Core AI model zoo

Music / audio generation (Stable Audio) on Core AI — port notes

Text→music/audio is latent diffusion, same shape as LTX-Video / VoxCPM: a text encoder conditions a DiT that denoises a latent over a few steps, then a VAE decodes the latent to a waveform. Ported here as Stable Audio Open Small (stabilityai/stable-audio-open-small, 341M, Stability + Arm): ~11s of 44.1 kHz stereo from a prompt, fully on device.

Shape of the port (3 bundles + a host sampler)

Lessons (the ones that saved days)

EDGE check BEFORE porting a music/audio-gen model (hard gate)

Music-gen moves fast on Apple Silicon — check whether MLX / CoreML already ship it on iPhone, or it’s a worse-MLX port (the DiffusionGemma / Qwen3-Coder-Next trap). Concretely: