Open-source YuE2 beats Suno v5/v6 on WildSongBench with symbolic score planning
solyarisoftware · x · 2026-09-15
A repost argues open-source music models have hit an inflection point, with YuE 2.0 matching or beating closed alternatives like Suno, ElevenLabs and Lyria 3.5 in certain capabilities.
Key points:
- Approach: YuE2 unifies symbolic and audio generation — it plans an editable ABC-notation score first, then renders vocals and accompaniment, supporting creation, covers, and agentic conversational editing.
- Specs: 3.59B parameters, 28 layers, from M·A·P, Tokenwave.AI, MBZUAI, and ACE STUDIO.
- Results: best-of-8 scores 6.9632 on WildSongBench (192 prompts), the highest of 17 evaluated settings; Suno v5 scores 6.8721, Suno v6 and v6 Wild score 6.5562 and 6.4195.
- Take: remaining gaps in controllability and coverage are marginal and fixable via local fine-tuning, while closed-source music models show diminishing returns.
Related event: Open-Source Music Model YuE2 Beats Suno v5/v6 on WildSongBench(4 posts)→
More from Multimodal
- LTX-2.5 trains 'surgically precise' LoRAs with rare subject generalization, says user — Aivan-Studios · 2026-11-02
- Synthesia now creates videos inside ChatGPT and Claude chats — synthesiaIO · 2026-11-02
- AI-generated 'mechanic rug' optical illusion goes viral: a handyman living under the coffee table — Delahuntagram · 2026-11-02
- Keeping one face on-model across 16 comic pages with Gemini: what held and what drifted — Money_Parfait8955 · 2026-11-02
- Pika's pika-h3-flex Space Trends on Hugging Face — Pika-Labs · 2026-11-02
- awa-video-reel: free Claude Code skill auto-edits shorts and records website B-roll — TawohAwa · 2026-11-02