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

Read on arXiv

Type: paper
Venue: arXiv / Qwen Team
Year: 2026
Source: arxiv
Access: free
Language: en
Added: 2026-08-14T16:32:00Z
Verified: 2026-08-14T16:32:00Z

Summary

Three-stage system: 25 Hz single-codebook Music Semantic Tokenizer (0.6B Conformer), Qwen-Music-LLM (3B dense Qwen3.5-Omni init with Melody-CoT planning), and Qwen-Music-Render (1.3B DiT + Spec-VAE + Band-Mode Refiner to 48 kHz stereo). LLM trained on >5M hours of multilingual music with a quality-graded curriculum then SFT/DPO/GSPO. On 600 ZH/EN prompts, best on 13/16 SongBench/SongEval/AudioBox-Aesthetic metrics; professional A/B prefers it over MiniMax Music 2.6 (66.7%) and Suno V5 (55.4%), comparable to Suno V5.5 (50.3%). Cover mode reports lower Melody MAE than Suno V5.5/V5 and MiniMax Cover on an AI-generated reference set.

Keywords

qwen-music · text-to-music · cover-song · melody-cot · tokenizer · dit · songbench · suno

Topics

music generation, audio, song synthesis

Research notes

Primary: arxiv abs (default nonexclusive-distrib 1.0, cs.SD). Discord #data-source-dump posted AlphaXiv 2607.11699v1. SOTA and A/B claims are the authors'. External AA Music-with-Vocals listing as JazzCat was not re-checked. No official weights/code linked from the abs page; did not copy an Apache-2.0 claim from secondary writeups.