What each precision costs this model, measured rather than assumed.

The Neural Engine runs float16 and nothing else, so reaching the accelerator means accepting whatever float16 does to a particular network — and that varies by more than an order of magnitude between architectures that look alike. Measured across fourteen families on an untrained forward: convolutional stacks land near 1e-3, ViT at 1e-2, and MaxViT at 1.6e-1. Nothing in an export says which kind a given model is.

So this exports it twice and compares both against the eager model, relative to each output's own magnitude. The packages are written to a temporary directory and removed: the answer is the point, not the artifacts.

The float16 figure is Core ML's, which is better than running the same model in half precision throughout — a ViT does that at 1.5 relative, against 1e-2 here — because the runtime keeps the parts that need range in float32. It is not a bound on your own half precision arithmetic elsewhere.