How to Run YuE2-3B Locally: Open Source Music Generation That Beats Suno v5 on Benchmarks

YuE2-3B is a 3.63 billion parameter open source music model from m-a-p, released September 9, 2026 under a non-commercial license. Give it lyrics and a style prompt and it writes a full song, vocals and instruments included, plus an editable score you can revise and regenerate. It needs a 24GB GPU, and on the WildSongBench benchmark its best-of-8 setting beats Suno v5.

Written by Priya Raghunathan, local-AI hardware reviewer. I check parameter counts and hardware requirements against the primary source before recommending an install path.

A three-day-old model already in HF's trending top 20

YuE2-3B entered Hugging Face's trending top 20 across all pipelines within days of its September 9, 2026 release. At time of writing it has 174 likes and 971 downloads on the Hugging Face repo, plus 6,851 stars on the multimodal-art-projection/YuE GitHub repo it ships from and 12 likes on the top community Space, mrfakename/yue2-3b. A four-figure GitHub star count within days points to a lab with an established following, not a first release: m-a-p's original YuE model from 2025 built that audience already.

Specs at a glance

Hardware: 24GB GPU, published numbers

Unlike a lot of new HF entrants, m-a-p actually publishes hardware and speed numbers instead of leaving you to estimate. The quick start lists a 24GB NVIDIA GPU with BF16 support plus 24GB of available host RAM as the baseline for generating one song at a time, and the README backs that with measured runs.

Peak measured VRAM on an RTX 4090 tops out around 11.2 GiB, well under the card's 24GB, but m-a-p still calls for a 24GB card as the baseline: maximum-context runs in their own testing peaked at 14.08 GiB, and that's before you add headroom for a desktop environment or a second app. If you're shopping for a card specifically to run models like this, see our guide on setting up local AI on 24GB of VRAM.

What makes it different: an editable score, not just audio

Most text-to-music models output a waveform and nothing else. YuE2-3B's AR-NAR backbone writes a symbolic plan (an ABC-notation score plus semantic tokens) before flow matching turns that plan into audio and a VAE renders it as 48kHz stereo. That plan is a real intermediate artifact you can inspect, hand-edit, or route through an agent that reharmonizes a section or adapts the lyrics, then feed back in to regenerate. The README calls this workflow agentic editing, and its demo walks a single song through 9 edit steps and 14 versions, from Mandarin pop to an English jazz arrangement with a saxophone solo.

Benchmarks: where it beats Suno v5, where it doesn't

On WildSongBench, a 192-prompt full-song generation benchmark, YuE2-3B at best-of-8 posts a SongBench average of 6.9632, ahead of Suno v5's 6.8721, Suno v5.5's 6.7150, and MiniMax Music 3's open-model score of 6.2830. It's not the top score on every axis: Suno v5 scores higher on MuLan and AllMusicCaps (style and caption alignment), and Suno v4.5 has a lower phoneme error rate. Mureka 9, a proprietary model, edges out YuE2-3B's non-best-of-8 setting on SongBench average, 6.9377 to 6.7316, though best-of-8 YuE2-3B still comes out ahead of it too.

YuE2 is an open music generation model that rivals Suno v5.

Against its own predecessor the jump is large: the original YuE scored a 4.9165 SongBench average on the same benchmark, so YuE2-3B's non-best-of-8 score of 6.7316 is roughly a 37% improvement generation over generation.

Install it

YuE2-3B ships as a wheel file downloaded from its own HF repo rather than a PyPI package. You'll need Linux, Python 3.10 or newer, and that 24GB GPU.

Generate your first song

Load the pipeline once, then generate from a style prompt and lyrics. The example below pulls one of the model's own demo prompts so you can hear the intended output before writing your own.

from pathlib import Path
from yue2 import YuE2Pipeline
import json
from huggingface_hub import hf_hub_download

pipe = YuE2Pipeline.from_pretrained("m-a-p/YuE2-3B", device="cuda")

repo = "m-a-p/YuE2-3B"
prompt_path = hf_hub_download(repo, "examples/tonight-awake.json")
demo = json.loads(Path(prompt_path).read_text(encoding="utf-8"))
style, lyrics = demo["style"], demo["lyrics"]

song = pipe(style=style, lyrics=lyrics, cot="full", seed=demo["seed"])
song.save("song.flac")
song.save_artifacts("outputs/song")

save_artifacts writes out the ABC score, tokens, latents, and generation settings alongside the audio, which is what you'll edit if you want to revise the song later.

The cot setting: full, melody or off

Cover an existing song

Covers work from a melody score and a lyrics file rather than a reference audio clip. Transcribe the melody with m-a-p's own SheetSage2 model into ABC notation without chord symbols, get the lyrics from an agent search or a speech-to-text pass, then supply both to the pipeline with a new style prompt and cot="melody".

cover = pipe(
    style="Jazz-funk, warm lead vocal, Rhodes piano, electric bass, tight drums",
    lyrics=Path("cover_lyrics.txt").read_text(encoding="utf-8"),
    abc=Path("melody.abc").read_text(encoding="utf-8"),
    cot="melody", seed=831001,
)
cover.save("cover.flac")
cover.save_artifacts("outputs/cover")

Edit a song after generating it

Because YuE2-3B exposes the intermediate plan, you can export it, edit the ABC score by hand or through an agent, and regenerate with the same seed and lyrics to hear the revision. Copy the score.abc file from your first song's output folder, edit it, then rerun with a new style prompt describing what changed.

plan = pipe.plan(style=style, lyrics=lyrics, cot="full", seed=demo["seed"])
plan.save("original_plan")
# edit outputs/song/score.abc, save the copy as edited.abc

edited_style = (
    "Jazz, expressive lead vocal, piano, tenor saxophone, upright bass, "
    "brushed drums, no guitar, spacious modern harmony"
)
song = pipe(style=edited_style, lyrics=lyrics, cot="full", seed=demo["seed"],
            abc=Path("edited.abc").read_text(encoding="utf-8"))
song.save("edited.flac")
song.save_artifacts("outputs/edited")

FAQ

What is YuE2-3B used for?

Generating complete songs, vocals and instrumentation included, from a text style prompt and lyrics. It also covers existing songs in a new style and supports agent-driven editing, where you or an AI agent revise the musical score and regenerate the audio.

How many parameters does YuE2-3B have?

3.63 billion parameters, per the safetensors metadata in the Hugging Face API.

What GPU do I need to run YuE2-3B?

A 24GB NVIDIA GPU with BF16 support and 24GB of available host RAM, per the model card's quick start. Measured peak VRAM on an RTX 4090 is around 11.2 GiB for a typical song, but m-a-p's own maximum-context testing reached 14.08 GiB, which is why the 24GB baseline has real headroom built in.

Is YuE2-3B better than Suno v5?

On the WildSongBench benchmark, YuE2-3B's best-of-8 setting scores a 6.9632 SongBench average against Suno v5's 6.8721, and it also leads on musicality and caption-alignment metrics. Suno v5 still scores higher on style similarity (MuLan) and caption match (AllMusicCaps). Benchmark averages aren't a guarantee your specific prompt will sound better in either tool.

Can I use YuE2-3B commercially?

No, not without a separate license. The weights are released under CC BY-NC 4.0, a non-commercial license, so shipping YuE2-3B output in a paid product or service requires a separate agreement with m-a-p.

Where to go from here

• Shopping for the GPU this needs? How to set up local AI on 24GB of VRAM covers the hardware tier YuE2-3B is built for.
• Want speech instead of full songs? How to run tencent/AuK locally and how to run Kokoro-82M locally cover open text-to-speech models instead.