NotaGen is an open-source model for symbolic music generation that applies large language model training paradigms to classical sheet music, producing compositions with notably strong musicality. Released under the MIT license, it has about 1,200 stars on GitHub and was updated in 2025.
Instead of generating raw audio, NotaGen composes in symbolic notation, the digital equivalent of writing a score. It borrows the recipe that made LLMs successful and adapts it to music in three stages: large-scale pre-training on about 1.6 million musical pieces, fine-tuning on roughly nine thousand high-quality classical works, and a final alignment stage that steers the model toward more musical output. The result is a generator that writes coherent classical-style pieces in ABC notation that can be rendered, edited and performed like any other score. Model weights are published on Hugging Face, a web demo lets anyone listen to generated pieces, and the method is documented in an arXiv paper.
Because the output is notation rather than audio, generated scores can be converted to MIDI, rendered for listening, or opened in notation software and edited like any human-written composition. The project positions symbolic generation as a complement to audio generators: musicians can inspect the harmony and voice leading, correct passages and arrange the piece, instead of accepting a fixed recording.
NotaGen is completely free and open source under the MIT license. The code, weights and demo are available at no cost; you only need your own compute for generation.
NotaGen is for music AI researchers, composers and developers interested in generative music that produces editable scores rather than opaque audio files.