Wiring MLX to Swift: Running Fine-Tuned Models on Apple Silicon with Zero CoreML Overhead

Direct MLX Swift bindings bypass CoreML's model compilation and ANE scheduling latency — covers MLX array operations, model loading from Hugging Face Hub via swift-transformers, KV-cache management in Swift 6 actors, and benchmark comparisons against CoreML for quantized LLMs on M-series chips