YuE2-3B Beats Suno Generating Editable Full Songs on a Single GPU
An open 3B music model from M-A-P matches Suno v5 on quality, adds editable ABC scores, and runs on a 24GB GPU.
- M-A-P released YuE2-3B, an open 3B music model that generates 5-minute songs with vocals.
- Best-of-8 scores 6.96 on WildSongBench, edging out Suno v5 (6.87) and Suno v6 (6.56).
- Writes an editable ABC score first, then renders 48 kHz stereo audio via flow matching.
- Runs on a 24GB GPU: 3.6-minute song in 71 seconds on an RTX 4090, no quantization.
- Supports agentic editing loops where an LLM revises score, style, and lyrics iteratively.
- Weights are CC BY-NC 4.0 (non-commercial); inference code is Apache-2.0.
YuE2-3B generates editable full songs on an RTX 4090
The Multimodal Art Projection team has released YuE2-3B, an open-weight, three-billion-parameter model that turns lyrics and a style prompt into a complete song with vocals and accompaniment. YuE2 first generates an editable symbolic score, then renders 48 kHz stereo audio from that musical plan.
WildSongBench evaluates full-song quality and text alignment. YuE2 posts the highest SongBench average among the open and proprietary systems in the published results when allowed to generate eight candidates and select the best one.
| Model | Evaluation mode | Average |
|---|---|---|
| YuE2-3B | Best-of-8 | 6.9632 |
| Mureka 9 | Single generation | 6.9377 |
| Suno v5 | Single generation | 6.8721 |
| YuE2-3B | Single generation | 6.7316 |
| Suno v6 | Single generation | 6.5562 |
| Suno v6 Wild | Single generation | 6.4195 |
The leading YuE2 result uses roughly eight times the generation compute of a single call. Eight serial generations would take about 9.5 minutes on an RTX 4090 at the release’s reported speed, before selection overhead. The single-generation result trails Mureka 9 and Suno v5.
An editable score sits in the middle
Models such as ACE-Step 1.5, DiffRhythm 2, and LeVo 2 generate audio tokens directly from text. YuE2 inserts an ABC score between the prompt and the waveform. ABC is a plain-text notation format that can represent notes, timing, and harmony, making the model’s musical plan inspectable and editable.
A single autoregressive and non-autoregressive Mixture-of-Transformers backbone generates the score and semantic tokens, which encode the song’s higher-level audio content. Flow matching converts that representation into acoustic latents, and a variational autoencoder decodes the latents into stereo audio. The combined backbone uses sequential token generation where order matters and parallel prediction where several elements can be produced together.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
