YuE2

An open song-generation model from HKUST and Multimodal Art Projection (M-A-P) that composes a readable score before it renders audio. One 3.58B-parameter AR–NAR mixture-of-transformers first writes a lead sheet with melody, chords, key, meter, tempo and form, a step the authors call symbolic composition planning; it then predicts semantic music tokens autoregressively and generates continuous acoustic latents with flow matching, which a VAE decodes to 48 kHz stereo. The model was pretrained from random initialization on about 346,000 hours of music, mostly CC0 recordings and synthetic data (most of it licensed from Tokenwave.AI), packing whole songs into a 24,576-position context, on 64 H800 GPUs. Because ordinary recordings come without aligned scores, the team also built MERT2, a music representation model that leads 14 of 15 MARBLE metrics and supplies the semantic tokens, and SheetSage2, an audio-to-lead-sheet transcriber that leads 12 of 15 benchmark–metric pairs in the authors' comparison and supplies the score supervision; both are released alongside it, with the WildSongBench evaluation set. Because the score is plain text, a person or an external language model can edit it and re-render the song, and the release packages this workflow as an agent skill.

On WildSongBench's 192 prompts, YuE2 scores 6.73 on the SongBench average, above every model with public weights (next best LeVo 2 at 6.32; the January 2025 YuE scores 4.92) and above Suno v4.5, v5.5 and v6. Choosing the best of eight candidates raises it to 6.96, the highest of the 17 evaluated settings, just ahead of Mureka 9 (6.94) and Suno v5 (6.87), though the authors note that the gap between the top means is not statistically significant. In expert listening, planning first won 49.3% of overall-quality judgments against 34.6% for the same checkpoint generating without a score, best-of-eight was preferred over Suno v4.5, and preferences against Suno v5 were nearly balanced. The same checkpoint follows score edits and makes zero-shot covers: conditioned on a source work's score, covers of 948 works held out of training reach 0.647 CLEWS mAP (a cover-identification measure), against 0.006 without a score. Weights shipped on September 9, 2026 (the date used here) under CC BY-NC 4.0 with an added permission for individual creators to monetize their outputs; companies need a commercial license, and the code is Apache 2.0. The technical report followed on arXiv on September 27. The YuE repository, which hosted the original YuE and now serves YuE2, has about 10,600 GitHub stars, and the Hugging Face model has over 1,000 likes. The corresponding authors are Wei Xue and Yike Guo of HKUST; lead author Ruibin Yuan is affiliated with HKUST, ACE Studio and M-A-P, and other contributors come from Tokenwave.AI, NYU, Stanford, MBZUAI and HKGAI.