All rankings / Music

Music · instrumental

Blind listener preferences published by Carnegie Mellon University’s Music Arena. Vocal and instrumental results are separate populations.

15 entries · Music Arena score · Higher is better

Source publication: 2026-07-31. Retrieved 2026-11-05.

Source: Carnegie Mellon University · Music Arena · MIT.

Ranks use confidence bounds. Several models can share a rank; the rows within a shared rank are displayed by score.

Music · instrumental · Music Arena score
Significance rankModel / evaluated variantMusic Arena score95% confidence intervalVotes
1Lyria 3 ProGoogle DeepMind1,3941,336 – 1,452241
1Stable Audio 3 (Medium)Stability AI1,3331,295 – 1,371444
3Riffusion FUZZ v1.0Producer.ai1,2251,159 – 1,291220
3Sonauto v2.2Sonauto1,1501,109 – 1,191469
3Riffusion FUZZ v1.1Producer.ai1,1191,076 – 1,162452
5ElevenLabs Music v1ElevenLabs1,0551,015 – 1,095330
5Lyria 3 ClipGoogle DeepMind1,003929 – 1,07785
6Sonauto v3 PreviewSonauto950855 – 1,04541
7Magenta RealTime (Large)Google DeepMind911885 – 937829
7ACE-Step 1.5 Turbo (1.7B)ACE Studio / StepFun894839 – 949176
9Stable Audio OpenStability AI842812 – 872694
9MusicGen MediumMeta838807 – 869569
9Magenta RealTime 2 (Base)Google DeepMind809740 – 878200
11MusicGen SmallMeta794759 – 829463
15Stable Audio Open SmallStability AI682638 – 726435

What this ranking can tell you

  • The date is the end of the source’s evaluation window, not the date this page was fetched.
  • These results cover the source’s tested model versions. They do not establish the quality of later releases.
  • Confidence bounds are reconstructed from the source’s rounded score and ± interval. Small differences should not be interpreted as decisive.
  • No audio is redistributed. A benchmark result does not grant rights to a model or its generated music.

Check the source

The original evidence is retained byte for byte. The downloadable board includes the source URLs, revision and SHA-256 hashes.

Download this board and provenance →

Source files and licences

Explore other model types →Read the benchmark guide →

Evidence & Ask