YuE
An open-source music generation model series designed to produce complete musical pieces, including vocal singing and instrumental accompaniment, from text prompts or lyric inputs. The models handle the mapping of lyric content to vocal output as a core part of their generation pipeline.
Core Facts
YuE is an open-source music generation model series capable of generating complete music clips with vocal singing and accompaniment based on text prompts or lyrics
Based on lyrics and style prompts provided by the user, Yue2 generates a symbolic arrangement of melodies and chords presented in ABC notation format.
Timeline (last 90 days)
This porting solution has been open-sourced on GitHub, contains a six-step process, and requires pulling a 12.2G model and verifying its hash.
The official command-line and T8 code paths produce audio of identical duration bit-for-bit under the same seed and lyrics, both 174.91866 seconds.
FP8 quantization must absolutely not be enabled in this porting solution, as it causes relative errors as high as 90% to 117% and will silently produce garbage audio.
Four capability test results: song generation producing 175 seconds of audio, audio-to-score transcription of a 4-second source taking 36 seconds, sheet music rendering outputting a PDF and a playback audio file, and reference voice timbre completed in 47 seconds.
Reference voice timbre is a post-processing step added by the T8 author (separating vocals, swapping timbre, then remixing) that has zero code references to the YuE model itself and differs from the official cover-song feature.
Using a single AMD 9070XT GPU with 16G VRAM, a complete 2-minute-55-second Chinese song titled 《那年夏天的诺言》 was locally generated via the YuE model, including lyrics, composition, vocals, and accompaniment.
The YuE model weights were jointly released by the Moonnight 2 (m-a-p) team, the Hong Kong University of Science and Technology, and MAP, under a non-commercial license that permits only research and educational use.
In this real-world test, peak VRAM usage was only 11–12G, far below the official requirement threshold of 24G.
Porting YuE to the ROCm platform requires only five source-level code changes in total.
After changing the VAE chunked frame count from 1024 frames to 512 frames, the processing time for the same request dropped from 613 seconds to 393 seconds, a 35.9% speedup, with VAE decoding reduced from 313 seconds to 107 seconds.
4 more timeline events
All Facts (14)
YuE is an open-source music generation model series capable of generating complete music clips with vocal singing and accompaniment based on text prompts or lyrics
70%VerifiedBased on lyrics and style prompts provided by the user, Yue2 generates a symbolic arrangement of melodies and chords presented in ABC notation format.
65%UnverifiedPorting YuE to the ROCm platform requires only five source-level code changes in total.
50%UnverifiedThe official command-line and T8 code paths produce audio of identical duration bit-for-bit under the same seed and lyrics, both 174.91866 seconds.
50%UnverifiedFP8 quantization must absolutely not be enabled in this porting solution, as it causes relative errors as high as 90% to 117% and will silently produce garbage audio.
50%UnverifiedFour capability test results: song generation producing 175 seconds of audio, audio-to-score transcription of a 4-second source taking 36 seconds, sheet music rendering outputting a PDF and a playback audio file, and reference voice timbre completed in 47 seconds.
50%UnverifiedReference voice timbre is a post-processing step added by the T8 author (separating vocals, swapping timbre, then remixing) that has zero code references to the YuE model itself and differs from the official cover-song feature.
50%UnverifiedThe YuE model weights were jointly released by the Moonnight 2 (m-a-p) team, the Hong Kong University of Science and Technology, and MAP, under a non-commercial license that permits only research and educational use.
50%UnverifiedIn this real-world test, peak VRAM usage was only 11–12G, far below the official requirement threshold of 24G.
50%UnverifiedAfter changing the VAE chunked frame count from 1024 frames to 512 frames, the processing time for the same request dropped from 613 seconds to 393 seconds, a 35.9% speedup, with VAE decoding reduced from 313 seconds to 107 seconds.
50%UnverifiedThe YuE series follows an open-source, self-hostable approach, differentiating itself from closed-source commercial products such as Suno and Udio.
50%UnverifiedYuE2-3B is an iterative version of the YuE series.
50%UnverifiedUsing a single AMD 9070XT GPU with 16G VRAM, a complete 2-minute-55-second Chinese song titled 《那年夏天的诺言》 was locally generated via the YuE model, including lyrics, composition, vocals, and accompaniment.
50%UnverifiedThis porting solution has been open-sourced on GitHub, contains a six-step process, and requires pulling a 12.2G model and verifying its hash.
50%