Run YuE2 locally: VRAM, install steps, Mac and GGUF

Model files download from Hugging Face on first use. The generation checkpoint is m-a-p/YuE2-3B and the default decoder is m-a-p/YuE2-Vae; YuE2-Vae-legacy exists for reproducing the paper’s benchmark, and SheetSage2 is the transcription model used for covers.

Maximum-context testing peaked at 14.08 GiB. The measured peaks are well under 24 GB, but the authors still specify a 24 GB GPU and 24 GB of available host RAM as the supported baseline, and their agent skill warns against silently shortening a song or lowering settings to hide an out-of-memory error.

git clone https://github.com/multimodal-art-projection/YuE.git
cd YuE
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install .
python examples/generate.py --output outputs/first-song

The first run writes outputs/first-song/audio.flac and keeps the score, semantic tokens, acoustic latents and generation settings beside it. A staged Python API exposes plan(), generate_semantic(), synthesize() and decode() when you want to edit the score between steps.

The original YuE model is not the same code: it is preserved on the YuE-v1 branch of the same repository with its own documentation and licence.

The official quick start lists Linux and NVIDIA only, with no macOS build. Running YuE2 on a Mac currently means a third-party MLX port. These projects are independent of the YuE2 authors, and this site has not tested them.

The YuE2 Studio author reports a 3:20 song taking 337 seconds at 32 steps on a MacBook Pro M5 Pro with 24 GB, with a 10.4 GiB memory peak. Treat that as one user’s measurement rather than a specification. Intel Macs are not supported by either project.

The reference runtime runs without quantization. Smaller footprints come from other projects:

• ComfyUI’s native nodes load their own checkpoint, yue2_3b_int8_convrot.safetensors, which is an int8 build going by its filename.
• yue2.cpp is a community C++ implementation that loads GGUF files and runs on CPU, CUDA or Vulkan. It includes an embedded web UI served from a local port.

The maintainer describes Q8_0 as near lossless and publishes nothing below Q5. These are third-party conversions that this site has not tested; the weights keep their CC BY-NC 4.0 licence.

The code, agent skill and documentation are Apache 2.0. The model weights are CC BY-NC 4.0 with an additional creator permission. The authors summarise it this way:

This is a summary, not legal advice. The full terms are in the MODEL_LICENSE file of the upstream repository. Tracks made in this studio are covered by the yue2.pro Terms.

A local install pays off when you want the editable score, covers from your own transcriptions, or many generations on hardware you already own. If you only want to hear what the model does with a set of lyrics, the browser studio sends a style brief and lyrics to hosted YuE2-3B inference and returns an MP3, with the credit cost shown before you submit.