#1 on WildSongBench
WildSongBench is a 192-prompt automatic song benchmark. On SongBench average, YuE2-3B best-of-8 (pick the best of eight generations) ranks first. A single YuE2 run does not beat Mureka 9 or Suno v5.
This page is published by Noiz. YuE2 is co-developed by Noiz. Scores and the figure are transcribed from M·A·P; Noiz did not re-run the bench here.
WildSongBench song-quality index versus text-alignment index. Bubble area is AudioBox PQ. Outlines mark Pareto optima.
| System | Setting | SongBench Avg |
|---|---|---|
| YuE2-3B | best-of-8 | 6.9632 |
| Mureka 9 | delivered | 6.9377 |
| Suno v5 | delivered | 6.8721 |
| YuE2-3B | single run (2 candidates) | 6.7316 |
| Suno v6 | 2-candidate protocol | 6.5562 |
Automatic metrics transcribed from M·A·P, results dated 12 Sep 2026. Best-of-8 selects by SongBench Musicality, then prompt control, then phoneme error rate. Gaps are descriptive, not statistical significance. On SongEval average, Suno v6 scores 4.3086, above YuE2 best-of-8 at 4.2960 and a single YuE2 run at 4.2625. Full results YuE2-3B weights Figure PDF