OpenMOSS releases YuE2-3B for music generation

OpenMOSS has published YuE2-3B, a 3-billion-parameter music generation model, on Hugging Face. The model targets full-song generation and supports both English and Chinese lyrics, and it arrives under a non-commercial CC-BY-NC-4.0 license.

What distinguishes YuE2 from a straightforward text-to-audio system is its stated approach: the model combines symbolic planning with agentic editing. In practice that means the system can reason about musical structure before committing to audio, then revise its output rather than producing a single one-shot render.

Why it matters

Open music models have lagged behind their text and image counterparts, in part because coherent long-form composition is hard and high-quality training data is scarce. A mid-sized, openly available model that leans on structured planning is a useful data point for anyone building or studying generative audio tools.

• Roughly 3B parameters, a dense (non-MoE) architecture
• Bilingual support for English and Chinese
• Symbolic planning plus agentic editing for iterative refinement
• Released under CC-BY-NC-4.0, limiting commercial use

The non-commercial license means YuE2-3B is best understood as a research and experimentation release rather than a drop-in tool for commercial production. Full details and weights are available on the model's Hugging Face page.