YuE2
A song-generation model from m-a-p: given a style prompt and lyrics, it generates a full song with vocals and instrumentation, up to about six minutes. Known here secondhand, through yue2-studio, a Windows app built around it — this page carries what that app's README says about the model, not a reading of m-a-p's own model card.
What's known
- 3B-parameter backbone plus a separate VAE, distributed as safetensors and, via Serveurperso's
quantizations (
Serveurperso/YuE2-GGUF), as GGUF at BF16, Q8_0, Q6_K and Q5_K_M. - Runs through yue2.cpp (
ServeurpersoCom/yue2.cpp), a native C++/CUDA inference port — the model itself doesn't require Python at inference time in that path. - CC BY-NC 4.0: non-commercial use only, unlike the MIT-licensed app and inference code around it.
- Takes a style prompt and lyrics; yue2-studio adds an ABC-notated, editable score on top and can feed it a melody transcribed by sheetsage2 for cover generation.
What isn't known
No independent benchmark of YuE2's output quality turned up in this ingest — the only performance figure anywhere in the source is yue2-studio's own render-speed number (46 seconds for a 3:38 song on an RTX 4090), which says nothing about quality. m-a-p's Hugging Face profile is credited with the MERT encoders and the original YuE alongside YuE2, but this ingest did not confirm whether YuE2 shares a paper with YuE (arXiv 2503.08638, cited on sheetsage2's page) or has its own — left open rather than guessed.
Re-checked 2026-10-11: YuE2 has its own paper, the m-a-p technical report arXiv 2609.33757 (27 September 2026). It reports a SongBench Global Avg of 6.73 on its WildSongBench prompts (6.96 best-of-8) and expert listening preferences against Suno, and it concedes that Suno v6 leads on SongEval and phoneme error rate. Those are the authors' own numbers on a benchmark they built. No outside evaluation has appeared, so the first sentence of this section still holds.
Related
yue2-studio · multimodal-art-projection · sheetsage2 · musicgen