On June 4, 2026, Google DeepMind's Magenta team officially open-sourced Magenta RealTime 2 (MRT2). It is fundamentally different from the "prompt → song" tools of the past: the model has only 2.4B parameters, yet can continuously output 48kHz stereo audio with end-to-end control latency of about 200ms, while simultaneously accepting text, audio examples, and MIDI signals as "navigation" — a creator holds a chord, and the model fills in the entire band live in the background, which is exactly what all previous text-to-music models could not do.
Three engineering deliverables
- Fully open-sourced: weights CC BY 4.0, code Apache 2.0
- Native Apple Silicon: with the MLX backend and C++ inference engine, runs real-time on MacBook; the mrt2_small (230M parameters) runs real-time from M1 Air, mrt2_base needs M2 Pro and above
- Plug-and-play: the official AUv3 plug-in and independent apps like Jam and Collider let the model be embedded directly into mainstream DAW workflows, putting "an instrument that can play" into musicians' hands for the first time
The dual-track strategy is worth pondering
Google simultaneously maintains "open-source weights + closed-source Lyria RealTime API": local weights serve musicians and researchers, cloud API serves consumer products like MusicFX DJ Mode. This "research open, business closed" combination essentially turns the model itself into the bedrock of the community ecosystem, and the API into the monetization gateway.
Less than 24 hours after Hugging Face and the GitHub repo went up, the community had already built a gesture-controlled MIDI improvisation demo — DeepMind researcher Jesse Engel used his MacBook camera to gesture-direct MRT2 to play on a plane offline, and a peer called it "Oh, I'm living in the future."
What's really scarce today isn't "can the model write a song," but "can the model play with you." The "tool era" of music generation is indeed ending.