YuE2

YuE2-3B is an open-source music generation model.

YuE is an open-source music generation model series capable of generating complete music clips with vocal singing and accompaniment based on text prompts or lyrics

Based on lyrics and style prompts provided by the user, Yue2 generates a symbolic arrangement of melodies and chords presented in ABC notation format.

Audio is a time-series signal, and music generation models need to maintain coherence over long time spans; integrating them into a workflow engine primarily designed for images/video requires specialized adaptation in sampling scheduling, memory management, and output format

YuE2 Studio offers three audio modes for users to choose from: Full, Melody, and Direct.

Yue2 is a music generation model that adopts a white-box symbolic planning approach, decomposing music generation into two stages: symbolic arrangement generation and audio rendering.

YuE2 underlyingly uses a 28-layer Mixture-of-Transformers (MoT) architecture, running autoregressive (AR) and non-autoregressive (NAR) expert types in parallel, which share attention computation but use separate normalization, projection, and feed-forward layers.

YuE2's generation process is decomposed into three representations: symbolic tokens (ABC scores), 25Hz dense semantic music tokens, and 25Hz continuous acoustic latents decoded into 48kHz stereo audio by a variational autoencoder.

SheetSage2 performs full-song lead-sheet transcription, converting audio recordings into ABC notation, providing the symbolic supervision target s for YuE2 training.

When directly compared with Suno v5, YuE2's best-of-8 configuration produced nearly balanced preference results.

YuE2's symbolic planning supports three types of interaction: agentic music editing, zero-shot cover, and targeted local editing, and these interactions rely on ABC scores as a human-readable, machine-editable explicit intermediate representation.

On the WildSongBench (WSB) benchmark, YuE2 achieved a SongBench global average score of 6.73, the highest among all public baselines at the time of the report's release; with the best-of-8 strategy the score rose to 6.96, the highest mean observed among all evaluated systems, including commercial systems.

YuE2 is released by the Multimodal Art Projection team (MAP) and adopts a staged generation pipeline that first generates ABC notation scores and then renders them into audio, which the authors call symbolic planning.

YuE2's AR component processes the symbolic and semantic sequences in a causal manner, and the NAR component realizes the acoustic sequence through bidirectional flow matching, with the joint probability expressed as p(s, c, z | y) = p_AR(s, c | y) × p_NAR(z | y, s, c).

In the expert listening evaluation, participants' preference rate was 49.3% for symbolic-planning outputs, 34.6% for non-planning generation, and the remaining about 16% had no preference.

When deploying local AI models, a smoke test using a short audio clip of approximately 30 seconds should be performed first to verify at low cost whether the environment and model are compatible, avoiding crashes after a long generation process

YuE2 is an open-source AI music generation model jointly developed by a Hong Kong team in collaboration with Hong Kong universities

The default model path for YuE2 scripts is the models/月二-3b folder, and the vast majority of errors about models not being found are due to the model being placed in the wrong location

YuE2 provides a dedicated FP16 version to cover users with Turing architecture GPUs that do not support BF16

The YuE2 model itself does not support commercial use, and after a generation is interrupted it does not support resuming from a checkpoint, requiring a restart from the beginning