Can YuE2-3B be used commercially?

YuE2-3B weights are released under CC BY-NC 4.0, which does not permit commercial use. The project's first-party code, agent skill and documentation carry a separate Apache-2.0 license, so the software around the model can be reused freely while the model itself stays limited to research and non-commercial projects.

What hardware does YuE2-3B need?

YuE2-3B is documented for Linux with Python 3.12 and an NVIDIA GPU that supports BF16 and has 24 GB of VRAM. The team tested generation on an RTX 4090 and used H800 80GB cards for server deployments. Installation is a standard pip install from the cloned repository.

How does YuE2 compare to Suno v6?

On WildSongBench, YuE2 with best-of-8 selection scores 6.9632 against 6.5562 for Suno v6 and 6.8721 for Suno v5; a single YuE2 run scores 6.7316. WildSongBench is a 192-prompt set built and published by the YuE team, so the comparison comes from the model's authors rather than an independent evaluator.

What was YuE2 trained on?

The YuE2 team states its models are trained primarily on CC0 music and synthetic data, with most of the synthetic training data licensed from Tokenwave.AI. Training hours are listed per component: 346K hours for YuE2 itself, 700K hours for the MERT2 encoder and 28.4K hours for SheetSage2.

Can a song be changed after YuE2 generates it?

Yes. YuE2 exposes the symbolic score it planned, so melody and chords can be read, played and edited before the audio is rendered. The system also supports agentic editing, where an existing track is refined through conversation, plus zero-shot covers built from a song that SheetSage2 has transcribed.

Multimodal Art Projection · 2026-09-09 · major

YuE2-3B — open music model tops Suno v5 and v6 on WildSongBench

YuE2-3B turns lyrics and a style prompt into a full 48 kHz song with vocals. It writes an editable melody-and-chord score first, then renders the audio, and scores 6.96 on WildSongBench against Suno v6's 6.56.

Hugging Face model card banner for YuE2-3B by Multimodal Art Projection

An open 3B music model that writes an editable score first, then renders a full song with vocals.

Quick facts

Maker	Multimodal Art Projection (m-a-p)
Size	3.59B parameters, 28 layers
Code license	Apache-2.0
Weights license	CC BY-NC 4.0 (non-commercial)
Audio output	48 kHz stereo
Hardware	NVIDIA GPU, 24 GB VRAM, BF16
Release	yue2-v0.1.6, 9 September 2026

Benchmarks

WildSongBench — SongBench average (192 prompts)

YuE2 (best-of-8)		6.9632 / 10
Suno v5		6.8721 / 10
YuE2		6.7316 / 10
Suno v6		6.5562 / 10
Suno v6 Wild		6.4195 / 10

source ↗

What is it?

YuE2-3B adds an editable symbolic score to open music generation: the model plans melody and chords you can read, play and change, and only then renders audio. Multimodal Art Projection released it on 9 September 2026 with weights on Hugging Face. The same system also produces zero-shot covers of a transcribed song and refines an existing track through conversation.

How does it work?

A 3B AR-NAR Mixture-of-Transformers backbone generates the symbolic score and semantic tokens. Flow matching then produces acoustic latents, which a VAE decodes to 48 kHz stereo audio. Two helper models ship alongside YuE2-3B: SheetSage2 transcribes audio back into a score, and MERT-v2-FullSong supplies music representations.

Why does it matter?

Open music generation has trailed the closed services, and the WildSongBench table puts YuE2 ahead of both Suno v5 and Suno v6 on a 192-prompt set. The score-first design also answers a common complaint about text-to-music: when the melody is wrong you edit the plan instead of rerolling the prompt. The weights are non-commercial, so this is for research and personal work rather than shipping products.

Who is it for?

audio ML researchers and music producers

Frequently asked questions

Can YuE2-3B be used commercially?: YuE2-3B weights are released under CC BY-NC 4.0, which does not permit commercial use. The project's first-party code, agent skill and documentation carry a separate Apache-2.0 license, so the software around the model can be reused freely while the model itself stays limited to research and non-commercial projects.
What hardware does YuE2-3B need?: YuE2-3B is documented for Linux with Python 3.12 and an NVIDIA GPU that supports BF16 and has 24 GB of VRAM. The team tested generation on an RTX 4090 and used H800 80GB cards for server deployments. Installation is a standard pip install from the cloned repository.
How does YuE2 compare to Suno v6?: On WildSongBench, YuE2 with best-of-8 selection scores 6.9632 against 6.5562 for Suno v6 and 6.8721 for Suno v5; a single YuE2 run scores 6.7316. WildSongBench is a 192-prompt set built and published by the YuE team, so the comparison comes from the model's authors rather than an independent evaluator.
What was YuE2 trained on?: The YuE2 team states its models are trained primarily on CC0 music and synthetic data, with most of the synthetic training data licensed from Tokenwave.AI. Training hours are listed per component: 346K hours for YuE2 itself, 700K hours for the MERT2 encoder and 28.4K hours for SheetSage2.
Can a song be changed after YuE2 generates it?: Yes. YuE2 exposes the symbolic score it planned, so melody and chords can be read, played and edited before the audio is rendered. The system also supports agentic editing, where an existing track is refined through conversation, plus zero-shot covers built from a song that SheetSage2 has transcribed.

Try it

git clone https://github.com/multimodal-art-projection/YuE && cd YuE && python -m pip install .