YuE2 is a frontier music generation framework designed for high-quality, editable song composition. Utilizing an AR–NAR Mixture-of-Transformers backbone, the model autoregressively predicts symbolic scores and semantic tokens before generating acoustic latents via flow matching, which are ultimately decoded into 48 kHz stereo audio by a VAE. A key technical advancement is its "white-box" symbolic planning, allowing users to inspect and modify melody and chord plans (via ABC notation) before final rendering. This enables advanced capabilities such as zero-shot covers—where a transcribed recording is reimagined in a new style—and agentic editing through conversational refinement of scores and lyrics. In benchmarks, YuE2 (best-of-8) achieved a SongBench Avg of 6.9632 on WildSongBench, outperforming models like Suno v6 and Mureka 9. The system supports various workflows, including direct text-to-audio generation and score-conditioned synthesis, and is compatible with ComfyUI and the Maestro toolkit. For optimal performance, the model requires an NVIDIA GPU with 24 GB VRAM and BF16 support.