Running 24B parameters locally, the MRT2_base specification indicates that the inference potential of Apple chips is seriously underestimated, entering the real-time era of generative music.

Google DeepMind releases open-source real-time music generation model Magenta RealTime 2 (MRT2), optimized for M series Macs, with low latency locally, under 200 milliseconds. It can perform real-time playing and control via MIDI, text prompts, or audio clips, supporting three modes: MIDI guidance, text-to-synthesis, and audio cloning. Two versions are available: MRT2_small with 230 million parameters and MRT2_base with 2.4 billion parameters. The inference library Magenta-RT and C++ engine Magentart::core are open-sourced on GitHub, with standalone macOS applications and DAW plugins provided.