Alibaba's Wan Team Ships WanSong v1.0, 25B Pure-Diffusion Music Model Beats Suno V5 and Mureka on PER

Alibaba's WanSong v1.0 technical report drops a ~25B end-to-end pure-diffusion song generator that directly outputs demixed vocals + BGM stems for tracks up to 5 minutes, trained on 6M+ hours of multilingual song data with a novel dual-stem token scheme and DPO+ReFL RLHF alignment. On the WanSong benchmark it posts 7.43% Pronunciation Error Rate vs 22.80% for Suno V5 and 12.7% for Mureka V7.6, and a Musicality score of 5.49 against Suno V5's 4.18. Authored by Binghui Chen, Pandeng Li, Yu Liu and Jingren Zhou.