iOS 27 / Core AI — any plans, or is MLX-at-runtime the long-term direction?

[tushgup]

Thanks again for this package!

With iOS 27, Apple's guidance is that MLX is a build-time tool and Core AI is the runtime for shipping apps. We're evaluating exporting Kokoro's upstream PyTorch weights via coreai-torch instead, mainly because MLX on iOS has cost us:

• background GPU restrictions — we gate generation on resignActive to avoid crashes
• MLX aborting on the simulator, so voices can't be tested without a device
• version sensitivity — we're pinned to mlx-swift 0.29.1 for audio quality, and mlx-swift 0.30.2 pin fails to build with Xcode 26 / iOS 26 SDK #30 suggests newer pins have their own SDK problems

ANE would address the first two structurally. Is Core AI export something you're considering, or would a PR exploring it be welcome? Happy to share what we learn either way.