Local LLM troubleshooting on Mac, iPhone and iPad

Apple hardware fails at local AI along one axis above all others: memory. There is no separate VRAM to run out of, so nothing tells you that you have asked for too much. On a Mac the request succeeds and macOS starts swapping a multi-gigabyte working set to SSD, which looks exactly like a hang. On an iPhone there is no swap at all, so iOS terminates the app the instant it crosses its limit — and because that is a jetsam kill rather than a crash, you get no dialog, just the home screen.

The second Apple-specific problem is code signing. Most local-AI tooling is small open-source projects without an expensive certificate, so Gatekeeper reports them as damaged when they are merely unsigned. The fixes are safe if you verify the download first and reckless if you do not, and every page here that touches xattr says so plainly.

The rest are narrower but expensive to guess at: the Metal working-set ceiling that reserves a third of your memory away from the GPU, an x86 Python running under Rosetta that makes MLX refuse to install, and sustained-load throttling that halves your tokens per second ten minutes into a long job. Two pages at the end cover iPhone and iPad specifically.