Google launches Magenta RealTime 2, a low-latency music model for live jamming

Google's Magenta team has launched the Magenta RealTime 2 (MRT2) model, a local real-time music tool designed for live jamming with users. It comes in two sizes: mrt2_base with 2.4 billion parameters for high quality, and mrt2_small with 230 million parameters for speed. Both are optimized for Apple Silicon, with mrt2_small running in real time on M1 and later Macs.

The big upgrade is low latency. The previous version needed TPUs or powerful GPUs and had about 3 seconds of delay, processing audio in 2-second chunks. MRT2 generates frame by frame, each lasting 40 milliseconds, cutting latency to around 200 milliseconds—15 times faster. It supports text and audio style prompts, plus MIDI notes and drum control, and can follow note changes continuously. An Auto-Strum mode handles plucking timing automatically.

Google also released a free app called Jam, where users describe a music style and the model generates audio, with pitch adjustable via keyboard or MIDI. A plugin lets creators use MRT2 inside their favorite DAW. The model is based on the SpectroStream audio codec, processing 48kHz stereo sound at 25 frames per second.