YUE2 VS SUNO
YuE2 vs Suno: what the benchmark says and what it does not
The YuE2 authors report that their open model is competitive with Suno v5 and v6 on an automatic benchmark. The headline number is real, and so are its conditions. This page shows the scores as published, explains how they were produced, and compares the two on the points a benchmark cannot measure.
Reviewed October 4, 2026. yue2.pro is an independent studio and is not affiliated with the YuE2 authors, WaveSpeed, or any other project named on this page.
- 6.9632
- YuE2 best-of-8 · SongBench Avg
- 6.8721
- Suno v5 · SongBench Avg
- 6.7316
- YuE2 standard · SongBench Avg
- 192
- WildSongBench prompts
SHORT VERSION
Close on the published scores, different in kind
- On WildSongBench, YuE2 with best-of-8 selection has the highest SongBench average of the 17 evaluated settings. Standard YuE2 scores below Suno v5 and above Suno v6.
- These are automatic metrics reported by the YuE2 authors on September 12, 2026. This site has not reproduced them.
- YuE2 is a downloadable model with an editable score. Suno is a proprietary hosted service. That difference matters more than a few hundredths of a point for most decisions.
WILDSONGBENCH
The scores as published
| Model | SongBench Avg ↑ | Musicality ↑ | MuLan ↑ | PER ↓ |
|---|---|---|---|---|
| YuE2 (best-of-8) | 6.9632 | 6.2666 | 0.5051 | 9.79% |
| Mureka 9 | 6.9377 | 6.0488 | 0.4394 | 11.69% |
| Suno v5 | 6.8721 | 5.9918 | 0.5428 | 8.10% |
| YuE2 | 6.7316 | 5.9075 | 0.5068 | 8.44% |
| Suno v5.5 | 6.7150 | 5.8087 | 0.5089 | 5.96% |
| Suno v4.5 | 6.6995 | 5.8317 | 0.5022 | 5.80% |
| Suno v6 | 6.5562 | 5.6558 | 0.4916 | 7.58% |
| Suno v6 Wild | 6.4195 | 5.5644 | 0.4999 | 7.45% |
| MiniMax Music 2.6 | 6.3222 | 5.4437 | 0.4251 | 24.55% |
| MiniMax Music 3 | 6.2830 | 5.3482 | 0.3928 | 6.27% |
| ACE-Step 1.5 | 6.0118 | 5.1588 | 0.4372 | 7.46% |
SongBench Avg averages seven quality dimensions. MuLan measures how closely the audio matches the style text. PER is phoneme error rate, a proxy for how intelligible the sung lyrics are; lower is better.
CAVEATS
How to read those numbers
Best-of-8 is a selection protocol
The top YuE2 row picks the best of eight candidates per prompt. The standard YuE2 row selects from two. One pipeline call produces one candidate; selection is a separate step.
Standard YuE2 sits between Suno versions
At 6.7316 it is below Suno v5 (6.8721) and above Suno v6 (6.5562) and v6 Wild (6.4195).
Suno v5 leads on text alignment
Its MuLan score of 0.5428 is the highest in the table, ahead of both YuE2 rows.
Older Suno versions sing more clearly on this metric
Suno v4.5 and v5.5 post phoneme error rates under 6%, lower than either YuE2 setting.
It is an automatic benchmark
The authors label these as automatic results under stated protocols and are running a separate blind listening arena for human preference.
Protocols differ between systems
The model card notes that earlier proprietary systems keep their delivered-candidate protocols, and that the benchmark used a specific YuE2 decoder (YuE2-Vae-legacy).
A track generated in this studio is a single result from hosted YuE2-3B inference. It is not a best-of-8 selection, so do not read the top row as a promise about any one generation.
BEYOND THE SCORE
Practical differences
| YuE2 | Suno | |
|---|---|---|
| What it is | An open music generation model with downloadable weights | A proprietary hosted music service |
| Where it runs | Your own GPU, ComfyUI, or a hosted endpoint | Suno’s own platform |
| Composition control | Melody and chords are written as an ABC score you can read and edit before rendering | Defined by Suno’s product; see its documentation |
| Covers | From a transcribed melody score, in a new style | Defined by Suno’s product |
| Local hardware | 24 GB NVIDIA GPU for the reference runtime | None |
| Terms | Code Apache 2.0; weights CC BY-NC 4.0 with an additional creator permission | Suno’s own terms and plans |
| Output | 48 kHz stereo | See Suno’s documentation |
This page does not list Suno’s prices or feature set because they change and are best read on Suno’s own site. The YuE2 column is taken from the authors’ repository and model card.
ALSO COMPARED
YuE2 vs MiniMax Music
The same table includes MiniMax Music 3 at a SongBench average of 6.2830 and MiniMax Music 2.6 at 6.3222, both below standard YuE2 at 6.7316. MiniMax Music 3 has the lowest phoneme error rate among the open-model group at 6.27%, compared with 8.44% for YuE2. The model card notes that MiniMax Music 3 was run with its official caption rewriter.
CHOOSING
Which one fits the job
- Choose YuE2 when you want to inspect or edit the melody and chords, run the model on your own hardware, or build on open weights.
- Choose a hosted service when you want a finished product experience and do not need access to the model.
- If you are comparing sound, compare with your own lyrics. A benchmark average does not tell you how a model handles your genre, language or phrasing.
QUESTIONS
Short answers.
Answers are drawn from the sources listed below and from how this studio works today.
Is YuE2 better than Suno?
On the authors’ WildSongBench results, YuE2 with best-of-8 selection scores 6.9632 against 6.8721 for Suno v5, while standard YuE2 scores 6.7316. These are automatic metrics reported by the YuE2 team, not an independent listening verdict.
What does best-of-8 mean?
Eight candidates are generated for each prompt and the best is selected by Musicality, then prompt adherence, then phoneme error rate. Standard YuE2 in the same table selects from two candidates.
Is YuE2 free?
The weights are free to download under CC BY-NC 4.0 with an additional creator permission, and the code is Apache 2.0. Running it needs your own GPU. yue2.pro is a paid hosted studio that quotes credits before each generation and gives new accounts a welcome trial.
Can I use YuE2 output commercially?
The upstream licence summary says personal users, content creators and musicians may use YuE2 and monetize generated outputs with no fees or royalties to the authors, while companies should contact them for a commercial licence to the weights. Read the full MODEL_LICENSE; this is not legal advice.
Does yue2.pro use best-of-8 selection?
No. Each generation returns one result from hosted YuE2-3B inference.
Is yue2.pro affiliated with Suno or the YuE2 authors?
No. yue2.pro is an independent studio. Suno, YuE2 and the other names on this page belong to their respective owners.
KEEP READING
Related guides.
Prompting guide
What belongs in the style field, how to tag lyrics, style briefs by genre, and the mistakes that waste a generation.
Local install & VRAM
Official requirements, measured VRAM and speed, install commands, and the community Mac and GGUF options.
YuE2 ComfyUI guide
Required version, the checkpoint file and folder, both official workflows, planning modes, and the duration budget.
SOURCES
Where the facts come from.
Model details, measurements and licence terms on this page are taken from the projects’ own pages, read on October 4, 2026. They can change; check the source before relying on a figure.
YUE2-3B IN THE BROWSER
Judge it with your own lyrics.
Run the same words and style brief you would use anywhere else, and listen to what YuE2-3B returns.
Open the YuE2 studio