[P] On-device Qwen3-TTS (1.7B/0.6B) inference on iOS and macOS via MLX-Swift — voice cloning, voice design, and streaming TTS with no cloud

About this article

Hey r/MachineLearning. I'm a solo dev working on on-device TTS using MLX-Swift with Qwen3-TTS. 1.7B model on macOS, 0.6B on iOS, quantized to 5-bit to fit within mobile memory constraints. No cloud, everything runs locally. The app is called Speaklone. Short demo video: https://www.youtube.com/watch?v=05gne9oPaaY The most interesting technical challenge has been MLX's lazy evaluation on memory-constrained devices. Computation graphs silently accumulate memory through strong references between...