Magenta RealTime 2: Open & Local Live Music Models
Summary
Magenta RealTime 2 introduces a live, low-latency on-device music model with 40 ms frame-level inference and MIDI/text/audio conditioning. The project includes an open-weights model (2.4B), an open-source Python library, a C++ inference engine using MLX, and a suite of example apps for musicians and developers. It emphasizes real-time control, frame-by-frame conditioning, and cross-modal guidance to enable flexible live performances and DAW integrations.