The open-source research collective Multimodal Art Projection (M-A-P), with the Hong Kong University of Science and Technology and other partners, released YuE2 on Sept. 9, 2026. YuE2 is a song generator that links text and audio: from lyrics and a style description, it first writes the melody and chords as an editable score in ABC notation, then turns that score into a full song with vocals and accompaniment. Writing the score first gives musicians a plan they can change, instead of only a finished recording from a closed commercial service.
The version on Hugging Face, YuE2-3B, is listed at about 3.6 billion parameters, the numbers a model tunes as it learns. The team says one model writes the score and then builds the sound from it in stages, and a separate component turns the result into stereo audio. Its trained files are free to download for non-commercial use under a CC BY-NC 4.0 license, and the code for running it is on GitHub. The team says YuE2 made a 3.6-minute song in 71 seconds on a single Nvidia RTX 4090, a graphics card with 24 GB of memory, using about 11 GiB of it at peak.
The project page says YuE2 was trained on about 346,000 hours of material, mainly copyright-free CC0 music and licensed synthetic data, with Tokenwave.AI supplying most of the synthetic set. Partners listed on the official page include NYU, Stanford, MBZUAI, NOIZ and ACE Studio.
In the team's own automatic song ratings, updated on Sept. 12, 2026, YuE2's best of eight tries scored higher than Suno v5 and Suno v6. The team notes that the models did not get the same number of tries to choose from. Its full technical report is listed as coming soon, and for now the team asks people to cite its 2025 paper on the original YuE.
No outside listening test has been published, so it is not yet clear whether people hear the gap the automatic ratings show. The license allows only non-commercial use, which could limit what companies build with YuE2.