YuE2: How to Run This Open-Source Music Generation Model Locally
YuE2 is an open-weight 3B music model that reportedly outscores Suno V6 on benchmarks. Here's how to run it locally and what quality to expect.
What is YuE2?
YuE2 is an open-weight music generation model with roughly 3 billion parameters that you can download and run on your own GPU. It takes lyrics and a style prompt and produces a full song, complete with vocals, instrumentation, and structure (intro, verse, pre-chorus, chorus, bridge, outro). Unlike most music generators that just spit out a waveform, YuE2 also produces an editable symbolic score, essentially sheet music in ABC notation, that sits in the middle of its pipeline. That means you can, in theory, go back and adjust melody, chords, or lyrics before regenerating the final audio.
According to benchmark charts shared on the model’s card, YuE2 (and a variant called YuE2 BO8) lands in what’s called the “pareto optimal” zone on a chart plotting song quality against prompt adherence, outperforming Suno V6 and edging past Suno v5.5 on those two axes simultaneously. That’s a notable claim since Suno is a closed, paid commercial product, and YuE2 is free and open. Benchmarks like this always deserve skepticism since sample sets are curated, but the fact that an open 3B model is even in the conversation with Suno is worth paying attention to.
TL;DR
- YuE2 is a 3B parameter open-weight model that generates full songs from lyrics and a style prompt, and can also do cover generation and score editing.
- Benchmark charts show it beating Suno V6 and nudging past Suno v5.5 on a chart measuring song quality against prompt adherence, though these figures come from curated samples and deserve independent verification.
- The pipeline is genuinely different from typical music generators: it first writes a symbolic score (ABC notation, essentially sheet music) before converting that into audio, which means the intermediate output is human-editable.
- Local testing showed VRAM usage under 8GB, and a roughly 2-minute song took about 2 minutes to generate on a 48GB card, suggesting it’s approachable on consumer-grade hardware too.
- Multilingual quality varies a lot by language: Russian, Bengali, Brazilian Portuguese, French, Indonesian, and Tagalog came out sounding solid, while Hindi and Urdu vocals were noticeably weaker in hands-on testing.
- Vocals still sound identifiably synthetic in most languages even when the instrumentation and arrangement are convincing, which remains the biggest giveaway that a track is AI-made.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
How does YuE2 actually generate a song?
YuE2’s architecture breaks song generation into three separate stages rather than jumping straight from text to audio.
First, an “AR expert,” a causal left-to-right autoregressive model, takes your lyrics and style prompt and writes a symbolic score in ABC notation, essentially a written-out piece of sheet music. That score then gets converted into semantic tokens.
Second, an “N expert” model takes those semantic tokens and performs flow matching to generate acoustic latents, a compressed internal representation of audio that isn’t directly listenable but encodes the acoustic detail needed to reconstruct sound.
Third, a variational autoencoder (VAE), referred to as a V decoder, converts those latents into an actual 48kHz waveform, the audio file you actually hear.
The practical upshot of this three-stage design is that the middle output, the symbolic score, is a checkpoint you can inspect and edit like real sheet music before the model renders final audio. That’s a meaningfully different workflow from end-to-end diffusion or autoregressive audio models that only let you tweak the input prompt and re-roll the output.
How do you install and run YuE2 locally?
Running YuE2 locally involves downloading the YuE2 code package from Hugging Face, which provides the pipeline needed to load and run the model. Once installed, the model and an example configuration (a JSON file specifying lyrics and style) get pulled down automatically on first run.
In testing on an Nvidia GPU with 48GB of VRAM, actual consumption stayed under 8GB for a full end-to-end song generation, suggesting the model is well within reach of mid-range consumer GPUs, not just data-center cards. A roughly 212-second (about 3.5 minute) song took around 2 minutes to generate.
The input format is straightforward: a title, a style prompt (for example, “city pop, upbeat, danceable, groovy bass”), and full lyrics broken into standard song sections. No fine-tuning or extra configuration was required to get a complete, structured song out the other end.
Is the music quality actually good?
This is where results get more mixed, and where benchmarks alone can’t tell the full story.
On the instrumental side, arrangements were frequently convincing. A city pop track generated from Mandarin lyrics and a style prompt came out with a cohesive groove and structure. English, French, Brazilian Portuguese, and Indonesian tracks in particular had arrangements and percussion that sounded close to human-produced music, and a Tagalog track was singled out as carrying real emotional weight in its delivery.
Vocals are the weaker link. Across nearly every language tested, the singing voice retained an identifiable synthetic quality, even when the underlying music didn’t. This lines up with where most current music generation models, open or closed, still struggle: instrumentation has gotten very good, but AI-generated singing voices remain the easiest way to tell a track wasn’t made by a human.
Language performance was also uneven. Russian, Bengali, Brazilian Portuguese, French, German, Swedish, Indonesian, Dutch, Tagalog, and Persian all produced passable to good results in informal listening tests. Hindi and Urdu vocals, however, came out notably rough, described as sounding like a low-quality cover band rather than anything close to Bollywood-style production. The model’s own documentation doesn’t specify which languages it officially supports, so this kind of manual language-by-language testing is currently the only way to gauge real-world coverage.
Is YuE2 a real alternative to Suno?
For anyone who wants full control, local hosting, and zero subscription cost, YuE2 is a legitimate option to try, especially given how it undercuts Suno on price (free and open versus a paid closed service) while claiming to match or beat it on quality metrics. The editable symbolic score is also a genuine differentiator: no mainstream closed model currently gives you sheet-music-level access to edit melody or chords before final rendering.
That said, benchmark charts showing YuE2 beating Suno V6 should be treated as a starting point, not a verdict. Curated demo samples on a model card are, by definition, the best-case output. Hands-on generation across many languages showed a wider range of quality, from genuinely impressive (Brazilian Portuguese, Tagalog, French) to rough around the edges (Hindi, Urdu). If your use case involves a specific language, it’s worth generating a few test tracks yourself before drawing conclusions from any leaderboard position.
Frequently Asked Questions
How many parameters does YuE2 have?
YuE2 is reported as roughly a 3 billion parameter model, small enough to run locally on a single consumer or prosumer GPU rather than requiring data-center hardware.
How much VRAM does YuE2 need?
In local testing on a 48GB Nvidia GPU, end-to-end song generation consumed under 8GB of VRAM, suggesting it should run comfortably on GPUs with significantly less memory than high-end data-center cards.
Can YuE2 generate music in languages other than English?
Yes. It was tested across many languages including Mandarin, Russian, Polish, Hindi, Urdu, Arabic, Bengali, Spanish, Brazilian Portuguese, French, German, Swedish, Indonesian, Dutch, Tagalog, Persian, and others. Quality varies significantly by language, with Brazilian Portuguese, French, Tagalog, and Bengali sounding notably strong, and Hindi and Urdu sounding weaker.
Does YuE2 let you edit a generated song?
Yes. Because the pipeline generates an intermediate symbolic score in ABC notation before rendering final audio, you can edit melody, chords, or lyrics at that stage and regenerate the song, something most end-to-end music generation models don’t offer.
Is YuE2 actually better than Suno?
Benchmark charts on YuE2’s model card show it outperforming Suno V6 and slightly edging out Suno v5.5 on measures of song quality and prompt adherence. Independent hands-on testing found instrumentation often convincing but vocals still identifiably synthetic in most languages, so real-world quality depends heavily on language and use case.
