Google's Magenta Releases Real-Time AI Music Generation Model With Up To 2.4B Parameters
Summary
Google's Magenta team releases Magenta RealTime 2, a cutting-edge open-weights AI music generation model available in 230M and 2.4B parameter sizes, enabling real-time music streaming on Apple Silicon Macs and offline inference on NVIDIA GPUs.
Key Points
- Google's Magenta team releases Magenta RealTime 2 (MRT2), a state-of-the-art open-weights model for real-time music generation, featuring two model sizes: a 230M parameter small model and a 2.4B parameter base model.
- Real-time streaming is exclusive to Apple Silicon Macs, with the smaller model running on any M-series chip and the larger model requiring a Pro Max chip, while both models support offline inference on Apple Silicon or NVIDIA GPUs via a Python library.
- The project includes a Python inference library with JAX and MLX backends, a C++ inference engine optimized for Apple Silicon, and a suite of example applications including DAW plugins and standalone macOS apps, with supervised fine-tuning support planned for future updates.