seven buckets Q1 to Q7 for curriculum learning (warm up on lower-quality tiers, anneal toward the high-quality Q1). Post-training alignment runs SFT cold start, offline DPO, then online GSPO.
In professional rater A/B blind tests, Qwen-Music's win rate is 66.7% over MiniMax Music 2.6, 55.4% over Suno V5, and 50.3% over Suno V5.5 (essentially tied). It ranks third on the Artificial Analysis vocals leaderboard under the name JazzCat.
| System | Win rate vs Qwen-Music |
| MiniMax Music 2.6 | 66.7% |
| MiniMax Music 2.5+ | 59.1% |
| Mureka V8 | 58.3% |
| Suno V5 | 55.4% |
| Suno V5.5 | 50.3% |
On objective metrics it is best in 6 of 7 SongBench dimensions and all 5 SongEval dimensions, with lyric intelligibility (PER) at 6.10, second lowest. Cover generation (restyling an existing song with a new vocal) reaches melody MAE of about 1.44 semitones on real pop songs, better than the comparison.
Across 8 genres it ranks first in Electronic/EDM, Jazz and Blues, Punk and Hardcore, R&B, and Rock; it places second in Hip-Hop/Rap, behind Suno V5.5.
This is "plan before you generate" moved from text to music. Melody-CoT makes the melody skeleton a controllable intermediate instead of letting an end-to-end model emit waveform from a black box. For anyone working on music generation or controllable audio, the decoupling of semantic and acoustic layers means you can iterate and evaluate each separately, and tasks like cover generation and style transfer get a clean attachment point.
Cover fidelity still has a gap: section-level melody MAE around 1.5 semitones means the cover drifts from the original enough to be heard. Commercial rivals stay close: Suno V5.5 is essentially tied overall and leads in Hip-Hop. The subjective scores come from professional raters, but sample size and rater composition are not fully specified, so absolute numbers deserve a discount; the objective metrics (SongBench, SongEval) are themselves model-scored, not human gold labels. And the report says little about the provenance and licensing of the 5-million-hour training set.