YuE2 renders audio from a score you can edit first

The system

YuE is an open foundation model family for lyrics-to-song generation released 26 January 2025 by a research collective spanning HKUST, M-A-P and several partner institutions, documented in the original repository and a technical report submitted that March. Its successor, YuE2, described in the project's current main repository, adds a distinct step: it first plans a song as an editable symbolic melody-and-chord score, then renders that score into full-band audio with vocals, rather than generating audio directly from a prompt.

What the documents establish

The original YuE report states the family is built on the LLaMA2 architecture, scaled to trillions of tokens, and generates up to five minutes of music using track-decoupled next-token prediction to keep vocal and accompaniment signals from interfering during generation; the project reports YuE's learned representations also perform competitively on the independent MARBLE music-understanding benchmark. The current YuE2 repository describes a different backbone, which it calls an AR–NAR Mixture-of-Transformers: one model predicts the symbolic score and semantic tokens autoregressively, then a separate stage generates acoustic detail through flow matching before a decoder produces stereo audio. On evaluation, the repository reports YuE2 as competitive with Suno's v5 and v6 models on its own WildSongBench benchmark, a claim the same document immediately qualifies by noting the gap between top-scoring systems does not establish statistical significance.

Craft and rights

The editable symbolic score is the capability worth isolating for a working musician: unlike a pure audio-diffusion or token-prediction system, YuE2's documented pipeline lets a person inspect and change the melody and chords before the final render, which is a meaningfully different consent-and-control point in the workflow than approving or rejecting a finished waveform. On rights, the current repository states the code is Apache 2.0 licensed while the model weights carry a CC BY-NC 4.0 licence with an added creator permission, and that companies seeking commercial use must contact the project directly — terms a user should read in full before assuming open access extends to commercial release.

Outcomes and open questions

Every comparative benchmark figure in these documents, including the Suno comparison, is the project's own self-reported result on its own benchmark and evaluation protocol, not an independent test; the repository's own caveat about statistical significance is a useful check against over-reading a single leaderboard position. Readers should watch whether third-party evaluations of YuE2 against commercial systems appear, and whether the commercial-licensing terms change as the project matures.

• Does an editable symbolic score change who is meaningfully the author of the final recording?
• How much weight should a self-reported benchmark carry against a system's own named competitors?
• What exactly does the model-weights licence permit before you would need to contact the project directly?

YuE2's architecture is a genuine departure from the direct-to-audio systems earlier in this lineage, but its performance claims against named commercial competitors remain, on these documents, a claim the project makes about itself.