YuE2: open music model edges Suno v5, runs on a 24GB GPU
The M-A-P team's weights expose an editable ABC score, do zero-shot genre covers, and fit one card — but they ship under a noncommercial license.
Copy markdown
It edges Suno v5 on the open bench
M-A-P open-sourced YuE2, a 3B AR-NAR model that turns lyrics plus a style prompt into a full vocal track at 48kHz stereo. With best-of-8 sampling it scores 6.96 on WildSongBench’s 192 prompts, nudging past Suno v5’s 6.87 — the first open model to trade blows with a frontier paid one.
You edit the score, not just re-roll the prompt
YuE2 plans a song as an editable symbolic layer first — melody and chords exposed in ABC notation — so you can revise the actual notes before it renders audio. That turns music generation from a prompt slot machine into something you can direct bar by bar.
Zero-shot covers and agentic re-edits
Ask for a jazz-funk or heavy-metal version of a track and it reharmonizes zero-shot; an agentic loop lets you iterate across rendered takes instead of starting from scratch. English and Mandarin vocals both work out of the box.
One 24GB card, native in ComfyUI
The checkpoint is 7.3–7.8GB and runs unquantized on a single 24GB NVIDIA GPU with Python 3.10+, and native ComfyUI support landed the same day. A weekend of glue and you have a local, editable Suno on your own box.
The catch: noncommercial weights
YuE2 ships under CC BY-NC 4.0, so it is free for prototypes, demos, and personal tracks but off-limits for anything you would sell without asking M-A-P first. Read the license before you wire it into a product.