Open-source music generator YuE2 tops SongBench, beating Suno v5 with symbolic planning
solyarisoftware · x · 2026-09-10
YuE2 is a new open-source music generation model that unifies symbolic and audio generation: it first writes an editable ABC symbolic score, then renders vocals and accompaniment, supporting creation, covering, and conversational agentic editing.
Key facts:
- 3.59B parameters, 28 layers; AR and NAR experts share attention with separate normalization
- On WildSongBench (192 prompts, 15 evaluated settings), YuE2 best-of-8 scores 6.9632 — the highest observed mean, beating Suno v5's 6.8721, as well as Mureka v9 and MiniMax 3
- Symbolic plans let users inspect and edit melody, rhythm, and chords before rendering
- Built by M·A·P, Tokenwave.AI, MBZUAI, and ACE STUDIO, with interactive demos and scores
Related event: YuE2-3B Open-Source Music Model Launches, Claims to Rival Suno(9 posts)→
More from Multimodal
- Reusable prompt template: Watercolor Splash Fauvism style with two pure colors — LudovicCreator · 2026-09-10
- GPT6 + 3D generation builds a stunning foldable iPhone concept web demo — vista8 · 2026-09-10
- 20-Year Veteran Editor: GPT-6 Astra Edits in DaVinci via MCP, "Better Than Me" — petewoodbridge · 2026-09-10
- Watermark-free generated iPhone renders handy for cutouts and product shots — songguoxiansen · 2026-09-10
- Hot take: LTX 2.5 beats MiniMax H3 on speed, HDR 50fps video and LoRA compatibility — Straight_Hat1304 · 2026-09-10
- Gen2Balance: text-to-video synthesis fills long-tailed data for SOTA action recognition — dimadamen · 2026-09-10