YuE2 runs a Suno-class music model on one 24GB GPU
The open 3B model writes editable ABC scores before it sings and edges Suno v5 on WildSongBench — part of a wave of local media models surging on Hugging Face.
Copy markdown
Full songs on a GPU you already own
YuE2-3B turns lyrics plus a style prompt into complete stereo songs — vocals and accompaniment at 48kHz — running locally on a single 24GB NVIDIA card (7.3GB checkpoint, Linux + Python 3.10+). No credits, no cloud queue.
It drafts the score before the audio
Instead of a black box, YuE2 first writes melody and chords in editable ABC notation — or you hand it your own score to sing. On WildSongBench (192 prompts) it scored 6.96 to Suno v5's 6.87.
Native in ComfyUI today
Support landed natively in ComfyUI (PR #16250 merged to master), so you can drop music generation into a local pipeline and chain it with TTS or video nodes.
Elsewhere: open 4K video from Lightricks
LTX-2.5 (22B, open weights) is trending alongside it — text-, image-, and video-to-video plus audio up to 4K HDR, with the distilled model rendering a 10s clip in ~6-8s on top NVIDIA chips. Pair the two for a fully local A/V stack.