Score first, then sound
The model can write melody and chords before it renders audio. In the official pipeline you can inspect that ABC plan. Here you brief it in chat; planning still happens inside the run.
YuE 2 AI music
YuE2. Lyrics in. Full song out.
Paste lyrics and a style. YuE2 sings them back with a band.
Start from a brief
48 kHz
Stereo full-song output
6.96
WildSongBench best-of-8 average
YuE2-3B
Open weights you can inspect
Chat
Revise a take in conversation
What is YuE2?
YuE2 is an open music generation model from Multimodal Art Projection (M-A-P). Give it lyrics and a style prompt. It returns a complete song with a lead vocal and accompaniment in 48 kHz stereo.
Search results also show YuE 2 and YuE 2 AI. Those names point to the same family. The public checkpoint is YuE2-3B. This site is an online studio: you talk in chat, then listen to the take.
The model is not a sealed jukebox. Official inference writes a symbolic plan first (melody and chords as ABC), then audio. That middle step is why covers, reharmonization, and agent edits exist in the published workflow.
On WildSongBench (12 September 2026), the standard row lands at 6.7316 SongBench average and best-of-8 at 6.9632, above Suno v5 (6.8721) and Suno v6 (6.5562) under the reported protocol. Those are model scores, not a promise for every prompt you type here.
This studio is not the M-A-P lab. Weights stay on Hugging Face. We host a chat path so you can try a song without standing up a 24 GB GPU. Community LoRAs such as hum-to-song extend the stack further; we mention them as research, not as buttons on this page.
What YuE2 actually does
Plan the song, then perform it. Chat on this site is how you send a brief and ask for another take.
Vocals and a band
This stack is built for full songs, not 15-second loops. Lyrics become a sung performance with accompaniment. Review pronunciation on every take; published PER is higher than Suno v6.
Covers with a new setting
Official covers start from a transcribed melody and a new style. Demos include Auld Lang Syne as jazz-funk and Jingle Bells as metal. Ask for that kind of rewrite in chat.
Edit like a session
The published agent demo revises a song across many versions. This studio uses conversation the same way: hear a take, change the lyric or the brief, generate again.
More than one language
Published demos cover English, Mandarin, Japanese, and Spanish. Vowels and phrasing still need a human ear. Write section labels so the form stays visible.
Open weights, named scores
YuE2-3B is downloadable. WildSongBench tables are public. If you need to run locally, the model card is the source. This site is the hosted chat lane.
Directions YuE2 already sang
Stills for published moods. We do not host those audio files here. Use a card as a brief, then generate your words.
English · original
Cyber Metal
Stacked guitars, English vocal, a five-minute form. A flagship original from the public demo set.
Mandarin · nu-disco
今晚不眠
Mandarin funk and nu-disco. A night that does not end.
English · rock
Passion
Heartland rock from a lyric-and-style prompt.
Cover · jazz-funk
Auld Lang Syne
A cover: familiar carol, new harmonic clothes.
Spanish · vocal
Flamenco night
Spanish phrasing in a public studio example set.
Japanese · city pop
After midnight
City pop color that the public demos explore in Japanese.
How to use YuE2 here
1. Put the words down
Paste lyrics. Mark [verse], [chorus], and [bridge] so the form is visible. Keep musical notes out of the lyric field.
2. Name the room
Name the genre, the voice, the instruments, and the tempo. A short style paragraph beats a pile of tags.
3. Play the take
Sign in, spend credits, and send the chat. The model renders vocals plus accompaniment. Wait for the file; a progress bar is only an estimate.
4. Talk back
If a line sits wrong, rewrite it and ask for another pass. Same seed, new brief, or a cover-style rewrite: say it in the next message.
YuE2 vs Suno v6, as published
Figures come from the team WildSongBench table (12 September 2026, 192 prompts). The YuE2 column is the standard row, not best-of-8. These scores describe the model, not every song from this studio.
| Measure | YuE2 | Suno v6 |
|---|---|---|
| Musicality | 5.9075 | 5.6558 |
| SongBench average | 6.7316 | 6.5562 |
| Prompt adherence (Q3O) | 4.6819 | 4.6258 |
| Pronunciation errors (PER, lower is better) | 8.44% | 7.58% |
Who YuE2 is for
Writers with words, no band
You have a verse. Hear whether the chorus lands, then rewrite.
Producers who want another pass
Use a vocal demo, a genre flip, or a cover sketch. Official score editing lives in the model workflow; chat here is the hosted lane.
Builders inspecting an open model
Weights, planning APIs, and evaluation kits are public. Run locally if you have the GPU. Use this site when you want a take without the install.
Choose Your Plan
Create more music. Unlock commercial rights.
FREE
$0
Free trial
- 10 free songs
- Lossless audio quality
- Unlimited downloads
- Private songs
- 24×7 email support
Most popular
PRO
$99.00
- 90 songs
- Valid for 365 days
- Commercial license
- Lossless audio
- Private songs
BASIC
$49.00
- 40 songs
- Valid for 30 days
- No commercial use
- Lossless audio
- Private songs
YuE2 questions
An open music model from M-A-P. It turns lyrics and a style prompt into a song with vocals and accompaniment. The public checkpoint is YuE2-3B.
No. YuE 2, YuE 2 AI, and YuE2 refer to the same model family. YuE 2 AI is everyday search language for writing a song with this stack.
In the published WildSongBench table, musicality is 5.9075 vs 5.6558 and SongBench average is also higher. Suno v6 has a lower phoneme error rate (7.58% vs 8.44%). Listen to your own output.
Yes, on your hardware. Official quick start wants Linux, Python 3.10+, and a 24 GB NVIDIA GPU with BF16. Community ComfyUI graphs exist. This website does not require that install.
The official cover path uses a transcribed melody (often via SheetSage2) plus a new style. In this chat studio you describe the cover and lyrics; the model generates a new rendition. Melody lock is not guaranteed here.
The stage that writes melody and chords before audio. Modes include full plan, melody-only, and off. This chat composer does not expose those switches or ABC download.
A community LoRA (hum-to-song) does that on Hugging Face Spaces. It is not a button on this homepage. You can still describe a hummed melody in text and ask the model to follow it.
Yes. An account holds credits and your chat history. Sign in, paste lyrics, send. Free credits depend on the live plan table below.
Published demos include English, Mandarin, Japanese, and Spanish. Other lyric languages are worth a try; check vowels yourself.
Paid plans are meant for download and commercial use subject to our terms. Weights on Hugging Face use CC BY-NC 4.0. Hosted output follows this site's license, not the weight license, unless we say otherwise.