YuE2-3B: An Open Model for Music Generation with Editable Score

The model m-a-p/YuE2-3B with the text-to-audio task has been published on HuggingFace. The author is m-a-p, whose tools are already covered in our reference. The model weighs 3.6 billion parameters, the largest weights file takes up 6.8 GB in safetensors format, and the architecture is built on flow matching. The license is cc-by-nc-4.0, meaning use is permitted for non-commercial purposes only. The author describes YuE2-3B as a music generation model that turns song lyrics and a style description into a full track with vocals and accompaniment. A separate capability is editing the melody and chords via a score in ABC format. According to the developer, running on your own hardware requires a GPU with 24 GB of VRAM; without quantization, the model outputs 48 kHz stereo audio.

What it means

Our audio generators section doesn't yet have tools for creating music: the closest in task are XTTS from Coqui and Fish Speech from Fish Audio, but both handle text-to-speech, that is, speech synthesis rather than music. Closer in task type is ChatTTS, whose fact sheet lists the text-to-audio task, but it too is geared toward conversational speech. In terms of hardware requirements, YuE2-3B is noticeably heavier: ChatTTS asks for a minimum of 4 GB of VRAM, whereas here 24 GB is stated. On licensing, YuE2-3B matches ChatTTS — both models are distributed under cc-by-nc-4.0, which limits use to non-commercial scenarios. XTTS has stricter weights licensing: the Coqui Public Model License prohibits even indirect payment and training other models for commercial use. YuE2-3B is the first model we've seen that covers the music generation task with the ability to edit the score, and in terms of VRAM requirements it targets workstations with modern GPUs.