What it is
A single music model plans an explicit symbolic score, then renders semantic tokens and full-song audio. It unifies composition, editing, covers, and audio generation in one checkpoint. Instead of generating audio with composition left implicit, it plans a readable score first, then renders it to semantic tokens and audio in one AR-NAR MoT. The same checkpoint also supports score editing and zero-shot covers.
Key results
- 49.3% overall expert preference with planning vs 34.6% without
- 6.73 SongBench Global Avg on WildSongBench, best among public baselines
- Best-of-8 reaches 6.96 SongBench Global Avg, highest observed mean
Explicit symbolic planning improves perceived song quality while preserving a direct path to frontier full-song audio generation.