Google DeepMind Releases Magenta RealTime 2, Real-Time Music Model for Mac with Sub-200ms Latency

According to Beating, Google DeepMind released Magenta RealTime 2 (MRT2), an open-source real-time music generation model optimized for Apple Silicon MacBooks. The model enables musicians to control audio generation through MIDI signals, text prompts, or audio samples with latency below 200 milliseconds.

MRT2 supports three interaction modes: MIDI-guided accompaniment that automatically generates orchestral backing based on chord input, text-to-instrument synthesis for style-based generation, and audio cloning that mimics input sound characteristics. Two parameter versions are available—a 230-million-parameter mrt2_small for all Apple Silicon Macs, and a 2.4-billion-parameter mrt2_base for higher-end M-series Pro and Max chips. The inference library magenta-rt and C++-based inference engine are now open-sourced on GitHub, with standalone macOS applications and DAW plugins available.