YuE2 is a music model that writes a score, then a full song. The weights are CC BY-NC 4.0, and the card's scores sit beside the paper's listening test.
One checkpoint, three jobs
The paper describes one checkpoint of about 3.58 billion parameters, trained on about 346,000 hours. It writes a readable score first: melody, chords, key, meter, tempo, and form, in ABC notation. It then makes semantic tokens and 48 kHz stereo audio. Creation is the model writing that score from a style prompt and lyrics. Editing is you handing it a revised score. A cover starts with SheetSage2 transcribing a recording, then the same checkpoint rendering that score in a new style. The paper says this cover path was not trained on original-and-cover pairs. SheetSage2 and MERT2 are separate models.
The model card says the local run is Linux, Python 3.10 or newer, and a 24GB NVIDIA GPU with BF16, without quantization. The inference wheel on the card is yue2_infer-0.1.5, installed with huggingface-hub==0.36.2. Planning has three switches. cot="full" plans melody and chords, and is the default. cot="melody" plans melody only, and is what the card recommends for covers. cot="off" generates with no symbolic plan. For a cover, the card says to save a melody ABC without chord symbols. Melody mode does not strip chord symbols by itself.
The card's prose says a 3.6-minute song took 71 seconds on an RTX 4090. The table under that, for full planning, lists 71.04 seconds of generation for 214.85 seconds of audio, with a peak of 11.18 GiB. The same section says to use a 24GB GPU and 24GB of free host memory, and that a maximum-context test peaked at 14.08 GiB.
Two decoders, and two kinds of score
The card names two decoders. YuE2-Vae is the default, and the card says it has the better perceptual audio quality. The three demo songs on the card use that decoder. YuE2-Vae-legacy is the one the card says scores higher on musicality, and the WildSongBench table says both YuE2 rows use it. On that table, the plain YuE2 row selects from two candidates. Best-of-8 selects from eight.
Those automatic SongBench averages, on 192 prompts, are printed as YuE2 6.7316 and best-of-8 6.9632. On the same table Suno v5 is 6.8721, Suno v4.5 is 6.6995, Suno v5.5 is 6.7150, Suno v6 is 6.5562, Suno v6 Wild is 6.4195, and Mureka 9 is 6.9377. The card's headline calls the best-of-8 number the highest in that set. The plain YuE2 row sits under Suno v5 and Mureka 9, and over the other Suno rows listed there. The card dates that comparison 12 September 2026.
The paper's expert listening is a separate count. With the same checkpoint, 49.3% of overall-quality judgments preferred symbolic planning and 34.6% preferred generation without it. The rest were ties. Experts preferred best-of-8 over Suno v4.5, and the paper describes the overall preference against Suno v5 as nearly balanced. It also gives a mean tie-adjusted audio-quality preference of 58.9% across six proprietary systems.
The card's cover table, 948 works and no candidate picking, uses the legacy decoder. With a full score, song-identity mAP is 0.647 and Musicality is 5.104. Without chord symbols those are 0.598 and 5.490. With no score they are 0.006 and 5.691. The card says identity and quality should be read together.
Noncommercial weights, and a hosted page that says free
The model card says the weights are CC BY-NC 4.0, and that third-party code has its own notices. The same card calls YuE2 an open music model. The paper says training used mostly CC0 music and synthetic data, and that Tokenwave.AI supplied most of the synthetic data under license. Code is at multimodal-art-projection/YuE.
The project page links a try button to yue.noizai.net and labels it free, hosted by NOIZ. That hosted page prints "3 left" beside Generate and beside Generate cover, and it also prints "Free to use, endless to explore." The paper, the model card, and that hosted page do not print a dollar price.