YuE2 Review: The Open Music Model That Topped the Benchmark, Honestly Assessed

YuE2 arrived in September 2026 from m-a-p (Multimodal Art Projection), the research collective behind the original YuE, and it immediately did something no open model had done: it posted a higher average score on a public song benchmark than every commercial service its authors tested, including Suno.

That kind of headline invites skepticism, and it should. This review goes through what YuE2 actually is, what the benchmark does and doesn't prove, how its score planning and cover features work, what hardware it really needs, and where it falls short. Everything here is drawn from the official model card, the published benchmark tables, and the documented behaviour of the model. Where a claim comes from early users rather than the developers, we say so.

A disclosure up front: Song Creator Pro ships YuE2 as one of its two models, so we have an interest in it. That's exactly why the limitations section below is long.

The Verdict in Brief

Area	Rating	Why
Song quality	Excellent	Highest SongBench average on WildSongBench (self-reported)
Structure and control	Excellent	Plans melody and chords as an editable score before rendering
Covers and remixes	Excellent	Keeps the melody recognizable far better than other open models
Pronunciation	Good	Beaten by Suno v4.5/v5.5 and MiniMax Music 3 on phoneme error rate
Languages	Limited	English and Mandarin only
Ease of setup (official)	Poor	Linux, Python, 24 GB NVIDIA GPU, command line
Licence (official weights)	Mixed	CC BY-NC 4.0; individuals may monetize output, companies need a licence

If you have an NVIDIA or AMD GPU and sing in English or Mandarin, YuE2 is the best open-weight song model available at the time of writing. If you need many languages, the fastest possible single take, or a browser tool, look elsewhere.

What YuE2 Is

YuE2 is a roughly 3-billion-parameter lyrics-to-song model. You give it a style prompt and lyrics with section labels, and it produces a complete song, vocals and accompaniment together, as 48 kHz stereo audio. The developers' demo songs run from about a minute to five minutes.

The interesting part is the architecture. Most song generators go straight from text to audio tokens. YuE2 adds a planning stage first: an autoregressive model writes the song's melody and chord progression in ABC, a standard text notation for music, and only then generates the audio that performs that plan. The model card exposes three planning modes:

• Full (default): plans melody and chords.
• Melody: plans only the melody and leaves the harmony free. The developers recommend this for covers.
• Off: skips the score and generates directly.

Because the plan is plain notation, you can read it, edit it, or supply your own, then render audio from it. The developers even demo an agent revising a song over a chain of fourteen versions by editing the score, style and lyrics between renders.

Our YuE2 prompting guide covers how to write style prompts and lyrics for this pipeline in detail.

Quality: What the Benchmark Says

The headline numbers come from WildSongBench, a 192-prompt full-song evaluation published alongside the model. The SongBench average is a composite quality score; PER (phoneme error rate) measures how accurately the lyrics are sung, where lower is better.

System	SongBench average	Phoneme error rate
YuE2 (best of 8)	6.96	9.79%
Mureka 9	6.94	11.69%
Suno v5 (retired)	6.87	8.10%
YuE2 (standard)	6.73	8.44%
Suno v5.5	6.72	5.96%
Suno v4.5 (retired)	6.70	5.80%
Suno v6 (current)	6.56	7.58%
LeVo 2	6.32	26.12%
MiniMax Music 2.6	6.32	24.55%
MiniMax Music 3	6.28	6.27%
ACE-Step 1.5	6.01	7.46%

Source: YuE2 model card on Hugging Face, September 2026.

How to read this honestly:

1. It's a self-reported benchmark. The YuE2 team built WildSongBench and ran every system through it. The prompts and methodology are published, but no independent lab has reproduced the ranking yet.
2. The top number needs eight tries. Best-of-8 means generating eight songs and keeping the one the scorer likes most. At the standard setting (the model card's protocol picks the better of two candidates), YuE2 scores 6.73: above Suno v6, essentially level with Suno v5.5, and below Suno v5 and Mureka 9.
3. Pronunciation is not its strongest suit. Several systems sing lyrics more accurately on this measure, including Suno v4.5, Suno v5.5 and MiniMax Music 3. YuE2's advantage is in overall musicality and prompt adherence, not diction.
4. The comparison is a snapshot. Suno retired v5 and v4.5 on September 9, 2026, and Mureka and MiniMax ship new versions regularly.

Even with every caveat applied, YuE2 is in the same quality tier as the best commercial services, and it is clearly ahead of every other open model in the table. That's a first. For a deeper look at the Suno side of the comparison, see YuE2 vs. Suno, and for MiniMax, YuE2 vs. MiniMax Music 3.

Score Planning in Practice

The score layer changes how you work with an AI music tool. With a black-box generator, a bad chorus melody means rerolling the whole song and hoping. With YuE2, the melody exists as notation before any audio does, which has three practical effects:

• Structure is reliable. A [Chorus] label becomes a planned section of the composition, not a loose hint. This is also why formatting mistakes cost more: anything in the lyrics field gets planned and sung.
• Edits are targeted. Change a phrase in the melody or a chord in the bridge and re-render. The song stays the same song.
• You can bring your own melody. Supply ABC notation you wrote, and YuE2 will sing it.

The cost is that the planning stage adds work to every generation, and full planning is slower than turning it off. For most original songs, full planning is worth it.

Covers and Remixes

This is where YuE2 most clearly beats everything else that's open. The model card reports zero-shot cover results on SHS100K, a standard cover-song dataset, using automatic song-identification models to check whether the cover is recognizably the same song:

Method	Correct song identified first (Hit@1)
YuE2 (full score)	71.3%
YuE2 (melody only, no chords)	67.3%
SongEcho	48.4%
ACE-Step 1.5	2.4%
YuE2 (no score)	0.3%

The last row is the telling one: without the score, YuE2 covers are no more recognizable than chance. The planning stage is what makes covers work. The model transcribes the source song's melody, keeps it as notation, and re-performs it in whatever style you describe.

The official cover workflow is fiddly, though. You transcribe the song with a separate tool (SheetSage2), get the lyrics from a speech recognition model, organize them by section, and then generate. Each tool runs in its own environment. Song Creator Pro wraps those steps into one: upload a song, click Analyze, and the lyrics land in an editable field with section tags already in place. Our YuE2 remix walkthrough covers the full process, including rewriting the lyrics while keeping the melody.

Languages

YuE2 sings in English and Mandarin, the languages its developers' official samples cover. The developers publish no wider language list, and those two are all we've tested. That's a real limitation. Suno, Mureka and ACE-Step 1.5 all cover far more languages, and if you write in Japanese, Spanish, Korean or Hindi, YuE2 is the wrong model. In Song Creator Pro, that's what ACE-Step 1.5 is for: it handles vocals in 50+ languages on the same hardware.

Hardware Reality

The official release is a research release. Its stated requirements:

Requirement	Official YuE2
OS	Linux
Software	Python 3.10+, manual environment setup
GPU	NVIDIA with 24 GB VRAM and BF16 support
System RAM	24 GB available
Interface	Command line (Python API)

The model card reports a measured peak of about 11 GiB of VRAM on an RTX 4090, and a 3.6-minute song rendered in 71 seconds on that card with full planning. That's fast, but it's on the most expensive consumer GPU NVIDIA sells, under Linux.

For most people this is where YuE2 stops being an option. Mainstream gaming cards carry 8 to 16 GB, and most of those machines run Windows.

There are now three realistic ways to run YuE2:

• Official release. Free, full control, Linux plus a 24 GB card.
• ComfyUI. Native YuE2 nodes landed in September 2026. Good if you already live in node graphs; the model card notes separate dependencies, and you assemble the cover workflow yourself.
• Song Creator Pro. A Windows app that runs YuE2 on 8 GB NVIDIA or AMD cards through quantization and memory offloading, with text-to-song, planning modes, Remix and batch generation built in. It was among the first apps to do this. Generation on an 8 GB card is slower than the 4090 figure above. See how to run YuE2 on an 8 GB GPU for details.

Limitations

Here's everything we'd want to know before committing to YuE2:

• Two vocal languages. English and Mandarin are the only languages its official samples cover.
• Pronunciation trails the leaders. An 8.44% phoneme error rate at standard settings is good, but Suno v4.5 and MiniMax Music 3 do better.
• Best results need multiple takes. The 6.96 score is a best-of-8 number. Plan on generating a few candidates per prompt. Locally this costs time, not money.
• The official weights aren't fully open for business use. They're released under CC BY-NC 4.0 with an extra creator permission: individual creators and musicians who self-host can sell and monetize their songs, but companies using the weights commercially need a separate licence from the team. Song Creator Pro uses YuE2 with direct permission from its developers, and its users can use YuE2 output commercially; see our Terms of Service for details.
• No fine-tuning yet. The training pipeline hasn't been released, so community LoRAs aren't feasible at the time of writing.
• Strict input formatting. Every non-sung word in the lyrics field gets sung. Repeated choruses must be written out in full.
• Fewer generation knobs in Song Creator Pro. The app doesn't expose YuE2's cfg_scale setting; you steer with the style prompt, planning mode and seed.
• No section revision. To redo one section without touching the rest, Song Creator Pro's ACE-Step 1.5 is the tool (how to edit part of an AI song).

Who YuE2 Is For

A strong fit if you:

• Write lyrics in English or Mandarin and want the best open model for them
• Care about song structure and want to edit melody or chords rather than reroll
• Make covers, genre swaps, or rewritten-lyric versions of songs
• Want to generate locally, with no credits, uploads or subscriptions
• Have an NVIDIA or AMD GPU with at least 8 GB of VRAM (with Song Creator Pro), or a 24 GB NVIDIA GPU (official release)

Probably not for you if you:

• Need vocals in other languages
• Don't have an NVIDIA or AMD GPU, or work from a laptop with integrated graphics
• Want a browser tool with sharing and community features
• Want the single most accurately sung take, every time, with no curation

If you're weighing YuE2 against other open models, our roundup of the best open-source AI music models puts it in context, and best GPUs for AI music generation covers the hardware side.

Final Verdict

YuE2 is the first open-weight song model that belongs in the same conversation as Suno and Mureka on quality, and it does something none of them offer: it writes the song down before performing it. That makes it more controllable for originals and far better at covers than anything else you can run yourself.

The trade-offs are real. Two languages, a self-reported benchmark, best scores that depend on curation, official weights that companies can't use commercially without a separate licence, and official hardware requirements that exclude most PCs. If those don't rule you out, it's the most interesting music model released this year.

---

Ready to try YuE2 on your own PC? Get Song Creator Pro on the Microsoft Store for $49.99 (free trial included) or on itch.io for $44.99. One-time purchase, YuE2 and ACE-Step 1.5, unlimited generations on 8 GB NVIDIA or AMD GPUs, all running locally on Windows.