Android port: mrt2_small running in real time on a phone GPU (OpenCL, Adreno 840)
It's mrt2_small (230M) ported to OpenCL, running fully on-device on a Galaxy S26 Ultra (Snapdragon 8 Elite Gen 5, Adreno 840).
From a 10 min 37 s live session blending 10 prompts:
• Each 40 ms frame generated in ~31 ms (about 1.28x faster than real time)
• Zero dropouts, and no slowdown as the phone warmed up
• ~1.5 s control latency (I keep a bigger playback buffer than your ~200 ms setup, as a safety margin in case the phone throttles)
• 48 kHz stereo, fp16 weights
Steering is text prompts and prompt blends for now. The engine accepts MIDI note conditioning but I haven't tested it with a controller yet, and the MusicCoCa audio tower isn't ported. In the video I'm blending cinematic and synthwave prompts on a 2-D surface modeled on your Collider app.
Code: https://github.com/a8nova/adreno-llms/tree/main/src/models/magenta-realtime-2
Demo: https://www.reddit.com/r/OpenCL/comments/1wulfx7/ported_googles_magenta_realtime_2_to_android_with/