Open music models grew up in 2026: seven models for songs, instruments and sound effects
Seven open music and sound models arrived between January and September 2026, from song writers such as ACE-Step 1.5 to a live instrument from Google. Compare jobs, hardware and licenses.
This site did not cover music until now, and this year gave it plenty to cover. Between January and September 2026, seven open music and sound models arrived from seven different developers. They do four different jobs, so we added a Music & sound category and a list of the best open music and sound models.
Seven models, four jobs
| Released | Model | Developer | License of the weights | Job |
|---|---|---|---|---|
| 14 Jan 2026 | HeartMuLa | HeartMuLa team | Apache 2.0 | Songs with vocals |
| 23 Jan 2026 | ACE-Step 1.5 | ACE Studio and StepFun | MIT | Songs with vocals |
| 20 May 2026 | Stable Audio 3 | Stability AI | Stability AI Community License | Instrumental music and sound effects |
| 25 May 2026 | MOSS-SoundEffect v2.0 | OpenMOSS | Apache 2.0 | Sound effects |
| 4 Jun 2026 | Magenta RealTime 2 | Google DeepMind | CC BY 4.0 | Live, interactive music |
| 13 Aug 2026 | MiniMax Music 3 | MiniMax | MiniMax-Music3 Community License | Songs with vocals |
| 9 Sep 2026 | YuE2 | M-A-P | CC BY-NC 4.0 | Songs from an editable score |
The jobs matter more than the dates. A song model takes lyrics and a style description and sings them. Stable Audio 3 makes instrumental music and sound effects from a text prompt and can extend or edit existing audio. Magenta RealTime 2 plays continuously and reacts to text, audio or MIDI in about 200 ms, according to Google. MOSS-SoundEffect turns a sentence into a clip of up to 30 seconds.
The hardware range is wide
The four song models need very different machines. ACE-Step 1.5’s developers say its 2B turbo model runs in under 4 GB of VRAM and writes a full song in under 10 seconds on an RTX 3090. YuE2’s card asks for a 24 GB GPU and measures 11 to 14 GiB of peak memory, and MiniMax Music 3 fits a 24 GB GPU in BF16 or, with layer-by-layer offloading and a lot more time, an 8 GB card. HeartMuLa’s README gives no VRAM figure, and its files come to about 22 GB in 32-bit precision. Outside songs, Stable Audio 3 Medium needs about 8 GB, its Small models run on a CPU, and Magenta RealTime 2 Small runs on any Apple Silicon MacBook.
Licenses differ more than sizes
Three of the seven are plainly permissive: ACE-Step 1.5 is MIT, and HeartMuLa and MOSS-SoundEffect v2.0 are Apache 2.0, although HeartMuLa’s separate cover-song model is CC BY-NC. The weights of Magenta RealTime 2 are CC BY 4.0 with Google’s use terms on top. Two licenses draw a revenue line: the MiniMax-Music3 Community License asks products to show the model name and needs written authorization above US$20 million a year, and the Stability AI Community License is free below US$1 million a year, after registration and with a credit. YuE2 uses CC BY-NC, so it is for non-commercial use only. Free until you grow: where open licenses draw the line looks at how such thresholds work across open licenses.
Quality claims come from the developers
We cannot rank these models by sound. 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, with the hosted Suno v5 at 6.87. ACE-Step’s developers rate their model between Suno v4.5 and v5, and the HeartMuLa team says an unreleased 7B model matches Suno. Stability says Stable Audio 3 sets a new state of the art among open-weight systems, and Google says Magenta RealTime 2 was the only open-weights model for real-time generation with low-latency control at its release. Every one of those is the developer’s own claim, and MiniMax publishes no benchmark numbers for Music 3.
Training data is described to very different degrees too. ACE-Step’s developers say they used licensed, royalty-free and synthetic data, and the Stable Audio 3 report names licensed AudioSparx recordings and Creative Commons recordings from Freesound. Google says Magenta RealTime 2 learned from about 71,000 hours of stock music, while the MiniMax Music 3, YuE2 and HeartMuLa cards do not describe theirs. Ask a lawyer before you sell generated music.
Which to try first
- A song model with plain terms: ACE-Step 1.5 or HeartMuLa. ACE-Step is the one that runs on a small GPU.
- A larger song model with detailed prompts: MiniMax Music 3, if you can show its name and have a 24 GB GPU.
- Songs you can edit as a score: YuE2, for non-commercial work.
- Instrumentals and sound effects: Stable Audio 3, or MOSS-SoundEffect v2.0 if you need Apache 2.0.
- A live instrument: Magenta RealTime 2, on an Apple Silicon MacBook.
How to choose an open-source music generation model walks through the decision, ACE-Step 1.5 vs MiniMax Music 3 and Stable Audio 3 vs MOSS-SoundEffect v2.0 compare the closest pairs, and the glossary entry music generation explains the terms.
Sources
- HeartMuLa-oss-3B-happy-new-year model card (Hugging Face) huggingface.co
- ACE-Step 1.5 model card (Hugging Face) huggingface.co
- Meet Stable Audio 3.0, the model family built for artistic experimentation (Stability AI, 20 May 2026) stability.ai
- MOSS-SoundEffect-V2.0 model card (Hugging Face) huggingface.co
- Magenta RealTime 2: Open & Local Live Music Models (Magenta team, 4 Jun 2026) magenta.withgoogle.com
- MiniMax Music 3.0: Next-Generation Open-Weights, Production-Ready & Versatile Music Model (MiniMax, 13 Aug 2026) minimax.io
- YuE2-3B model card, WildSongBench comparison table (Hugging Face) huggingface.co
Models in this story
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| HeartMuLa HeartMuLa | 3.94B | – | Apache 2.0 (Permissive) | 26 GB Gaming GPU | |
| ACE-Step 1.5 ACE-Step | 2B – 4B | – | MIT (Permissive) | 4 GB Laptop or phone | |
| Stable Audio 3 Stability AI | 460M – 1.4B | – | Stability Community (Conditional) | 4.02 GB Laptop or phone | |
| MOSS-SoundEffect v2.0 OpenMOSS | 1.3B | – | Apache 2.0 (Permissive) | 13 GB Laptop or phone | |
| Magenta RealTime 2 Google DeepMind | 230M – 2.4B | – | CC BY 4.0 (Permissive) | 2.53 GB Laptop or phone | |
| 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 |