Jam

✨ Model Info & GPU Grant

🎹 Google Magenta RealTime 2 Demo

This Space runs an interactive live music synthesis & steering interface for google/magenta-realtime-2 by Google DeepMind. Drag the prompt puck to blend musical styles in real time and play melodies with your computer keyboard, mouse, or any physical MIDI controller!

⚡ Architecture Components

SpectroStream RVQ Codec: High-fidelity discrete audio codec (48kHz stereo, 25Hz frame rate, 16kbps).
MusicCoCa: Multimodal joint embedding projecting prompts and audio to 768-dim space (12 RVQ depth).
Decoder LLM (2.4B/230M): Autoregressive transformer generating music conditioned on style + MIDI notes.

🚀 ZeroGPU & Dedicated Hardware Grant

This Space repository includes the complete pure-PyTorch ZeroGPU streaming implementation (app.py and magenta_rt/). To upgrade this Space with free ZeroGPU hardware:

Open the Community tab on this Space and click New Discussion.
Set Title: Apply for a GPU community grant: Personal project
Once approved by Hugging Face, change sdk: static to sdk: gradio in the Space's README.md.

Model Card ↗ GitHub Repo ↗