Five days after Suno shipped v6 with three major labels standing behind it, Multimodal Art Projection (M-A-P) released YuE2-3B — open weights, 3 billion parameters, and a benchmark table that puts it ahead.
The score deserves its caveats, and we’ll get to them. The architecture is the more durable story.
White-box generation
Nearly every song-generation system is a black box in the same specific way: you supply lyrics and a style prompt, it returns finished audio, and if the result is wrong your only real move is to reroll and hope. There’s nothing in between the prompt and the mix to grab.
YuE2 splits that in half. It first performs symbolic planning — writing an actual musical score, in editable ABC notation, with melody and chords. M-A-P’s framing is that you can “read, play, and change the composition before rendering.” Only then does the model generate semantic tokens and acoustic latents, with a VAE decoding to stereo audio at 48 kHz.
The practical difference is the difference between rerolling and editing. If the chorus lands on the wrong chord, that’s now a fixable object rather than a reason to regenerate the entire track and lose everything that was working.
The backbone is described as “one AR–NAR Mixture-of-Transformers” that handles both the score and the semantic tokens, with acoustic latents produced via flow matching.
One checkpoint, three jobs
The same weights do generation, zero-shot covers (transcribe an existing song, reimagine it in another style), and agentic editing — taking musical feedback in ordinary language and converting it into revisions of score, style, and lyrics. That last capability is what makes the symbolic layer more than a viewing pane: the edits land on the score, and the score is what gets rendered.
About that benchmark
On WildSongBench (192 prompts), M-A-P reports:
| Setting | SongBench Avg |
|---|---|
| YuE2, best-of-8 | 6.9632 |
| Mureka 9 | 6.9377 |
| Suno v5 | 6.8721 |
| YuE2, standard | 6.7316 |
| Suno v6 | 6.5562 |
Two things to hold onto. First, WildSongBench is M-A-P’s own benchmark, published alongside the model that wins it — the oldest pattern in machine learning, and a reason to treat the ordering as a claim rather than a finding until someone independent reproduces it.
Second, look at the Suno rows: v6 scores lower than v5. That’s a strange result for a model release that the industry treated as a significant step up, and it points at either a real regression on the axes this benchmark weights, or a benchmark that doesn’t capture what v6 improved. Either way it’s a caution against reading the leaderboard as a straight quality ranking.
What survives the skepticism is the shape of the gap: a 3B open model is now in genuine contention with commercial systems, which was not true a year ago. The standard (non-best-of-8) score sitting below both Suno versions is also worth noting — the headline number depends on generating eight candidates and picking one.
Running it
Linux, Python 3.12, and an NVIDIA GPU with BF16 support and 24 GB of VRAM. That’s a 4090 or better — out of reach for a laptop, ordinary for a workstation or a rented hour of cloud time.
Licensing splits the way M-A-P’s releases usually do: code and documentation Apache 2.0, model weights CC BY-NC 4.0. Non-commercial. Anyone planning to put generated music into client work, a game, or a release should read that line twice — it’s the single biggest practical difference between this and the commercial services it’s being benchmarked against, and it isn’t a detail the benchmark table captures.