New on-device audio SDK lands
A modular MLX-Audio-Swift SDK for on-device audio intelligence on iOS, macOS and visionOS was announced on social channels, promising real-time capabilities and modular architecture for native apps announced. The SDK looks positioned to accelerate local audio features without sending raw audio to servers.
The Swift reference repo is published as mlx-audio-swift on GitHub, showing roughly 145 stars and an explicit modular package layout. (github.com) The broader MLX-Audio project (Python + tools) reports a GitHub repo with ~6.2k stars and forks, and a PyPI release version 0.4.1 published on Mar 14, 2026. (github.com) The SDK surface is split into focused modules — MLXAudioCore, MLXAudioTTS, MLXAudioSTT, MLXAudioCodecs and MLXAudioUI — with example apps and SwiftUI components listed in the repository README. (github.com) Model and runtime capabilities in the project README call out multiple TTS/STT model families (Kokoro, Soprano, Whisper/GLMASR) and explicit quantization targets (3-, 4-, 6-, 8-bit) to reduce footprint for on-device inference. (github.com) Performance targets emphasize Apple Silicon acceleration and MLX/Metal support for Neural Engine/GPU execution, per MLX framework docs and project claims about Apple Silicon optimization. (mlx-framework.org) Model weights and preconverted assets are surfaced via the mlx-community org on Hugging Face for easy integration, and several community forks and example apps (e.g., a mac/iOS app repo) demonstrate immediate local integration paths. (huggingface.co)
