YuE2

Purpose

YuE2 is an open music generation model developed by Multimodal Art Projection (M-A-P). Its core purpose is: input lyrics and a style prompt, and output a complete song containing lead vocals and accompaniment, in 48 kHz stereo format. YuE 2 and YuE 2 AI appearing in search results all point to the same model series, and the public checkpoint is YuE2-3B. This website is an online studio: users submit creative instructions via chat, then audition the generated results.

Main capabilities

Plan first, perform later

The model can write out the melody and chords (in ABC notation format) before rendering the audio. In the official inference pipeline, this intermediate plan can be viewed. In this site's chat interface, the planning process still happens internally, but users cannot directly access ABC downloads or switch planning modes (full plan, melody only, off, etc.).

Vocals and band

This stack is aimed at complete songs rather than 15-second loops. Lyrics are turned into a sung performance with accompaniment. Site tips should note the pronunciation in each audition; the officially published phoneme error rate (PER) is higher than Suno v6.

Covers and new styles

The official cover path starts from a transcribed melody (often with the help of SheetSage2) plus a new style. Demo tracks include a jazz-funk rendition of "Auld Lang Syne" and a metal rendition of "Jingle Bells". Users can make similar rewrite requests in chat.

Edit like a recording studio

The official agent demo can revise a song across multiple versions. This site's studio implements a similar process through conversation: audition a version, modify the lyrics or prompt, and generate again.

Multilingual

Published demos cover English, Mandarin, Japanese, and Spanish. Lyrics in other languages can also be tried, but vowels and phrasing still need manual checking. It is recommended to write section labels in the lyrics to keep the song form clear.

Open weights and public scores

YuE2-3B is downloadable, and the WildSongBench table is public. For local running, the model card is the authoritative source. This site provides a hosted chat channel.

How to use

1. Input lyrics: Paste the lyrics, mark sections with [verse], [chorus], [bridge], etc., and keep the song form visible. Do not include notes in the lyrics field.
2. Specify style: State the genre, voice, instruments, and tempo. A short style description is better than aStack labels.
3. Generate a preview: log in, spend credits, send a chat message. The model renders vocals plus accompaniment. Wait for the file to be generated; the progress bar is only an estimate.
4. Revise via conversation: if a line is unsuitable, rewrite it and ask for generation again. You can reuse the same seed, change to a new prompt, or ask for a cover-style rewrite; just explain in the next message.

Use cases

• Creators with lyrics but no band: listen to the chorus first, then rewrite.
• Producers who want extra versions: useful for vocal demos, style transfer, or cover drafts. The official scoring editing feature is in the model workflow; the chat on this site is a hosted channel.
• Developers who want to study open models: weights, planning API, and evaluation suite are public. With a GPU you can run locally; if you do not want to install anything, you can use this site to get previews.

Limitations

• The chat composer on this site does not expose symbolic planning mode toggles or ABC download.
• Melody lock cannot be guaranteed for covers.
• Phoneme error rate is higher than Suno v6, so pronunciation needs manual checking.
• Community LoRA (such as hum-to-song) runs on Hugging Face Spaces, not as a button on this homepage; users can still describe a hummed melody in text and ask the model to follow it.
• Using this requires registering an account; the account stores credits and chat history. Free credits depend on the live plan table.
• Running the official quick start locally requires Linux, Python 3.10+, and a 24 GB NVIDIA GPU (BF16). Community ComfyUI graphs are available. This website does not require that installation.
• The weights use the CC BY-NC 4.0 license on Hugging Face. The license for hosted output follows this website's terms, not the weights license, unless otherwise stated. Paid plans are intended to provide downloads and commercial use, subject to the terms.

Performance comparison (based on published data)

In WildSongBench (September 12, 2026, 192 prompts), the YuE2 standard row received a 6.7316 SongBench average score, and the best-of-eight was 6.9632, higher than Suno v5 (6.8721) and Suno v6 (6.5562). The musicality score was 5.9075 versus 5.6558. Suno v6 had a lower phoneme error rate (7.58% vs 8.44%). These are model scores, not a guarantee for every prompt. Please rely on actual previews.

FAQ summary

• YuE 2 and YuE 2 AI are different products? No, they are the same model series; YuE 2 AI is the everyday search term.
• Can it be run locally or used in ComfyUI? Yes, on your own hardware. Follow the official quick start requirements.
• Does it support covers? The official path uses transcribed melody plus a new style; in the chat studio, the user describes the cover and lyrics, and the model generates a new rendition, with melody lock not guaranteed.
• What is symbolic planning? The stage before audio where melody and chords are written out; modes include full plan, melody only, and off. The chat composer does not expose these toggles or ABC download.
• Can humming be turned into a song? The community LoRA (hum-to-song) can do this on Hugging Face Spaces, not as a button on this homepage. You can describe a hummed melody in text and ask the model to follow it.
• Is an account required? Yes, the account stores credits and chat history.
• Which languages are supported? Published demos include English, Mandarin, Japanese, and Spanish.
• Who owns the copyright? Paid plans are intended for download and commercial use, subject to this website's terms. Hugging Face weights use CC BY-NC 4.0. Hosted output follows this website's license unless otherwise stated.