making edge models faster deserves as much attention as making them smarter
i built metal-autotune to make MLX models run faster on Apple Silicon for a specified workload. it traces a model, identifies promising regions to optimize, and has an AI agent write and refine custom Metal kernels, tested directly on your Apple device. on a KernelBench suite I ported over to Metal/MLX, it found wins up to 2.34x over the compiled baseline.