magenta-rt 2.0.2

Real-time music generation models.

Magenta RealTime 2

[!NOTE] Looking for Magenta RealTime v1? The original model and code have been moved to the v1_legacy branch.

Magenta RealTime 2 (MRT2) is a state-of-the-art open-weights model for real-time music generation. It contains several key components

• An open-weights model
• A Python library magenta-rt for inference with JAX and MLX backends
• A C++ inference engine magentart::core for efficient streaming audio generation on Apple Silicon MacBooks
• A suite of example applications built on the inference engine.

Use this project to run inference, build a DAW plugin, or embed the model into new applications of your imagination. Future updates will support supervised fine-tuning.

📖 Full documentation: https://magenta.github.io/magenta-realtime/ (or build it locally — see docs/README.md).

Repo Highlights

• magenta_rt/ — Python inference library (JAX / MLX backends).
• cpp/ — C++ inference library (magentart::core).
• examples/mrt2/auv3 — All-in-one AUv3 plugin for DAWs.
• examples/mrt2/standalone — All-in-one standalone macOS app.
• examples/jam/ — App for exploring note control.
• examples/collider/ — App for exploring prompt space.
• notebooks/ - Notebook for trying Python API.

Hardware requirements

Real-time streaming requires Apple Silicon (M-series). We offer two model sizes:

• mrt2_small (230M parameters) — runs real-time on any Apple Silicon Mac, including Air models.
• mrt2_base (2.4B parameters) — higher quality; requires a Pro Max chip for real-time streaming.

The table below shows which devices support real-time streaming (generating audio faster than playback):

Device	mrt2_small (230M)	mrt2_base (2.4B)
M5 Max	✅	✅
M3 Max	✅	✅
M2 Max	✅	✅
M4 Pro	✅	✅
M2 Pro	✅	❌
M1 Pro	✅	❌
M4 Air	✅	❌
M3 Air	✅	❌
M1 Air	✅	❌

Note: Both models can also run offline (non-real-time) inference on any Apple Silicon Mac or NVIDIA GPU via the Python library. See more details on docs/models.md.

Quickstart on Apple Silicon

# Install uv if you haven't and create a venv
curl -LsSf https://astral.sh/uv/install.sh | sh
uv venv --python 3.12
source .venv/bin/activate

# Install dependencies (Python dev)
uv pip install "magenta-rt[mlx]"

# Download resources: style model and codec model
# (i.e., MusicCoCa and SpectroStream)
mrt models init
# Download the streaming model you want to use
mrt models download
# Generate 4s music (change to `mrt2_small` for small model)
mrt mlx generate --prompt "disco funk" --duration 4.0 --model=mrt2_base

Python Development

For local development, clone the repo instead of installing from PyPI:

git clone --recurse-submodules https://github.com/magenta/magenta-realtime.git
cd magenta-realtime
uv pip install -e ".[mlx]"

C++ App Development

To get started building C++ apps, perform the following setup:

# Install dependencies
uv pip install "cmake<3.28"

# Build hello_mrt2 (a basic command-line interface)
cmake . -B build
cmake --build build --target hello_mrt2 -j10

# Generate 4s music
./build/examples/hello_mrt2/hello_mrt2 \
    ~/Documents/Magenta/magenta-rt-v2/models/mrt2_base/mrt2_base.mlxfn \
    ~/Documents/Magenta/magenta-rt-v2/resources \
    100 \
    --prompt "ambient pads with sub bass"

See the full documentation:

• Installation
• Models & checkpoints
• Inference
• Exporting models
• Latency benchmark
• Testing

Other resources

• Get Started
• Blog Post
• Hugging Face

Citing this work

Please cite our previous technical report:

BibTeX:

@article{gdmlyria2025live,
    title={Live Music Models},
    author={Caillon, Antoine and McWilliams, Brian and Tarakajian, Cassie and Simon, Ian and Manco, Ilaria and Engel, Jesse and Constant, Noah and Li, Pen and Denk, Timo I. and Lalama, Alberto and Agostinelli, Andrea and Huang, Anna and Manilow, Ethan and Brower, George and Erdogan, Hakan and Lei, Heidi and Rolnick, Itai and Grishchenko, Ivan and Orsini, Manu and Kastelic, Matej and Zuluaga, Mauricio and Verzetti, Mauro and Dooley, Michael and Skopek, Ondrej and Ferrer, Rafael and Borsos, Zal{\'a}n and van den Oord, {\"A}aron and Eck, Douglas and Collins, Eli and Baldridge, Jason and Hume, Tom and Donahue, Chris and Han, Kehang and Roberts, Adam},
    journal={arXiv:2508.04651},
    year={2025}
}