YuE2-3B drops as an open-source music generator with editable scores

AI research group m-a-p just dropped YuE2-3B, an open-source music model that turns lyrics and style prompts into complete, studio-quality songs with full vocals and accompaniment.

Most AI music generators treat song creation like an all-or-nothing lottery: you type a prompt, wait for an audio file, and pray the bridge sounds right. If it fails, you reroll the dice. YuE2 takes a completely different road by inserting a symbolic planning phase between your prompt and the final track. It generates an editable ABC score first, outlining melody, rhythm, and chord progressions before synthesizing the actual audio.

What you can do with it

At its simplest, you feed YuE2 a style prompt and lyrics to get a finished song. But the real magic happens when you crack open that intermediate score.

Because the plan lives in plain ABC notation, you can edit it yourself or hand it to an LLM agent. You can reharmonize chords, tweak vocal lines, speed up tempos, or insert an instrumental solo, and then tell YuE2 to render the updated track.

Covers are built-in too. If you pair it with an audio transcription tool like SheetSage2, you can lift the vocal melody from an existing track and feed it into YuE2 under a completely new genre prompt. The team demonstrated this by turning Jingle Bells into heavy metal and reworking Chinese pop ballads into tight jazz-funk.

They also published an agentic editing walkthrough where an AI assistant iterated on a single track across 9 steps and 14 revisions, gradually swapping lyrics, harmonic structures, and lead instruments from Mandopop into modern English jazz.

Specs, performance, and running it locally

Under the hood, YuE2-3B uses a hybrid AR-NAR Mixture-of-Transformers backbone to generate score and semantic tokens, followed by flow matching and a custom VAE to produce 48 kHz stereo audio.

On the WildSongBench evaluation, YuE2's best-of-8 setting hit an average score of 6.9632. That edges past proprietary heavyweights like Suno v5 at 6.8721 and Suno v6 at 6.5562, while comfortably leading open baselines like MiniMax Music 3 and ACE-Step 1.5.

Local requirements are very approachable for a model in this class. Without quantization, an unquantized run on a single 24GB RTX 4090 synthesizes a 3.6-minute track in about 71 seconds, with peak VRAM hovering around 11.2 GiB. You will want at least 24GB of host RAM.

The weights and inference pipeline are up on Hugging Face under m-a-p/YuE2-3B, installable via standard pip wheels. If you do not want to set up an environment right away, the team also launched a blind listening arena where anyone can vote on model outputs without an account.

Why it matters

For indie musicians, game developers, and creators tired of wrestling with black-box audio generators, YuE2-3B changes the paradigm from blind gambling to actual arrangement. Being able to touch the underlying score means AI music can finally fit into existing DAW workflows instead of standing outside them.

We think this is one of the smartest structural moves in open generative audio this year. Giving users an inspectable, symbolic bridge between text and sound solves the biggest control bottleneck in music generation. If you have an RTX 4090 sitting under your desk, it is well worth cloning the repo this weekend.