Qwen-Music generates full vocal songs from text and lyrics
Alibaba's Qwen-Music produces complete songs with vocals from text prompts, trained on over 5 million hours of multilingual music data and outperforming proprietary systems on 13 of 16 objective metrics.
Alibaba has released Qwen-Music, a music generation model that creates complete songs with vocal performances from text descriptions, lyrics, and musical attributes. The system handles two core tasks: generating entirely new songs from prompts and producing cover versions that reinterpret existing tracks with different styles and vocal characteristics. Training drew on more than 5 million hours of multilingual music data spanning hundreds of languages.
The architecture combines three components. Qwen-Music-Tokenizer compresses audio into a 25 Hz single-codebook stream of semantic tokens that preserve melodic information for language model prediction. Qwen-Music-LLM performs autoregressive modeling on those tokens, introducing a melody-token-based chain-of-thought mechanism that plans melodies before generating the full song—a design choice that improves structural coherence and enables melody cloning from reference audio. Qwen-Music-Render then converts discrete semantic tokens into high-fidelity stereo waveforms through generative rendering, addressing the fidelity limitations of purely discrete representations.
Training followed a quality-aware pre-training curriculum, then progressive post-training combining supervised initialization, offline DPO, and online GSPO to improve musicality and instruction-following. Across 600 Chinese and English prompts, Qwen-Music achieved state-of-the-art results on 13 of 16 objective musicality and audio-quality metrics. Professional evaluators preferred its output over leading proprietary systems, and for cover song generation, the model preserved reference melodies more accurately than proprietary alternatives.
The technical report does not specify whether weights will be released publicly or what hardware requirements local inference would demand. Practitioners should watch for a model card on Hugging Face or an official Alibaba blog post clarifying licensing terms and compute footprint—details that will determine whether the system can run outside cloud APIs.
