Skip to main content

m-a-pAudio modelOpen weights

YuE2-3B

Popularityno data yet

n/ano search or usage data

Summary

YuE2-3B generates complete songs from lyrics, style directions, and musical instructions. The open music checkpoint supports vocals and accompaniment, plus editable scores, melody and chord control, covers, and agentic editing.

Its documentation cites 24 GB of GPU memory for local execution. The weights use CC-BY-NC-4.0, so they are not unrestricted authorization for commercial music production.

GraduallyBench Audio

YuE2-3B

Technical test
  • Only a technical test is available.
  • The selection of test inputs is limited.
  • The test used quantized model weights.
  • The tested weights have noncommercial usage terms.
  • The functional test does not assess content quality.
Show test details

Technical measurement

Local · Language not documented

Execution Apple Metal. Precision Q4_0 · Decoder precision FP16.

2 valid outputs from 2 attempted and 2 planned test slots.

No further scored metric is available for this test.

WER compares words and CER compares characters with the reference. The speech output was re-transcribed with GPT Transcribe. Its exact backend revision is unconfirmed. Read speech and digital silence do not cover spontaneous speech or real background noise. Error rates can exceed 100 %. Where used, the test references were not independently reviewed by humans. Test coverage and methods differ by model. Each metric stands on its own; there is no combined overall score across tests. Data compiled on 11/10/2026. How we test · Download all Decision and Audio results as JSON

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 51 published benchmark results

Page 1 of 5

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

PER (Show measurement, test conditions, and source)
Source value
9.79
Metric
phoneme error rate
Unit
%
Benchmark version
best-of-8
Category
audio
Direction
Lower is better
Source type
vendor-reported
Evaluator
YuE2 model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Same 192-prompt run
Context
Source-specific observation; it does not establish a compatible cross-model cohort.
Source YuE2 model card
WildSongBench expert preference for melody-and-chord planning, overall quality (Show measurement, test conditions, and source)
Source value
49.3
Metric
Share of judged responses favoring symbolic planning, with ties retained
Unit
%
Benchmark version
YuE2 arXiv:2609.33757v1 Section 4.2, direct planning ablation
Category
audio
Direction
Higher is better
Sample
211
Source type
vendor-reported
Evaluator
YuE2 paper authors, arXiv
Status
active
Retrieved at
2026-10-17
Methodology
Vendor-reported result. Same exact YuE2 checkpoint, prompts, lyrics, two-candidate budget and waveform decoder. The planned condition first generates a melody-and-chord score; the unplanned condition generates semantic tokens and acoustic latents directly. The expert panel retained 211 judged overall-quality evaluations: 49.3% favored planning, 34.6% favored no planning and 16.1% were ties.
Context
This is an ablation of two generation modes within the same checkpoint, not a cross-model quality score. The preference share retains ties, so 50% is not the stated neutral threshold; the paper gives a significance test but no confidence interval in this section.
Source YuE2 paper authors, arXiv
VRAM16 YuE2-3B Japanese lyric character match, int8, with score, short Japanese lyrics (Show measurement, test conditions, and source)
Source value
73
Metric
Whisper large-v3 character match against lyrics
Unit
%
Direction
Higher is better
Sample
1
Source type
independent-evaluator
Evaluator
VRAM16, YuE2-3B local evaluation by ulluna
Status
active
Retrieved at
2026-10-22
Methodology
Independent evaluation. YuE2-3B, int8, with score, short Japanese lyrics. One generated song was transcribed with Whisper large-v3; the backing track remained during recognition. Tested 2026-10-02 on the page's RTX 5080 / ComfyUI 0.38.0 setup. The page says the standard workflow matches the official ComfyUI template audio_yue2_text2music (32 steps, CFG 1, sampler dpm_2 / sgm_uniform, seed 1234) with score in full mode; the cited no-score row is an explicit no-score variant.
Context
Only one of two generated songs per setting was transcribed. Whisper large-v3 was run with the backing track still present; the tested lyrics were Japanese only, and the source calls these match rates rough.
Source VRAM16, YuE2-3B local evaluation by ulluna
VRAM16 YuE2-3B Japanese lyric character match, bf16, with score, short Japanese lyrics (Show measurement, test conditions, and source)
Source value
84
Metric
Whisper large-v3 character match against lyrics
Unit
%
Direction
Higher is better
Sample
1
Source type
independent-evaluator
Evaluator
VRAM16, YuE2-3B local evaluation by ulluna
Status
active
Retrieved at
2026-10-22
Methodology
Independent evaluation. YuE2-3B, bf16, with score, short Japanese lyrics. One generated song was transcribed with Whisper large-v3; the backing track remained during recognition. Tested 2026-10-02 on the page's RTX 5080 / ComfyUI 0.38.0 setup. The page says the standard workflow matches the official ComfyUI template audio_yue2_text2music (32 steps, CFG 1, sampler dpm_2 / sgm_uniform, seed 1234) with score in full mode; the cited no-score row is an explicit no-score variant.
Context
Only one of two generated songs per setting was transcribed. Whisper large-v3 was run with the backing track still present; the tested lyrics were Japanese only, and the source calls these match rates rough.
Source VRAM16, YuE2-3B local evaluation by ulluna
VRAM16 YuE2-3B Japanese lyric character match, int8, no score, short Japanese lyrics (Show measurement, test conditions, and source)
Source value
100
Metric
Whisper large-v3 character match against lyrics
Unit
%
Direction
Higher is better
Sample
1
Source type
independent-evaluator
Evaluator
VRAM16, YuE2-3B local evaluation by ulluna
Status
active
Retrieved at
2026-10-22
Methodology
Independent evaluation. YuE2-3B, int8, no score, short Japanese lyrics. One generated song was transcribed with Whisper large-v3; the backing track remained during recognition. Tested 2026-10-02 on the page's RTX 5080 / ComfyUI 0.38.0 setup. The page says the standard workflow matches the official ComfyUI template audio_yue2_text2music (32 steps, CFG 1, sampler dpm_2 / sgm_uniform, seed 1234) with score in full mode; the cited no-score row is an explicit no-score variant.
Context
Only one of two generated songs per setting was transcribed. Whisper large-v3 was run with the backing track still present; the tested lyrics were Japanese only, and the source calls these match rates rough.
Source VRAM16, YuE2-3B local evaluation by ulluna
VRAM16 YuE2-3B Japanese lyric character match, int8, with score, long Japanese lyrics (Show measurement, test conditions, and source)
Source value
84
Metric
Whisper large-v3 character match against lyrics
Unit
%
Direction
Higher is better
Sample
1
Source type
independent-evaluator
Evaluator
VRAM16, YuE2-3B local evaluation by ulluna
Status
active
Retrieved at
2026-10-22
Methodology
Independent evaluation. YuE2-3B, int8, with score, long Japanese lyrics. One generated song was transcribed with Whisper large-v3; the backing track remained during recognition. Tested 2026-10-02 on the page's RTX 5080 / ComfyUI 0.38.0 setup. The page says the standard workflow matches the official ComfyUI template audio_yue2_text2music (32 steps, CFG 1, sampler dpm_2 / sgm_uniform, seed 1234) with score in full mode; the cited no-score row is an explicit no-score variant.
Context
Only one of two generated songs per setting was transcribed. Whisper large-v3 was run with the backing track still present; the tested lyrics were Japanese only, and the source calls these match rates rough.
Source VRAM16, YuE2-3B local evaluation by ulluna
VRAM16 YuE2-3B Japanese lyric character match, bf16, with score, long Japanese lyrics (Show measurement, test conditions, and source)
Source value
81
Metric
Whisper large-v3 character match against lyrics
Unit
%
Direction
Higher is better
Sample
1
Source type
independent-evaluator
Evaluator
VRAM16, YuE2-3B local evaluation by ulluna
Status
active
Retrieved at
2026-10-22
Methodology
Independent evaluation. YuE2-3B, bf16, with score, long Japanese lyrics. One generated song was transcribed with Whisper large-v3; the backing track remained during recognition. Tested 2026-10-02 on the page's RTX 5080 / ComfyUI 0.38.0 setup. The page says the standard workflow matches the official ComfyUI template audio_yue2_text2music (32 steps, CFG 1, sampler dpm_2 / sgm_uniform, seed 1234) with score in full mode; the cited no-score row is an explicit no-score variant.
Context
Only one of two generated songs per setting was transcribed. Whisper large-v3 was run with the backing track still present; the tested lyrics were Japanese only, and the source calls these match rates rough.
Source VRAM16, YuE2-3B local evaluation by ulluna
SHS100K zero-shot cover generation, full score, CLEWS Hit@1 (Show measurement, test conditions, and source)
Source value
71.3
Metric
CLEWS Hit@1 for preserved-song identity after source exclusion.
Unit
%
Direction
Higher is better
Sample
3792
Source type
vendor-reported
Evaluator
YuE2-3B model card
Status
active
Retrieved at
2026-10-23
Methodology
Vendor-reported result. SHS100K zero-shot cover generation, full score condition. The card reports 3,792 generated outputs per method without candidate selection; source-score extraction is separate. CLEWS and Discogs-VINet assess preserved-song identity against 10,545 recordings after source exclusion; MuLan measures target-style similarity, and Musicality is from SongBench.
Context
The card identifies YuE2-Vae-legacy for these YuE2 variants but gives no immutable weights revision. Values apply to the named condition and should not be equated with other datasets or selection protocols.
Source YuE2-3B model card
SHS100K zero-shot cover generation, without chords, CLEWS Hit@1 (Show measurement, test conditions, and source)
Source value
67.3
Metric
CLEWS Hit@1 for preserved-song identity after source exclusion.
Unit
%
Direction
Higher is better
Sample
3792
Source type
vendor-reported
Evaluator
YuE2-3B model card
Status
active
Retrieved at
2026-10-23
Methodology
Vendor-reported result. SHS100K zero-shot cover generation, without chords condition. The card reports 3,792 generated outputs per method without candidate selection; source-score extraction is separate. CLEWS and Discogs-VINet assess preserved-song identity against 10,545 recordings after source exclusion; MuLan measures target-style similarity, and Musicality is from SongBench.
Context
The card identifies YuE2-Vae-legacy for these YuE2 variants but gives no immutable weights revision. Values apply to the named condition and should not be equated with other datasets or selection protocols.
Source YuE2-3B model card
SHS100K zero-shot cover generation, without score, CLEWS Hit@1 (Show measurement, test conditions, and source)
Source value
0.3
Metric
CLEWS Hit@1 for preserved-song identity after source exclusion.
Unit
%
Direction
Higher is better
Sample
3792
Source type
vendor-reported
Evaluator
YuE2-3B model card
Status
active
Retrieved at
2026-10-23
Methodology
Vendor-reported result. SHS100K zero-shot cover generation, without score condition. The card reports 3,792 generated outputs per method without candidate selection; source-score extraction is separate. CLEWS and Discogs-VINet assess preserved-song identity against 10,545 recordings after source exclusion; MuLan measures target-style similarity, and Musicality is from SongBench.
Context
The card identifies YuE2-Vae-legacy for these YuE2 variants but gives no immutable weights revision. Values apply to the named condition and should not be equated with other datasets or selection protocols.
Source YuE2-3B model card
WildSongBench Musicality (Show measurement, test conditions, and source)
Source value
6.2666
Metric
score
Unit
No unit provided
Benchmark version
best-of-8
Category
audio
Direction
Higher is better
Source type
vendor-reported
Evaluator
YuE2 model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. 192 prompts; candidates selected Musicality then Q3O then PER
Context
Source-specific observation; it does not establish a compatible cross-model cohort.
Source YuE2 model card
WildSongBench SongBench average (Show measurement, test conditions, and source)
Source value
6.9632
Metric
score
Unit
No unit provided
Benchmark version
best-of-8
Category
audio
Direction
Higher is better
Source type
vendor-reported
Evaluator
YuE2 model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. 192 prompts; candidates selected Musicality then Q3O then PER
Context
Source-specific observation; it does not establish a compatible cross-model cohort.
Source YuE2 model card

Page 1 of 5

All specifications

Architecture

Model class
Music generation
Source: Model class
Checkpoint
m-a-p/YuE2-3B
Source: Checkpoint

Input and output

Input
Lyrics, style, and musical direction
Source: Input
Output
Songs with vocals and accompaniment
Source: Output
Capabilities
Complete songs, editable score, melody and chord control, covers, and agentic editing
Source: Capabilities

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking
Unknown
no source
Watermarking technique
Not known
no source

Licence and access

Access
Local weights
Source: Access
License
CC-BY-NC-4.0
Source: License
Shutdown
None announced
no source

Further details

Hardware note
The documentation cites 24 GB of GPU memory for local execution
Source: Hardware note
Repository
Official YuE repository
Source: Repository

Sources and data date

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.

Without vendor claims and historical or disputed results; those stay labelled in the single values.

TypeEvidence and data date
Research statusResearch date October 4, 2026. 3 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not published

confidence-intervals-for-standard-wsb-metrics

Note

The paper gives the common 192-prompt denominator for its WildSongBench tables but no per-metric confidence intervals there. The direct planning comparison has 211 judged responses and a p-value, but the section does not give an interval.

Additional sourceYuE2-3B model card (retrieved October 3, 2026; October 4, 2026; October 17, 2026; October 23, 2026) · YuE2-3B model card · Editorial description reviewed October 3, 2026
Additional sourceYuE repository (retrieved October 3, 2026) · YuE repository · Editorial description reviewed October 3, 2026
Additional sourceYuE2 paper authors, arXiv (retrieved October 17, 2026; October 18, 2026)
Additional sourceVRAM16, YuE2-3B local evaluation by ulluna (retrieved October 22, 2026)