Qwen-Music Music Generation Model Report
Qwen · hf · 2026-07-20
Alibaba released a technical report for Qwen-Music, a music generation model capable of producing high-quality songs with full vocal performances. It supports two types of tasks: text-to-music generation, and covering existing songs with different styles or timbres.
The model architecture consists of three parts: Qwen-Music-Tokenizer compresses audio into 25Hz single-codebook music semantic tokens; Qwen-Music-LLM performs autoregressive generation based on these tokens, introducing Melody-CoT to plan the melody before generating the entire song; and Qwen-Music-Render handles generative stereo rendering to fill in audio quality details. For training, the team used over 5 million hours of multilingual music data, combined with quality-aware pre-training, supervised initialization, offline DPO, and online GSPO. The report claims Qwen-Music achieves SOTA in 13 out of 16 objective music/audio metrics and is preferred by professional reviewers. Compared to mainstream closed-source systems, it also better preserves reference melodies during covers.
More from Models
- Local benchmark: Flint matches Qwen3.6 quality at 50/52 while running 3.6x faster on M4 Pro — RichardSocher · 2026-10-27
- ICML 2026 position paper: LLM 'thinking traces' are a harmful anthropomorphism — rao2z · 2026-10-27
- One stray 'u' and Claude starts spelling like a Brit — cantrell · 2026-10-27
- Cohere's Invent-A-Dataset paper: zero-seed synthetic data beats frontier APIs on quality and diversity — idanbeck · 2026-10-27
- ARC Prize redesigns leaderboard: highest reasoning effort by default, CSV export — mhmazur · 2026-10-27
- Cohere: open-weight Qwen3.8-Flash-Next tried uploading probe file to Beijing cloud bucket in sandbox — wstone_bd · 2026-10-27