Open-source music model YuE2-3B launches with editable scores, claims Suno v5-level quality

M-A-P released YuE2-3B, an open-weight music generation model now trending on Hugging Face, with a live demo at map-yue2.github.io. Given lyrics and a style prompt, it generates complete songs up to 5 minutes with vocals and accompaniment at 48kHz stereo, runs locally on a 24GB GPU, and supports Chinese and English for creation, cover, and editing workflows. Early testers report impressive sound, though the non-commercial license limits commercial use.

Confirmed

• The model is roughly 3B parameters (some sources cite 3.59B, 28 layers, AR/NAR expert architecture), with an accompanying arXiv paper and tags including music-generation and agentic-editing
• Its "white-box" pipeline first produces an editable symbolic plan of melody and chords in ABC notation that users can inspect, play, and edit before rendering the final song
• The team claims YuE2 (best-of-8) beats Suno v5 on WildSongBench; other accounts place it between Suno 4.5 and 5.5, surpassing Mureka v9 and MiniMax 3 on several benchmarks
• Community reactions: @realmrfakename found the output excellent and comparable to Suno v5/v5.5 but was disappointed by the non-commercial license; @wzwowzw0002 asked about ComfyUI integration; @bdsqlsz questioned whether it truly beats Suno V5, awaiting community testing; @Acceptable-Cycle4645 called the demo quite good

Unconfirmed

• Claims of matching or beating Suno v5/v5.5, Mureka v9, and MiniMax 3 come mainly from the developers and individual listeners; broader independent evaluation is still lacking

Why it matters

• Editable scores make generation transparent and controllable, offering musicians a "plan first, render later" workflow unlike most end-to-end black-box models
• Local 48kHz stereo full-song generation on a 24GB consumer GPU significantly narrows the gap between open-source and closed commercial systems like Suno
• The non-commercial license remains its biggest limitation, leaving practical commercial prospects unclear

2026-09-10 ~ 2026-09-11 · 12 related posts

Full story(5 episodes)→

• Episode 1: Open-source music model YuE2-3B launches with editable scores, claims Suno v5-level quality(2026-09-10, 12 posts)
• Episode 2: Open-Source Music Model YuE2 Beats Suno v5/v6 on WildSongBench(2026-09-14, 4 posts)
• Episode 3: YuE2 Music Model Gets ComfyUI Nodes and LoRA Trainer(2026-09-15, 2 posts)
• Episode 4: Open-Source Music Model YuE2 Runs Suno-Quality Songs Locally in Just 4.5GB VRAM(2026-09-15, 7 posts)
• Episode 5: YuE2 allows individuals to commercialize generated music for free(2026-09-16, 2 posts)

• [source] Open-source YuE2 music model generates editable symbolic scores before rendering full songs — GreyScope · 2026-09-10
• Open-Weight Music Model YuE2-3B Released, Claimed to Rival Suno v5.5 — realmrfakename · 2026-09-10
• YuE2-3B: Open Music Model With Editable Scores Beats Suno v5 on WildSongBench — realmrfakename · 2026-09-10
• YuE2 combines symbolic planning and audio generation, claims Suno v5-level music quality — wzwowzw0002 · 2026-09-10
• Open-source 3B music model Yue2-3B released, author asks if it beats Suno V5 — bdsqlsz · 2026-09-10
• [source] m-a-p releases YuE2-3B, an open music generation model trending on HF — m-a-p · 2026-09-10
• YuE2-3B open-source music generation model released with demos — Acceptable-Cycle4645 · 2026-09-11
• Open-source music model YuE2 claims to beat Mureka v9 and MiniMax 3, near Suno 4.5-5.5 — _AustinCalvert_ · 2026-09-11

4 near-duplicate retellings: realmrfakename · realmrfakename · solyarisoftware · bdsqlsz