Google’s Magenta team released Magenta RealTime 2, an open‑weights music model designed for live, low‑latency synthesis on a laptop. The 2.4B parameter engine runs on Apple Silicon, delivering sub‑200 ms response to MIDI, text or audio cues. Users can download a standalone Mac app or embed the model in a DAW via a C++ inference engine.

Unlike earlier generative music tools that render a full track from a prompt, MRT2 streams audio frame‑by‑frame, using a causal sliding‑window attention mechanism to keep memory bounded. Frame‑aligned conditioning injects MIDI and style prompts every 40 ms, yielding ~15× lower latency than the original RealTime. The release includes a Python library (magenta‑rt), example apps, and a suite of playable instruments.

All components run locally, eliminating cloud latency and privacy concerns. The C++ engine leverages MLX to compile the model into an .mlxfn container, enabling efficient GPU execution on M‑series Macs. By open‑sourcing weights and tooling, Magenta invites developers to craft custom instruments, plugins, and real‑time jam sessions today.

The team also announced upcoming finetuning support, allowing musicians to train MRT2 on personal datasets, and previewed collaborations with artists like Manaswi Mishra. A challenge at the Boston Music Technology Hackathon will showcase novel uses of the model. Magenta RealTime 2 thus transforms AI from offline composer to responsive instrument.