Magenta RealTime 2 transforms the MacBook into a generative instrument

Google DeepMind's Magenta RealTime 2 open-weights model runs locally on Apple Silicon Macs to generate continuous, low-latency music from MIDI and text.

Google DeepMind's Magenta team conceives of its new model as a full-fledged instrument. Magenta RealTime 2 produces a continuous real-time music stream, with a latency of approximately 200 ms, and runs natively locally on Apple Silicon Macs. Google DeepMind presents it as the only open-weights model capable of continuous music generation driven at such low latency. The code is published open source under Apache 2.0, the weights as open weights under CC BY 4.0, and the whole succeeds Magenta RealTime, an open cousin of Lyria RealTime.

The control combines three inputs: a MIDI keyboard, with note-by-note control or Auto-Strum mode, text style prompts, and audio excerpts serving as reference, with multiple prompts able to be mixed and weighted on the fly. The output is 48 kHz stereo. Two sizes coexist, Base with its 2.4 billion parameters, which requires an M3 Pro chip or higher for true real-time, and Small, 230 million parameters, which runs on any Apple Silicon Mac.

A ready-to-use bundle includes three applications: Jam for improvising, Collider for hybridizing genres on a 2D interface, and an AU plugin to load into DAWs like Logic or Ableton. The weights are downloaded on first launch, and everything runs locally, without passing through the cloud. The code is on GitHub, the model on Hugging Face.