YuE2-3B: a free, open music model that rivals Suno, on a 4090
The 3.6B weights turn lyrics into full 48kHz songs, do zero-shot covers, and edit as ABC scores — topping an in-house benchmark, but non-commercial.
Copy markdown
Beats Suno v5 and v6 — on its makers' own benchmark
On WildSongBench (192 prompts), YuE2's best-of-8 scored 6.9632, edging Suno v5 (6.8721) and Suno v6 (6.5562) — the highest mean across 17 tested settings. It's the team's own eval, so read it as a strong signal, not gospel: the headline is that an open model now sits in the same tier as the best paid one.
Three things it does that Suno can't
YuE2 plans a song as an editable ABC score before rendering audio, so you can hand-tweak melody and chords; it does zero-shot covers (restyle any track's tempo, lyrics, arrangement); and it edits agentically across turns instead of one-shot regen. Output is 48 kHz stereo, English or Mandarin lyrics.
Runs on a single 4090
The weights are 7.3 GB in BF16 and fit on one 24 GB GPU; a 4090 renders a 3.6-minute song in about 71 seconds. Build this weekend: a local cover-song machine or a score-to-audio pipeline that never touches a paid API.
The catch: the weights are non-commercial
YuE2-3B ships under CC BY-NC 4.0 — research and personal use only, no commercial deployment, a step back from the original YuE's Apache 2.0. The upside: it was trained on 346k hours of mostly CC0 and synthetic audio, so provenance is cleaner than most.