Magenta RealTime

Google DeepMind's live music model generates a continuous stereo stream that can be steered while it plays, using text prompts, audio examples, or weighted blends of both, so a performer can morph between styles mid-jam. The current Magenta RealTime 2 release ships two open-weights checkpoints, a 230M-parameter Small and a 2.4B-parameter Base, with inference libraries in JAX, MLX, and C++ plus macOS example apps including an AUv3 plugin and the Jam and Collider tools. Real-time streaming runs on Apple Silicon, where Small works on any M-series Mac including Airs while Base needs a recent Pro or Max chip; offline generation also runs on any Apple Silicon Mac or on NVIDIA GPUs. Install is a uv pip install of the magenta-rt package followed by an mrt CLI command that pulls weights from Hugging Face. Code is Apache-2.0 and the weights carry CC-BY-4.0, so commercial use is allowed with attribution, and training used licensed stock music rather than scraped catalogs. It stands as the reference open option for latency-sensitive, interactive music generation.