This Open Model Out-Scored Suno v5. You Can't Sell the Songs.

TL;DR

YuE2 landed on September 9 from a group co-led by HKUST and M-A-P, and it does something no other frontier song model does: it writes an editable melody-and-chord score first, shows it to you, then renders the audio from that plan. On WildSongBench, its best-of-8 setting posted 6.9632 SongBench Avg, the highest mean of all 15 systems evaluated, ahead of Mureka 9 at 6.9377 and Suno v5 at 6.8721. It runs locally on a 24 GB GPU at 48 kHz stereo with no quantization. And the weights ship under CC BY-NC 4.0, which means the one thing you cannot do with the best open song model available is earn money with it.

Score first, sound second

Every song generator you have used is a black box in one specific way: you get audio back, and if the chorus modulates somewhere you hate, your only lever is to regenerate and hope. There is no intermediate representation to grab.

YuE2 inserts one. Give it lyrics and a style prompt and it first produces a melody-and-chord plan in ABC notation, then converts that score into semantic tokens, then acoustic latents, then stereo audio through a VAE. The team calls this white-box generation. The plan is a text file you can read, play, and edit before a single sample gets rendered.

Think of most song models as a session band that nails a take in one pass and then loses the charts. YuE2 hands you the charts.

The behavior is a flag, not a separate model. cot="full" generates the full melody-and-chord plan and is the default for new songs. cot="melody" keeps a melody plan and leaves the accompaniment free, which the team recommends for covers. cot="off" goes straight from lyrics to audio like everything else on the market. You can also hand it your own ABC score with abc=... and make it play your composition.

The benchmark, and every asterisk on it

WildSongBench is 192 prompts run across 15 settings, scored automatically on September 5. Eight of those settings are public-weight models, five are proprietary, and two are YuE2.

Now the asterisks, most of which the team wrote down itself before anyone asked.

• Best-of-8 is best-of-8. That setting generates eight candidates and picks by SongBench Musicality, then prompt control, then phoneme error rate. The benchmark doc states outright that this "is not equivalent to one unselected pipeline call."
• Even standard YuE2 selects. It picks the lower-PER candidate out of two generations. Public baselines also used two candidates, so that part is matched; the proprietary systems kept their own delivered-candidate protocols.
• The lead is not significant. Their words: "the small gap between the highest means does not establish statistical significance." Both YuE2 rows are inside the proprietary range, which is the actual claim worth making.
• Suno still wins two columns. Suno v5 leads text alignment on MuLan at 0.5428 against YuE2's 0.5068, and Suno v4.5 has the cleanest lyric intelligibility at 5.80% PER against YuE2's 8.44%. Best-of-8 is worse there, at 9.79%, because it optimized for musicality instead.
• The scores used a different decoder. Benchmarks ran on YuE2-Vae-legacy; default listening uses YuE2-Vae. The repo adds that "the release names determine these roles, not the everyday meaning of legacy," a sentence written by someone who could already hear the GitHub issue coming.

If you want to reproduce any of it, the standard-YuE2 protocol is documented on the WildSongBench dataset card. That is a low bar that most model launches still fail to clear.

The number that proves the score is doing work

Benchmark leads this tight are easy to argue about. The cover evaluation is not. Across 948 SHS100K works, two requested styles and two seeds each, so 3,792 outputs per method with no candidate selection, the same general checkpoint was asked to reinterpret existing songs in new styles.

With the full transcribed score, cover identity holds at 0.647 CLEWS mAP and 71.3% Hit@1. Without a score, the same model scores 0.006 and 0.3%. That is not a tuning difference, it is the difference between reinterpreting a song and writing an unrelated one that happens to share lyrics.

There is an honest trade in the other direction, and the team publishes it: dropping the score improves target-style alignment (MuLan 0.474 against 0.382) and raw musicality, because a fixed transcription of one performance constrains how far you can bend toward a contrasting style. Hence the product advice to use melody-only covers and let the accompaniment move.

Editing lands where you point it

A separate paired study on ten works and 380 full-song recordings measured whether a requested edit actually happens. Local changed-note melody attainment went from 0.0083 to 0.9375; changed-duration harmony attainment from 0 to 0.8313. Ten works is a small cohort developed on itself, and the units do not combine into one score, both of which the doc says plainly. But going from effectively never to nearly always is not a subtle effect. Note also that editing regenerates the whole recording. It does not preserve your original waveform outside the edit.

The license is the catch

Code and documentation are Apache 2.0. The model weights are CC BY-NC 4.0. Non-commercial. You can download the best-scoring song model on this benchmark, run it on one consumer GPU, and legally do nothing commercial with the output.

Which is a strange place to land given the training story, because the training story is the cleanest part of this release. YuE2 saw 346K hours, SheetSage2 28.4K hours, MERT2 700K hours, and the team says its models are "trained primarily on CC0 music and synthetic data," with Tokenwave.AI licensing most of the synthetic set. Read "primarily" and "most" as the hedges they are: there is no per-source manifest here, and you should not treat this as an audited clean-data claim.

Compare the alternative. In August, Judge F. Dennis Saylor IV let UMG and Sony add a DMCA Section 1201 claim against Suno alleging it stream-ripped YouTube with yt-dlp to build its training set, a claim Suno is contesting on standing. Sony and Universal's separate bid to expand the suit from 560 recordings to 61,026 was denied. So the two poles of AI music right now are a commercial product fighting over where its audio came from, and a research release that can tell you where its audio came from but forbids you from selling anything you make with it.

What it takes to run

Linux, Python 3.12, an NVIDIA GPU with BF16 support and 24 GB of VRAM. It outputs 48 kHz stereo without quantization, which is the spec that matters: a 3090 or 4090 gets you full-rate stereo, not a downsampled preview. Three pieces ship alongside the generator:

• SheetSage2, audio-to-score transcription, SOTA on 10 of 13 metrics with 82.51% vocal melody pitch-class F1 on RWC-Pop. This is what makes zero-shot covers possible.
• MERT2, music representations and SheetSage2's encoder, SOTA on 14 of 15 MARBLE metrics with 91.72% genre accuracy on GTZAN.
• A packaged agent skill, yue2-music, that teaches a coding agent to generate, transcribe, cover, edit ABC scores, and check musical invariants. An editable score is exactly the kind of artifact an agent can reason over, which is probably the most underrated thing in this release.

The predecessor, YuE 1, scored 4.9165 on the same benchmark in March 2025 with a 36.38% phoneme error rate. Eighteen months later the open tier is trading blows with Suno and Mureka. The technical report is still listed as coming soon, so the architecture claims (28 layers, an AR-NAR Mixture-of-Transformers with shared attention and separate norms, projections, and MLPs) are currently repo-level description rather than peer-reviewed detail.

Key Takeaways

• An editable score is the actual product. YuE2 plans melody and chords in ABC notation before rendering, which gives you and your agents a handle no other frontier song model exposes.
• The benchmark lead is real but thin. 6.9632 SongBench Avg tops all 15 settings, and the team states the gap is not statistically significant. Suno still leads MuLan text alignment and phoneme error rate.
• Best-of-8 is a selection protocol, not a model. Eight candidates picked on musicality. Unselected single-shot behavior will be worse than the headline number.
• The score is what makes covers work. 0.647 CLEWS mAP with it, 0.006 without, same checkpoint. That is the strongest evidence in the release.
• CC BY-NC 4.0 on the weights blocks commercial use. Apache 2.0 covers only code and docs. Plan your pipeline accordingly, or plan to negotiate.
• 24 GB gets you 48 kHz stereo locally. No quantization, no API, no per-song credits. That is the part your homelab cares about.

Sources: YuE2 project page, multimodal-art-projection/YuE on GitHub, YuE2 benchmark documentation, yue2-v0.1.6 release, YuE2-3B model card, WildSongBench dataset, YuE 1 paper (arXiv 2503.08638), Music Business Worldwide on the Suno stream-ripping claim