YuE2, a next-generation open-source music generation system developed by a consortium spanning HKUST, M·A·P, Tokenwave.AI, NYU, Stanford, MBZUAI, NOIZ, and ACE Studio, has achieved the highest recorded SongBench Average score (6.9632) on the WildSongBench evaluation — surpassing all configurations of Suno (v5 at 6.8721, v5.5 at 6.7150, and v6 at 6.5562).
The system's key innovation is symbolic planning: given lyrics and a style prompt, YuE2 first generates an explicit melody-and-chord score in ABC notation, then realizes that plan as a complete 48 kHz stereo song with vocals and accompaniment. This white-box approach means the composition is readable, editable, and playable before any audio is rendered — a departure from purely black-box music generators.
YuE2 uses an AR-NAR Mixture-of-Transformers backbone that predicts the score and semantic tokens autoregressively, then generates acoustic latents via flow matching. The staged Python API exposes plan(), generate_semantic(), synthesize(), and decode() as separate steps.
The system also supports zero-shot covers: transcribing a source recording via the companion SheetSage2 model, then reimagining it in a new style using only the melody score. On 948 cover works, YuE2 achieved 0.647 CLEWS mAP with full-score conditioning, compared to 0.006 without a score.
The code and agent skill are licensed Apache 2.0; model weights are CC BY-NC 4.0. The project accompanies new state-of-the-art music understanding (MERT2) and audio-to-score transcription (SheetSage2) models released alongside.
