When to use MLX

Apple lists MLX as an open-source array framework for experiments, training, research, and fine-tuning on Apple silicon. The project lives at github.com/ml-explore/mlx. The Python API follows NumPy. There are also C++ and Swift APIs.

Use MLX when you want to try a model on a Mac, train it, fine-tune it, or shrink it. Arrays live in shared memory. Work can run on the CPU or the GPU. Use MLX for research and testing on the Mac. It is not the API you put in an App Store app to run the model.

import mlx.core as mx

x = mx.array([1.0, 2.0, 3.0])
y = x * 2

That is the public array API. Training code is longer. Keep training in MLX or another Mac trainer until the model does the job you want.

Then export if you will ship

An App Store app that runs your own neural model on iOS 27 or later should load a Core AI .aimodel. After the MLX work is done, export the model for Core AI. PyTorch to Core AI in Xcode shows one export path with coreai-torch. Use that course when you are ready to convert the trained model to a .aimodel.

MLX can also run models on a Mac for local tests. That is still a Mac test. It is not the iPhone app runtime. The iPhone app runtime for a custom neural model on 27 or later is Core AI.

Core AI models, typed lists official Core AI presets you may export after you study them. Model architectures in plain English helps you name the input and output before you train.

Next, put the three picks in one table.