YuE
Lyrics in, a full song out, runnable locally; VRAM and licence are the gates.
Model profile
- Vendor
- YuE 项目（M-A-P / 中国香港科技大学）
- Form
- 本地模型
- Input
- 歌词 · 文本
- Output
- 音频
- Licence
- 代码多为 Apache 2.0，权重按版本另标；较新权重默认非商用、个人创作者另有豁免，商用前读仓库许可
- Verified
- 2026-10-31
Related
Count the VRAM first
YuE is one of the few open-weight models that finishes a whole song: give it lyrics and a style and it returns vocals plus backing, several minutes long. It is not a browser tool; you install an environment, download weights and bring a large enough GPU before anything starts. Official and packaged setups often recommend 24GB of VRAM; quantised builds and community inference stacks can squeeze into 8–12GB at the cost of speed and stability. Settle that before judging the music.
How the lyrics get in
Input is line-broken lyrics plus a style description; verse, chorus and bridge tags can sit inside the lyric text. Output is a mixed stereo file, and some inference scripts write the backing and vocal separately. There is no instant one-line-to-song loop: a three-minute song usually takes tens of seconds to a few minutes on a consumer GPU, so experiments are measured in minutes. Newer builds first lay out melody and chords as a readable score plan, then render audio, which is easier to revise than a pure black box.
Long-form coherence is the point
Most open models manage a few dozen seconds; YuE's value is the whole song, with the lift into the chorus, the return after the bridge and the closing all in one pass. Chinese songs are a strength: pentatonic lines and sparse, long-held syllables hold up better than rapid English verses. For a dialect or guofeng demo it gives a structurally complete reference that you can arrange or hand to a singer.
Three walls: VRAM, diction, repetition
- VRAM and speed: 24GB is the comfortable line; under 12GB means quantisation and a higher chance of a long song collapsing.
- Diction: Mandarin tones blur on fast lines, stacked syllables and high notes. Check line by line against the lyric, not just the waveform.
- Repetition: past three minutes it often returns to earlier melody and orchestration, most audible in the second chorus.
- Stems: if a script writes vocals and backing separately, the vocal track carries reverb tails and residue and needs repair before remixing.
Start with a short song; do not open with five minutes.
Licence layers and alternatives
Read three layers: code, weights and companion models. Early releases were Apache 2.0; newer weights default to non-commercial with an exemption for individual creators, and companies need a separate licence. The repository licence file is the authority. Alternatives: Suno or MiniMax Music for a browser service; DiffRhythm for faster local full songs; Magenta for symbolic control. Each may cost you Mandarin diction or whole-song coherence, so choose by the job.
Listen in the library
All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).
Along the same line
Sources
- YuE 项目仓库与模型卡（M-A-P / 中国香港科技大学）关于输入形式、显存建议与权重许可的说明。
- YuE 技术报告对长曲生成与歌词对齐的公开描述。
- 公开用户报告中的显存、速度与咬字问题，只取现象，不引用具体评测分数。