Suno最强开源对手来了：4090一分钟出歌，把AI音乐从抽卡变成可编辑工程

YuE2, an open-source 3.6B-parameter Mixture-of-Transformers model from HKUST, M-A-P, NOIZ AI, and Stanford, scores 6.7316 on WildSongBench (192 prompts), approaching Suno v5 (6.8721) and Mureka 9 (6.9377) while beating Suno v6 (6.5562), and hits 6.9632 in 8-way settings to outperform all 17 systems. Its core innovation is symbolic planning - generating an editable ABC-notation score before rendering 48kHz audio - enabling Agentic Music Editing where targeted modifications increase pitch matching from 0.83% to 93.75% and chord consistency from 0 to 83.13%, plus Cover Song generation with a 71.3% first-hit rate versus 0.3% without scores. Runs on a consumer RTX 4090 producing a song in ~1 minute, with companion tools MERT2 and SheetSage2 also open-sourced.

why now

开源YuE2模型上线，4090可跑，超Suno v6。

evidence

4 sources carry this story · confidence medium

topic

AI Tech & New Models

source

量子位（公众号）