What Is YuE2? The Open AI Song Model, Explained

YuE2 is an open-weight AI music model that turns a style prompt and lyrics into a complete song, with vocals and full accompaniment. It was released on September 9, 2026 by m-a-p (Multimodal Art Projection) and HKUST, the same team behind the original YuE, and within a week it was the #1 trending repository on GitHub and later the top trending text-to-audio model on Hugging Face, mostly because of one number: on its creators' WildSongBench benchmark, it posted a higher average score than any Suno model.

What makes YuE2 different isn't only the score, though. It's the way it works. Before it renders any audio, YuE2 writes the song down: a melody and chord plan in standard music notation that you can read, play, and edit. Then it performs that plan. That one design choice is behind almost everything interesting about the model.

Most pages about YuE2 are either the research release itself or a setup tutorial. This page is the plain-language overview: what the model is, where it came from, what it can and can't do, what it needs to run, and how it compares to its predecessor and to ACE-Step.

YuE2 at a Glance

	YuE2
What it is	Open-weight model that generates full songs (vocals plus accompaniment) from lyrics and a style prompt
Made by	m-a-p and HKUST, with Tokenwave.AI, NYU, Stanford, MBZUAI, NOIZ, and ACE Studio
Released	September 9, 2026 (code release v0.1.6 and the YuE2-3B weights)
Model size	Published as YuE2-3B; roughly 3.6 billion parameters and 28 layers
Output	48 kHz stereo audio, full-length songs
Key feature	Writes an editable melody-and-chord score (ABC notation) before rendering audio
Languages	English, Mandarin, and Japanese
License	Code: Apache 2.0. Weights: CC BY-NC 4.0 with an additional creator permission
Official hardware	Linux, Python 3.12, NVIDIA GPU with 24 GB of VRAM
In Song Creator Pro	Windows, GPU with 8 GB of VRAM (NVIDIA; AMD experimental)
Benchmark	6.96 SongBench average on WildSongBench (best of 8), highest of 17 evaluated settings

Who Made YuE2?

YuE2 comes from m-a-p (Multimodal Art Projection), an open research collective, together with the Hong Kong University of Science and Technology (HKUST). The project lists six more contributing institutions: Tokenwave.AI, NYU, Stanford, MBZUAI, NOIZ, and ACE Studio.

It's the second generation of YuE, the lyrics-to-song model the same group released in January 2025. The original YuE is still available on its own branch of the repository; YuE2 replaced it as the main release.

The YuE2 release is really a family of models:

• YuE2-3B: the song generator itself.
• YuE2-Vae and YuE2-Vae-legacy: decoders that turn the model's output into audio. The first is the default for listening; the second is the one used for the published benchmark numbers.
• SheetSage2: a transcription model that turns an existing recording into a lead sheet (melody, chords, key, beats, and sections). It's what makes covers possible.
• MERT2: a music understanding model that SheetSage2 builds on.
• WildSongBench: the benchmark the team used to evaluate YuE2 against other systems.

According to the project page, YuE2 was trained on 346,000 hours of music, primarily CC0 (public domain) music and synthetic data, most of it licensed from Tokenwave.AI. A full technical report is listed as "coming soon".

How YuE2 Works

The project's tagline is "Compose in symbols. Create in sound." Generation happens in two conceptual steps:

1. Planning. YuE2 reads your style prompt and lyrics and writes a score in ABC notation: the vocal melody, the chord progression, and how your lyric sections map onto the song.
2. Rendering. It then turns that score into music tokens, and from those into audio, producing a full recording with vocals singing your words.

Under the hood this is a single model with two kinds of "experts": an autoregressive part that writes the score and a coarse musical sketch token by token, and a second part that fills in the detailed audio using flow matching. A separate decoder converts the result into 48 kHz stereo.

The score step can be switched on or off. The official release calls these planning modes:

Mode	What happens	Best for
Full (default)	Plans melody and chords, and the score is editable	Original songs
Melody	Plans only the melody; the accompaniment is free	Covers, where the arrangement should change
Off	Skips the score and generates directly	Quick experiments

You can also hand YuE2 your own score instead of letting it write one, which is the basis of both editing and covers.

What YuE2 Can Do

Write complete songs. Give it a style description and lyrics with section labels like [Verse] and [Chorus], and it produces a full structured song: intro, verses, choruses, bridge, outro. The official demos run from about one minute to five. Because the song is planned before it's rendered, the structure you write is the structure you get. Our YuE2 prompting guide covers how to write prompts and lyrics it responds to best.

Let you edit the composition. Since the score is a real, readable document, you can change a melody line, swap chords, or rewrite lyrics, and render the song again from the edited score. The developers' own editing study found that targeted melody and harmony edits mostly land as intended while the rest of the song stays close to the original. Note what this is and isn't: editing produces a new recording of the edited song, not a surgical patch on the old audio file.

Cover existing songs. Paired with SheetSage2, YuE2 can transcribe a recording into a melody score and re-perform it in a completely different style. A Mandarin pop song can come back as English jazz; "Jingle Bells" can come back as heavy metal. The melody survives because it's written into the score. This works with the general model, no special cover training. See how to remix a song with YuE2 for the full workflow.

Revise songs with an AI agent. The release includes an agent skill that lets an AI assistant read and edit the score for you. The official demo follows one song through nine rounds of feedback and 14 versions, from Mandarin pop to English jazz with new harmony and a saxophone solo.

Sing in English, Mandarin, and Japanese. The model card tags English and Chinese, and the official demo collection is mostly English and Mandarin with a set of Japanese songs. A few one-off demos in other languages exist, but these three are where it's dependable. The original YuE covered Cantonese and Korean too; YuE2's release doesn't claim them.

What it doesn't do: there's no built-in stem separation, no "revise just the second verse of this audio file" tool, and no official instrumental-only mode (though empty section labels work well, as covered in the prompting guide).

Benchmark Results

The YuE2 team evaluated 17 system settings on WildSongBench, a set of 192 real song requests (94 Chinese, 98 English), using automatic scoring. The headline metric is the SongBench average, the mean of seven quality dimensions. A selection of the results, as of September 12, 2026:

System	Open weights?	SongBench Avg
YuE2 (best of 8)	Yes	6.96
Mureka 9	No	6.94
Suno v5	No	6.87
YuE2 (standard)	Yes	6.73
Suno v5.5	No	6.72
Suno v4.5	No	6.70
Suno v6	No	6.56
Suno v6 Wild	No	6.42
ACE-Step 1.5	Yes	6.01
YuE (v1)	Yes	4.92

The standard YuE2 setting keeps the better of two generations (picked by the lowest pronunciation error rate); best-of-8 picks from eight. Suno v6 and the other open models were given the same two-take treatment.

The team also tested covers: across 948 songs, covers made from the full score were matched back to the original song by a cover-detection system 71.3% of the time on the first try, versus 0.3% when the score was left out. That's the clearest evidence that the score is what carries the melody.

Read these numbers with the caveats the developers themselves give: the benchmark comes from YuE2's own team, the scoring is automatic rather than human listeners, the gap between the top systems is small and not shown to be statistically significant, and other systems lead on some individual metrics (both Suno v6 variants have lower pronunciation error, for example). For a deeper look at the benchmark itself, see what WildSongBench is and how it scores, and for what it means in practice, our YuE2 vs. Suno comparison.

Hardware Requirements: Official vs. Song Creator Pro

	Official YuE2 release	YuE2 in Song Creator Pro
Operating system	Linux	Windows
Setup	Python 3.12 environment, command line	Normal app install, no command line
GPU	NVIDIA with 24 GB of VRAM and BF16 support	8 GB of VRAM or more (NVIDIA; AMD experimental)
System RAM	24 GB available	16 GB recommended
Precision	Full precision, no quantization	Quantized with memory offloading

The official figures are generous on purpose. The team's own measurements on an RTX 4090 show a 3.6-minute song generating in about 71 seconds, with peak GPU memory around 11 GB for a typical song and about 14 GB at maximum length. That's still out of reach for the 8 GB cards most gaming PCs ship with, which is why Song Creator Pro, among the first apps to run YuE2 on consumer hardware, uses quantization and memory offloading to fit the whole pipeline into 8 GB. The full story, including which cards qualify, is in how to run YuE2 on an 8 GB GPU.

YuE2 vs. YuE (v1)

YuE2 isn't a small update. Almost everything about the model changed:

	YuE (v1)	YuE2
Released	January 2025	September 2026
Architecture	Two stages: a 7B model plus a 1B model, with separate checkpoints per language	One model, roughly 3.6B parameters
Editable score	No	Yes (melody and chords in ABC notation)
Covers	Style transfer from a reference song	Score-based covers that keep the melody
Languages	English, Mandarin, Cantonese, Japanese, Korean	English, Mandarin, Japanese
Speed on an RTX 4090	About 6 minutes per 30 seconds of audio	About 71 seconds for a 3.6-minute song
Official memory guidance	24 GB for short runs; 80 GB recommended for full songs	24 GB
WildSongBench (SongBench Avg)	4.92	6.73 standard, 6.96 best of 8
Weights license	Apache 2.0	CC BY-NC 4.0 plus creator permission

The license row is worth noticing. The original YuE was relicensed to Apache 2.0, which allows almost any use. YuE2's weights are non-commercial by default, with an explicit carve-out for individual creators (more on that below).

YuE2 vs. ACE-Step 1.5

ACE-Step 1.5 is the other major open song model, and the second model built into Song Creator Pro. They're good at different things:

	YuE2	ACE-Step 1.5
WildSongBench (SongBench Avg)	6.73 standard, 6.96 best of 8	6.01
Editable melody and chords	Yes	No
Covers that keep the sung melody	Yes (score-based)	Remixes transform the audio directly
Vocal languages	English, Mandarin, Japanese	50+
Stem extraction and section revision	No	Yes
Speed	Slower	Faster
GPUs in Song Creator Pro	8 GB+ (NVIDIA; AMD and CPU-only experimental)	NVIDIA or AMD with 6 GB+, or CPU-only

In short: pick YuE2 for song quality, structure, editable compositions, and melody-faithful covers. Pick ACE-Step 1.5 for speed, languages beyond the three YuE2 sings, non-NVIDIA hardware, and audio tools like stem extraction.

The YuE2 License, in Plain Terms

The code, agent skill, and documentation are Apache 2.0. The model weights are CC BY-NC 4.0 with an additional creator permission, which the release summarizes like this:

• Personal users, content creators, and musicians can use YuE2 and monetize the songs they make, with no fees or royalties owed to the developers.
• Academic research and education is free for non-commercial use.
• Companies using the weights commercially need to contact the team for a commercial license.

Crediting YuE2 (or tagging #YuE2) is encouraged but optional. The creator permission prohibits illegal, harmful, or deceptive use, and the model comes with no warranties. This is a summary, not legal advice; the full terms are in the repository's model license file.

How to Run YuE2

There are a few ways to get YuE2 making music today:

• The official release. Install the Python package on Linux with a 24 GB NVIDIA GPU and run it from the command line or Python. This gives you every feature and full control, and assumes you're comfortable maintaining a research environment. Covers need SheetSage2 in a separate environment.
• The free online demo. The developers link to a hosted demo run by NOIZ where you can create a song in the browser. Nothing to install, but it runs on someone else's servers, so your lyrics and prompts leave your machine.
• ComfyUI. ComfyUI has native YuE2 nodes and official workflow templates for text-to-music and covers, if you already work in node-based tools.
• Song Creator Pro. A Windows app with YuE2 and ACE-Step 1.5 built in. It runs YuE2 on 8 GB NVIDIA cards (AMD experimental), handles model downloads and transcription for covers automatically, and supports batch generation so you can keep the best of several takes. Everything runs locally.

Where to Go Next

• YuE2 vs. Suno: an honest comparison of quality, features, cost, and ownership, plus what WildSongBench measures.
• YuE2 Prompting Guide: style tags, lyric formatting, and planning modes.
• How to Remix a Song with YuE2: the cover workflow, step by step.
• How to Run YuE2 on an 8 GB GPU: the official release needs a 24 GB GPU, and how Song Creator Pro runs it on 8 GB.