Google DeepMind releases Magenta RealTime 2, an open-source, low-latency music generation model designed to be played as a live instrument.
Google DeepMind has announced the release of Magenta RealTime 2 (MRT2), an open-source model that enables low-latency, live music synthesis. Designed to be played as a musical instrument, MRT2 runs natively on macOS and allows musicians to guide audio generation in real-time using text prompts, MIDI, or audio inputs. Built on three key components—SpectroStream, MusicCoCa, and a Large Language Model—the model utilizes a block autoregression architecture to generate continuous music faster than real-time, focusing on interactive, human-in-the-loop music creation.
While most generative audio models focus on producing static tracks, Magenta RealTime 2 is a significant step forward in making AI a collaborative partner for live musical performance.
* True Real-Time Jamming: The low-latency capability allows MRT2 to be played interactively like a physical instrument, moving away from "prompt-and-wait" audio generation.
* On-Device macOS Execution: Native support for MacBooks ensures privacy and reduces latency by eliminating network round trips, though it restricts compatibility for Windows and Linux users.
* Multimodal Control: By accepting text, audio, and MIDI simultaneously, the model gives musicians diverse, expressive ways to shape the synthesized performance.
* Open-Source & Open-Weights: Under Apache 2.0 and Creative Commons licenses, Google continues to enrich the developer ecosystem for generative music research.
DISCOVERED
3h ago
2026-06-04
PUBLISHED
3h ago
2026-06-04
RELEVANCE
AUTHOR
googlegemma