YuE2-3B Plans Full Songs with Editable Scores, but Only Its Best-of-8 Results Surpass Suno v5

The M·A·P team has released the [YuE2-3B weights and inference package](https://huggingface.co/m-a-p/YuE2-3B), breaking full-song generation into inspectable, editable intermediate representations. After a user provides a style and lyrics, the model can first produce an ABC score containing the melody and chords, then generate 25 Hz semantic music tokens and acoustic latents, before a VAE decodes them into 48 kHz stereo audio. Unlike black-box music generators that can only be refined by repeatedly revising text prompts, this approach lets an engineer or another agent directly adjust pitch, rhythm, harmony, and song sections before resynthesizing the audio.

According to the [project’s technical page](https://map-yue2.github.io/), the model has approximately 3.59 billion parameters and 28 layers and uses an AR–NAR Mixture-of-Transformers architecture. The autoregressive and non-autoregressive experts share attention computation while retaining separate normalization, projection, and MLP components; the later stage uses flow matching to generate acoustic latents. The interface supports three modes: full melody-and-chord planning, melody-only planning, and generation without a symbolic plan. It also exposes `plan`, semantic generation, acoustic synthesis, and decoding as separate stages, making it easier to build editing or batch workflows. The team says BF16 inference can run on an NVIDIA GPU with 24GB of VRAM.

On WildSongBench, which contains 192 prompts, standard YuE2 achieved an average SongBench score of 6.7316, below Suno v5’s 6.8721. Its score rose to 6.9632 only after selecting the best result from eight candidates based on musicality, prompt adherence, and lyric accuracy. This distinction matters: best-of-8 entails roughly eight times the candidate-generation cost and introduces an advantage from the selector, so it should not be treated as a measure of single-generation capability. The evaluation relies mainly on automated metrics published by the team, and independent listening tests and comparisons at equal compute cost are still lacking.

The model card uses the CC BY-NC 4.0 license, so the model cannot be used directly in commercial products. Inference also requires installing a custom wheel provided in the model repository. The team says the training data comprises approximately 346,000 hours, sourced primarily from CC0 music and licensed synthetic data, but the information currently available is insufficient to audit the dataset composition item by item. Key areas to watch include single-generation quality, the phoneme error rate for Chinese lyrics, real-world generation speed, and whether the community can reproduce the results across different GPUs and decoders.