Introduction

Feel that? The open lane just got heavier. Multimodal Art Projection (M·A·P) dropped YuE2—a frontier music model that doesn’t just spit audio and hope. It plans first: an editable symbolic score, then vocals and accompaniment that bring the composition to life. Research demos hit September 2026, and the claim is bold—song quality that rivals Suno v5, with numbers to match.

If you’ve been waiting for an open stack that thinks like a songwriter and performs like a studio, the pocket just swung open.

The Magic Behind YuE2

YuE2 unifies symbolic and audio music generation in one model (~3.59B parameters, 28 layers). Through symbolic planning it writes melody and chords you can actually inspect and edit (ABC scores), then converts that plan into semantic tokens, acoustic latents, and full-song stereo audio via a VAE decoder.

What you can do:

  • Create — lyrics + style prompt → complete song with vocals and accompaniment
  • Cover — reimagine an existing track with a new arrangement (melody-only planning recommended)
  • Agentic edit — talk through revisions; the agent reshapes score, style, and lyrics, YuE2 re-renders
  • Bring your own score — feed ABC, or plan then edit before synthesis

On WildSongBench (192 prompts), YuE2 best-of-8 hits a SongBench average of 6.9632, ahead of Suno v5 at 6.8721 in their published comparison. Local inference: 48 kHz stereo on a 24GB NVIDIA GPU without quantization—about a 3.6-minute song in ~71 seconds on an RTX 4090 in their HF package numbers.

Partners on the project page include Tokenwave.AI, MBZUAI, and ACE Studio. Companion models MERT2 (music representations) and SheetSage2 (audio→lead sheet) feed the symbolic side. Training is described as primarily CC0 music and synthetic data (Tokenwave.AI for most synthetic data under license)—a clearer ethical pitch than many closed stacks.

Weights are listed as CC BY-NC 4.0 on Hugging Face, so check the license before any commercial drop.

Pro Tip

Start with cot="full" for original songs so you get melody + chord plans you can fix before the final bounce. For covers, use cot="melody" and a clean ABC melody—don’t fight the model with messy harmony you didn’t mean to keep.

How to Get Started with YuE2

  1. Grab a Linux box with Python 3.10+ and a 24GB NVIDIA GPU (BF16).
  2. Install the inference wheel from Hugging Face (m-a-p/YuE2-3B).
  3. Load YuE2Pipeline, pass style + lyrics with cot="full", save FLAC + artifacts (ABC, tokens, latents).
  4. Explore demos and scores on the project page, then iterate covers/edits when you’re ready.

Install + load:

python -m pip install huggingface-hub==0.36.2
hf download m-a-p/YuE2-3B yue2_infer-0.1.5-py3-none-any.whl --local-dir .
python -m pip install ./yue2_infer-0.1.5-py3-none-any.whl

from yue2 import YuE2Pipeline

pipe = YuE2Pipeline.from_pretrained("m-a-p/YuE2-3B", device="cuda")
song = pipe(style="your style prompt", lyrics="your lyrics", cot="full", seed=42)
song.save("song.flac")
song.save_artifacts("outputs/song")

Demo hub: https://map-yue2.github.io/Model: https://huggingface.co/m-a-p/YuE2-3BVAE (default listening): https://huggingface.co/m-a-p/YuE2-Vae

Conclusion

YuE2 is the open music drop that finally treats composition like a plan you can hold—score first, performance second, agentic edits when the groove needs another pass. Frontier claims against Suno v5, local 48 kHz stereo, and a symbolic workflow producers can actually jam with.

Open the demos, pull the weights, and let that unfiltered motion hit your headphones.

—Anabel ♡