Magenta RealTime
Google’s open-weights model for real-time music generation, streaming live on Apple Silicon.
| Language | Python / C++ |
|---|---|
| Category | ML & generative |
| License | Apache-2.0 (code) · CC-BY-4.0 (weights) |
| Platforms | macOS Linux Windows |
| Install | uv pip install "magenta-rt[mlx]" |
| First released | 2025 (MRT2 in 2026) |
| Maintained | Yes |
| Links | Documentation · Source · Model |
Strengths
- Genuinely real-time — generates audio faster than playback, so you can steer it while it plays
- CC-BY-4.0 weights: commercial use permitted with attribution, the most permissive licence of any model in this category
- Apple Silicon is the primary target rather than an afterthought — mrt2_small streams on an M1 Air
- Two sizes: 230M for reach, 2.4B for quality
- Ships a C++ path as well as Python, for embedding in apps
Limitations
- Real-time streaming requires Apple Silicon — NVIDIA GPUs can only run offline inference
- The 2.4B model needs a Pro or Max chip to stream in real time
- Short-form generation; not a full-song model
- MRT2 replaced v1, which moved to the v1_legacy branch
Best for
Live performance and interactive apps where you steer generation as it plays — and the clear first choice if you are on a Mac or need commercially usable weights.
Minimal example
# Install and fetch the models
uv venv --python 3.12 && source .venv/bin/activate
uv pip install "magenta-rt[mlx]"
mrt models init # style (MusicCoCa) + codec (SpectroStream)
mrt models download # the streaming model itself
# Generate four seconds
mrt mlx generate --prompt "disco funk" --duration 4.0 --model=mrt2_baseReal-time streaming support. mrt2_small (230M) streams on every Apple Silicon Mac including Air models. mrt2_base (2.4B) streams on Pro and Max chips — M4 Pro, M2/M3/M5 Max — but not on Air or M1/M2 Pro. Both run offline on any Apple Silicon Mac or NVIDIA GPU.
Related tools
See all Magenta RealTime alternatives & how they compare →
See the whole ML & generative category →