FreeLLMFree AI music generators New
Text-to-music models ranked by people who listened to two tracks without knowing which model made them, and picked the better one. Green rows are open models: you can download them and make music on your own computer, free.
Ranking
Score from blind votes (1000 = average), axis starts at 750. Green = open weights you can download, grey = closed.
What this shows · How to read it
What it shows. Every music model with at least 200 blind votes on Music Arena, by how often listeners preferred it.
How to read it. Longer bar, preferred more often. 1000 is the average model; 100 points more means it wins about 64% of the time against an average one. Green bars are open models you can download.
| # | Model | Score | Votes | Weights | Size | Speed (open models) | Last vote |
|---|---|---|---|---|---|---|---|
| 1 | Stable Audio 3 Medium Stability AI |
1226 ±39 |
496 | Stability AI Community License free for individuals and companies under $1M a year in revenue |
2.31 B | faster than real time 0.02 s of compute per s of music |
30 Jul 2026 |
| 2 | Lyria 3 Pro (preview) Google |
1183 ±24 |
1K | Closed | — | — | 29 Jul 2026 |
| 3 | Riffusion FUZZ 1.0 Riffusion |
1181 ±28 |
863 | Closed | — | — | 20 Feb 2026 |
| 4 | Riffusion FUZZ 1.1 Riffusion |
1132 ±22 |
1K | Closed | — | — | 19 Feb 2026 |
| 5 | Sonauto v3 (preview) Sonauto |
1081 ±40 |
308 | Closed | — | — | 15 Apr 2026 |
| 6 | Lyria 3 (30 s) Google |
1054 ±32 |
470 | Closed | — | — | 15 Apr 2026 |
| 7 | Sonauto v2.2 Sonauto |
1028 ±16 |
2K | Closed | — | — | 31 Jul 2026 |
| 8 | Eleven Music v1 ElevenLabs |
1009 ±16 |
2K | Closed | — | — | 31 Jul 2026 |
| 9 | ACE-Step 1.5 Turbo ACE-Step |
953 ±19 |
1K | MIT | 1.7 B under 4 GB of VRAM, says the official card |
faster than real time 0.36 s of compute per s of music |
31 Jul 2026 |
| 10 | Magenta RealTime (large) Google |
918 ±20 |
1K | CC BY 4.0 credit Google |
— | faster than real time 0.9 s of compute per s of music |
31 Jul 2026 |
| 11 | Magenta RealTime 2 (base) Google |
883 ±36 |
443 | CC BY 4.0 credit Google |
2.4 B | slower than real time 1.13 s of compute per s of music |
31 Jul 2026 |
| 12 | Stable Audio Open 1.0 Stability AI |
882 ±22 |
1K | Stability AI Community License free under $1M a year in revenue |
1.21 B | faster than real time 0.34 s of compute per s of music |
4 Jun 2026 |
| 13 | MusicGen Medium Meta |
880 ±23 |
1K | CC BY-NC 4.0 non-commercial only |
— | slower than real time 2.21 s of compute per s of music |
31 Jul 2026 |
| 14 | MusicGen Small Meta |
853 ±27 |
726 | CC BY-NC 4.0 non-commercial only |
0.59 B | slower than real time 1.14 s of compute per s of music |
18 Feb 2026 |
| 15 | Stable Audio Open Small Stability AI |
800 ±29 |
703 | Stability AI Community License free under $1M a year in revenue |
0.5 B | faster than real time 0.03 s of compute per s of music |
13 Feb 2026 |
| 16 | ACE-Step v1 (3.5B) ACE-Step |
784 ±59 |
347 | Apache 2.0 | 3.5 B | faster than real time 0.29 s of compute per s of music |
15 Oct 2025 |
2 models with fewer than 200 votes
| # | Model | Score | Votes | Weights | Size | Speed (open models) | Last vote |
|---|---|---|---|---|---|---|---|
| 1 | Anonymous preview (jerboa) Undisclosed |
1077 ±61 |
127 | Anonymous | — | — | 23 Sept 2025 |
| 2 | Anonymous preview (ocelot) Undisclosed |
1075 ±59 |
134 | Anonymous | — | — | 25 Sept 2025 |
Listen to real examples
Tracks made in Music Arena’s blind tests from the same list of prompts, published under CC BY 4.0. They play from Hugging Face. No public tracks for Stable Audio 3 Medium, Stable Audio Open 1.0, Stable Audio Open Small: their makers did not allow the release.
Make music now, free
The official free demos on Hugging Face. They run on Hugging Face’s shared GPUs with a free daily quota for each visitor, so they can be busy; signing in to Hugging Face gives you more. Nothing runs on our servers.
Blind test: which track is better?
Two tracks for the same prompt, made by two different models. Listen, pick the better one, then see who made them and what the original Music Arena voter chose.
Open music models you can run yourself
- Stable Audio 3 Medium by Stability AI: Stability AI Community License, free for individuals and companies under $1M a year in revenue. 2.31 B parameters.
- ACE-Step 1.5 Turbo by ACE-Step: MIT, commercial use allowed. 1.7 B parameters. Runs with under 4 GB of VRAM (official card).
- Magenta RealTime (large) by Google: CC BY 4.0, credit Google.
- Magenta RealTime 2 (base) by Google: CC BY 4.0, credit Google. 2.4 B parameters.
- Stable Audio Open 1.0 by Stability AI: Stability AI Community License, free under $1M a year in revenue. 1.21 B parameters.
- MusicGen Medium by Meta: CC BY-NC 4.0, non-commercial use only.
- MusicGen Small by Meta: CC BY-NC 4.0, non-commercial use only. 0.59 B parameters.
- Stable Audio Open Small by Stability AI: Stability AI Community License, free under $1M a year in revenue. 0.5 B parameters.
- ACE-Step v1 (3.5B) by ACE-Step: Apache 2.0, commercial use allowed. 3.5 B parameters.
Licenses checked by hand on Hugging Face on 20 Oct 2026. Always read the license before using the music commercially. Speed: seconds of GPU time for each second of music on the arena's NVIDIA A6000; under 1 means faster than real time.
About the data
Votes come from Music Arena, a research project by Carnegie Mellon University, LMArena and Sony AI, published as an open dataset under CC BY 4.0: 8K battles from 2025-07-08 to 2026-07. We compute the scores ourselves with the Bradley-Terry method, as Arena does; “both bad” and “tie” count half a point to each model, so our order can differ slightly from Music Arena’s own page. We are not affiliated with Music Arena. Data updated 20 Oct 2026. Looking for voices instead? See free audio models.
Share this page:XLinkedInRedditWhatsApp
Spotted a wrong number or a broken link? Report a mistake