News · 2026-09-11

YuE2 is an open song generator that writes the sheet music first

The open-source research collective Multimodal Art Projection (M-A-P) released YuE2 on 9 September 2026, a song generator that first writes the melody and chords as an editable score and then renders the full song with vocals. The main model has about 3.6 billion parameters and is a 7.3 GB download, released under a non-commercial licence. It arrived the same day Suno, the best-known commercial AI music service, launched v6, a model family it says was "developed with our industry partners, including Warner Music Group, BMG and Believe."

Key facts

• 7.3 GB for the main YuE2-3B weights, plus a 0.5 GB audio decoder; the model card asks for a 24 GB NVIDIA GPU, and the team measured a peak of about 11 GiB on an RTX 4090.
• On that card, YuE2 generated a 3.6-minute song in 71 seconds.
• Weights are licensed CC BY-NC 4.0 (non-commercial); the code is Apache 2.0. The team says it trained "primarily on CC0 music and synthetic data."
• Primary sources: the model on Hugging Face, the project page and the code on GitHub.

Writing the lead sheet before recording

Most AI song generators go straight from a text prompt to audio. YuE2 adds a step a human musician would recognise. It first writes a score in ABC notation, a plain-text format for sheet music, containing the melody and chord progression. You can read it, change a chord or rewrite a phrase, and then ask the model to render audio that follows the edited plan. It is like a composer writing a lead sheet before the band goes into the studio: you can fix the bridge on paper instead of re-recording the whole take.

Under the hood, a single model predicts a compact "semantic" description of the audio from that plan, then fills in the acoustic detail using flow matching, and a decoder of the autoencoder type turns the result into stereo sound. The project is a collaboration led by M-A-P with HKUST and several partners, including Tokenwave.AI, which the team says provided most of its synthetic training data under licence. The technical report is listed as "coming soon," so the full method is not yet published.

The training-data question

For AI music, where the training data came from is the central legal question. Suno has faced lawsuits and court rulings over its models, including a Munich court finding that it memorised six songs and a Canadian rights society suing over 150 outputs. YuE2's answer is to train mainly on public-domain (CC0) recordings and synthetic music made under licence, some 346,000 hours in all, according to the project page.

Suno's answer is licensing deals. In its v6 announcement, chief executive Mikey Shulman says the new models were developed with Warner Music Group, BMG and Believe, and that "as v6 rolls out, we will retire our previous models." According to TechCrunch, the company said v6 is "not trained using the data it used to train previous versions." Suno's post does not itself say the models were trained only on licensed music, despite how some coverage has summarised it.

Does it beat Suno?

A popular Reddit thread says YuE2 "beats Suno v6." The team's own evidence is more modest. On 192 prompts scored with automatic measures, YuE2's best configuration ranked at the top, but that configuration picks the best of eight attempts, and the team describes the model as competitive with Suno rather than clearly better. No independent listening test has been published.

Why it matters

The anchor number is the hardware: a full song in about a minute on a single gaming GPU, from a model anyone can download. That puts serious music generation within reach of hobbyists and researchers. ComfyUI, the popular open-source generation tool, has an open pull request to add native YuE2 support.

The honest caveats: the non-commercial licence limits what users can do with the weights, the technical report has not been published, and the quality comparison rests on the team's own automatic measures. Whether a public-domain-plus-synthetic training set avoids every legal question is for courts, not model cards, to decide.

Primary source, verified: read the paper →

Ground Truth. (2026, September 11). YuE2 is an open song generator that writes the sheet music first. Ground Truth. https://groundtruth.day/news/yue2-is-an-open-song-generator-that-writes-the-sheet-music-first.html

@misc{groundtruth:yue2-is-an-open-song-generator-that-writes-the-sheet-music-first,
  title  = {YuE2 is an open song generator that writes the sheet music first},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/yue2-is-an-open-song-generator-that-writes-the-sheet-music-first.html}
}

Topics: open-weights · music-generation · generative-audio · suno · copyright · model-release