YuE2 Music skill 1.2.0: instrumental generation and covers from a description, ABC, or a recording

The YuE tool has released YuE2 Music skill 1.2.0, which adds generation of instrumental music and instrumental covers from a text description, an ABC score, or a reference recording. Details are in the GitHub release. The yue2-music.zip package is installed into the agent's skills directory; the recommended agent is GPT-6 Astra. By default, YuE2 writes the score itself and generates the audio, while the agent runs the workflow and checks the result. For a cover, vocal notes are transferred to an instrumental track while preserving pitch, onset, and duration; audio references go through SheetSage2. Inference requires a compatible NVIDIA GPU; the weights and dependencies are installed separately. The author reports 51 instrumental CPU tests and 35 regression checks of the packaging.

What it means

In our reference, YuE is described as an open source audio generator from M-A-P with the text-generation task. The code is publicly available under the Apache-2.0 license, and the weights are also Apache-2.0. The largest weights file is 4.6 GB, and all files together are 11.6 GB. The model is YuE-s1-7B-anneal-en-cot on Hugging Face, and the code language is Python. The last code change in the repository is dated 2025-06-04, and the repository was created on 2025-01-23. The authors describe YuE as an open model for generating full-fledged songs, similar to Suno.ai but with open source. Release 1.2.0 concerns specifically the yue2-music skill: in YuE's fact sheet on our tool page, this skill is not described separately, nor are the GPU requirements for inference, the exact model revisions, or the role of the GPT-6 Astra agent. The developer does not specify whether the base model weights change in the process or whether this is only a workflow on top of the existing model. VRAM requirements are not stated in the release. For those already working with YuE, the update adds a scenario for instrumental generation and for transferring a melody into a cover without vocals. To assess how much this changes the base model, the fact sheet does not yet have enough data.