How to choose an open-source music generation model

Compare open music and sound models by what they make, hardware and license: MiniMax Music 3, YuE2, ACE-Step 1.5, HeartMuLa, Stable Audio 3, Magenta RealTime 2 and MOSS-SoundEffect.

  • Beginner
  • 6 min read
  • Published
  • Updated

Short answer. For full songs with vocals under a permissive license, start with ACE-Step 1.5 (MIT, from about 4 GB of VRAM) or HeartMuLa (Apache 2.0). MiniMax Music 3 writes songs of up to five minutes and allows commercial products that show its name, and YuE2 scores best on its own team’s benchmark but is non-commercial. For instrumental music and sound effects, Stable Audio 3 runs on an 8 GB GPU or, in its Small versions, on a laptop CPU. Magenta RealTime 2 is a live instrument for Apple Silicon MacBooks, and MOSS-SoundEffect v2.0 makes sound effects under Apache 2.0. Quality is a matter of taste, so listen before you decide.

What to look for

An open music model does one of four jobs, and the job matters more than any score:

  1. Songs with vocals. Lyrics and a style description in, a finished song out. MiniMax Music 3, YuE2, ACE-Step 1.5 and HeartMuLa do this.
  2. Instrumental music. A text prompt in, music without singing out, often with inpainting or continuation. Stable Audio 3 does this.
  3. Live music. Audio is streamed continuously and reacts to text, audio or MIDI while it plays. Magenta RealTime 2 does this.
  4. Sound effects. A sentence in, a short clip out. MOSS-SoundEffect v2.0 and the sound-effect models of Stable Audio 3 do this.

Then compare the hardware, the license and what the developers say about their training data. Our music and sound generation list covers every model we track, and the glossary entry music generation explains the vocabulary.

The shortlist

Model Size License What it makes Hardware
MiniMax Music 3 11.1B MiniMax-Music3 Community License Songs with vocals, up to 5 minutes, 32 kHz A 24 GB GPU in BF16, or less with offloading
YuE2 3.6B CC BY-NC 4.0 Songs from an editable score, 48 kHz A 24 GB GPU, 11 to 14 GiB peak
ACE-Step 1.5 2B or 4B DiT plus a 0.6B to 4B planner MIT Songs of 10 seconds to 10 minutes, lyrics in 50+ languages About 4 GB (2B turbo), 12 to 20 GB (XL)
HeartMuLa 3.9B Apache 2.0 Songs of up to 4 minutes, 48 kHz About 22 GB of files, no VRAM figure
Stable Audio 3 1.4B Medium, 0.46B Small Stability AI Community License Instrumental music and sound effects, up to 6 min 20 s Medium about 8 GB, Small on a CPU
Magenta RealTime 2 2.4B Base, 230M Small Weights CC BY 4.0, code Apache 2.0 Live instrumental music, 48 kHz An Apple Silicon MacBook
MOSS-SoundEffect v2.0 1.3B Apache 2.0 Sound effects of up to 30 seconds, 48 kHz No VRAM figure, 11 GB of files

Pick by the job

  • A song draft on a gaming GPU. ACE-Step 1.5. Its developers say a full song takes under 10 seconds on an RTX 3090 and the 2B turbo model runs in under 4 GB of VRAM. It also does covers, repainting part of a track, stem separation and LoRA training on a few songs. Those speeds are the developers’ own figures.
  • Songs you may sell. ACE-Step 1.5 (MIT) and HeartMuLa (Apache 2.0) have the plainest terms. HeartMuLa writes songs of up to four minutes in Chinese, English, Japanese, Korean or Spanish at about real-time speed, but its reference-audio conditioning is still listed as to come.
  • Long, structured songs from detailed prompts. MiniMax Music 3 takes lyrics with [Verse] and [Chorus] tags plus a description of genre, tempo, key, vocals and arrangement, and writes up to five minutes. It is the biggest of the song models at 11.1B parameters and needs CUDA.
  • A song you can edit before it is rendered. YuE2 writes the melody and chords as a readable score first, so you can change the notes or bring your own score, and it can make covers from a transcription. It is non-commercial.
  • Background music, loops and drafts for video or games. Stable Audio 3 makes instrumental music of up to 6 minutes 20 seconds in about eight sampling steps, and it can inpaint, continue and edit audio. It cannot sing.
  • Sound effects. MOSS-SoundEffect v2.0 takes English or Chinese prompts under Apache 2.0 but uses 100 sampling steps, while the sound-effect models of Stable Audio 3 need eight and come with a revenue cap.
  • A live, playable instrument. Magenta RealTime 2 reacts to text, audio and MIDI in about 200 ms, according to Google, and it is built for Apple Silicon Macs.
  • A laptop with no discrete GPU. The Small models of Stable Audio 3 run on a CPU, and Stability measured about 6 seconds for two minutes of audio on a MacBook Pro M4. Magenta RealTime 2 Small runs on any Apple Silicon MacBook.

Read the license before you fall for the demo

Four patterns to know:

  • Plain permissive. ACE-Step 1.5 is MIT for code and weights, and HeartMuLa and MOSS-SoundEffect v2.0 are Apache 2.0. HeartMuLa’s separate cover-song model, MuLaCover, is CC BY-NC.
  • Credit and a revenue line. The MiniMax-Music3 Community License asks products to show “MiniMax-Music3”, needs written authorization above US$20 million a year and tells hosts to keep safeguards against infringing outputs. The Stability AI Community License behind Stable Audio 3 is free below US$1 million a year after registration, asks for a “Powered by Stability AI” credit and bars using the models or their outputs to build another foundation model. The glossary entry revenue cap explains the pattern.
  • Credit only. The weights of Magenta RealTime 2 are CC BY 4.0, and Google’s terms add responsible-use rules and forbid output that infringes others’ rights.
  • Non-commercial. YuE2 uses CC BY-NC, so selling its songs or running a paid service needs permission from the authors. See non-commercial license.

AI licenses explained covers the rules in general.

A license on the weights says nothing about the music a model learned from, and the developers describe it to very different degrees:

  • ACE-Step’s developers say they used licensed music, royalty-free and public-domain recordings and synthetic audio made from MIDI. We have not verified that.
  • The Stable Audio 3 report lists about 806,000 licensed AudioSparx recordings and about 470,000 Creative Commons recordings from Freesound.
  • Magenta RealTime 2 was trained on about 71,000 hours of stock music, mostly instrumental.
  • The cards of MiniMax Music 3, YuE2 and HeartMuLa do not describe their training data.

Ask a lawyer before you sell generated music, keep the prompts and settings you used, and avoid imitating a living artist’s voice or style.

Test it like a listener

Pick five prompts from your real use: a genre and mood, a tempo, a lyric in each language you need and, for sound effects, the sounds you actually want. Generate each with two or three candidates and listen blind. Time one full generation on your own hardware, because published speeds are best cases. Song models can drift from the tempo, key or lyrics you asked for, so listen for lyric clarity and structure as well as style. How to evaluate AI models yourself explains how to turn this into a repeatable test.

Frequently asked questions

Which open music model sounds the best?
We cannot give an independent ranking. The only table that compares several open song models is the YuE2 team's own WildSongBench, where YuE2 scores 6.73, MiniMax Music 3 6.28, HeartMuLa 6.25 and ACE-Step 1.5 6.01, against 6.87 for the hosted Suno v5. Those scores come from one developer. Generate the same five prompts with two or three models and listen blind.
Which open music models can I use in a commercial product?
ACE-Step 1.5 is MIT, and HeartMuLa and MOSS-SoundEffect v2.0 are Apache 2.0. MiniMax Music 3 allows commercial products that show the name MiniMax-Music3, and companies above US$20 million a year need written authorization. Stable Audio 3 is free for commercial use below US$1 million in annual revenue, after registration and with a Powered by Stability AI credit. The weights of Magenta RealTime 2 are CC BY 4.0, so credit Google and follow its terms. YuE2 is CC BY-NC 4.0 and so non-commercial only, and so is HeartMuLa's separate MuLaCover model.
How much GPU memory does a music model need?
From almost none to 24 GB. The Small models of Stable Audio 3 run on a laptop CPU, and the Small model of Magenta RealTime 2 on any Apple Silicon MacBook. ACE-Step's developers say its 2B turbo model runs in under 4 GB of VRAM, and Stable Audio 3 Medium needs about 8 GB. YuE2's card asks for a 24 GB GPU with BF16 support, and MiniMax Music 3 fits a 24 GB GPU in BF16, or an 8 GB card with layer-by-layer offloading, only more slowly.
Can I make sound effects with an open model?
Yes. MOSS-SoundEffect v2.0 (Apache 2.0) makes clips of up to 30 seconds from English or Chinese prompts. Stable Audio 3 has a Small SFX model for clips of up to 2 minutes and a Medium model that makes both music and sound effects for up to 6 minutes 20 seconds, under a license that is free below US$1 million in annual revenue.
Can an open model sing in languages other than English?
ACE-Step 1.5 says its lyrics work in more than 50 languages, and HeartMuLa's card lists Chinese, English, Japanese, Korean and Spanish. We have not checked those lists, so test your own language before you commit.
Is there an open model that works like Suno?
Four open song models take lyrics and a style description and return a finished song: MiniMax Music 3, YuE2, ACE-Step 1.5 and HeartMuLa. ACE-Step's developers rate their model between Suno v4.5 and v5, and the YuE2 team says YuE2 is competitive with Suno v5 and v6. Both are the developers' own claims, and Suno itself is a hosted service.

Key terms in this guide

  • Diffusion model A diffusion model generates images or video by starting from random noise and removing the noise step by step until a picture emerges.
  • LoRA LoRA, short for low-rank adaptation, fine-tunes a model by training a small set of added weights instead of changing all of its parameters.
  • Music generation Music generation models turn a text description, and often lyrics, into audio: full songs with vocals, instrumental tracks, live streams of music or sound effects.
  • Non-commercial license A non-commercial license lets you use a model for research, study and personal projects, but not to make money from it.
  • Revenue cap (revenue threshold) A revenue cap is a license condition that makes a model free below a certain income and requires a separate agreement above it.
  • Sampling steps Sampling steps are the number of passes a diffusion model makes to turn noise into an image or video. Fewer steps mean faster results, usually with some loss of detail or variety.

Models mentioned in this guide

Model Released Parameters Context License Smallest variant needs
MiniMax Music 3 MiniMax 11.1B – MiniMax Music3 Community (Conditional) 24 GB Gaming GPU
YuE2 M-A-P 3.63B – CC BY-NC (Non-commercial) 24 GB Gaming GPU
ACE-Step 1.5 ACE-Step 2B – 4B – MIT (Permissive) 4 GB Laptop or phone
HeartMuLa HeartMuLa 3.94B – Apache 2.0 (Permissive) 26 GB Gaming GPU
Stable Audio 3 Stability AI 460M – 1.4B – Stability Community (Conditional) 4.02 GB Laptop or phone
Magenta RealTime 2 Google DeepMind 230M – 2.4B – CC BY 4.0 (Permissive) 2.53 GB Laptop or phone
MOSS-SoundEffect v2.0 OpenMOSS 1.3B – Apache 2.0 (Permissive) 13 GB Laptop or phone