Back to home

About YuE2 & yue2music.com

Learn about the YuE2 foundation model, symbolic planning architecture, and the mission behind yue2music.com.

Last updated: 2026-09-16

About YuE2 & yue2music.com

YuE2 (Music with a Plan) is a next-generation open-weights foundation model for full-song music generation developed by Multimodal Art Projection (M·A·P) in collaboration with leading academic institutions including HKUST, Tokenwave.AI, NYU, Stanford, and MBZUAI.

yue2music.com is the dedicated interactive web workstation and community portal designed to bring the transformative power of YuE2 directly to musicians, producers, sound designers, and AI researchers worldwide.

The Breakthrough: Symbolic Planning

Traditional AI music generation systems (such as Udio or Suno) operate as "black boxes." When you enter a prompt and lyrics, the model attempts to map text directly into continuous acoustic tokens or spectrogram representations. This creates major hurdles:

  1. Zero Intermediate Controllability: You cannot adjust a single chord in the chorus without regenerating the entire song from scratch.
  2. Harmonic Incoherence: Models often hallucinate dissonant chords, drifting keys, and unstable rhythmic pulses.
  3. Black-Box Frustration: Professional musicians and producers cannot inspect, export, or print the underlying musical composition.

How YuE2 Changes Everything

YuE2 solves this dilemma through Symbolic Planning (Chain-of-Thought for Music):

  • Stage 1 (Symbolic Blueprint): The model first outputs a fully readable, deterministic ABC 2.1 musical score, specifying the exact key signature, time signature, tempo, bar lines, melody notes, and harmonic chord symbols (Dm7, G7, Cmaj7).
  • Stage 2 (In-the-Loop Human & Agent Editing): Before any expensive GPU audio synthesis takes place, users and AI agents can inspect, audit, and reharmonize the score with natural language commands.
  • Stage 3 (Acoustic Performance): Once the blueprint is approved, YuE2's acoustic diffusion and neural audio VAE render the score into 48kHz 24-bit studio-master audio featuring lifelike vocals and rich instrumental accompaniment.

The YuE2 Model Ecosystem

The YuE2 research initiative encompasses a cohesive family of open foundation models:

  • YuE2-3B: The primary full-song generation model unifying symbolic planning with flow matching diffusion.
  • YuE2-VAE: A high-fidelity 48kHz stereo neural audio autoencoder with low reconstruction artifacting.
  • SheetSage2: An all-in-one audio-to-score transcription model capable of hearing any audio recording and transcribing it into editable ABC sheet notation.
  • MERT2 & MERT2-FS: Self-supervised music representation models providing acoustic embeddings across the international MARBLE benchmark.
  • WildSongBench: An open evaluation benchmark consisting of 192 wild, diverse musical prompts for fair, reproducible comparisons.

Research Partnerships

YuE2 is an open research endeavor uniting premier global AI and music labs:

  • The Hong Kong University of Science and Technology (HKUST)
  • Multimodal Art Projection (M·A·P)
  • Tokenwave.AI
  • New York University (NYU)
  • Stanford University
  • Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)
  • NOIZ
  • ACE Studio

Our Mission at yue2music.com

Our goal is to make music creation transparent, controllable, and democratized. We believe artificial intelligence should empower human creativity rather than replace it. By combining the precision of musical theory with the expressive power of neural audio diffusion, we provide artists with an intelligent co-creator that speaks the language of music.

Whether you are writing a jazz ballad, scoring an indie game, experimenting with microtonal scales, or laying down cyber-rock guitar riffs, YuE2 gives you the steering wheel.

Contact & Community