Why AI apps strain a Mac
Apple silicon Macs use unified memory, so the CPU and GPU draw from one pool. A local language model loads its weights into that pool and keeps them there. A 7‑8 billion parameter model at 4‑bit quantization takes roughly 4‑5 GB, and larger ones take far more. On a 16 GB machine that already runs a browser, Slack and a few Electron apps, that is enough to push macOS into swapping to disk.
Storage is the quieter problem. Model files are large, and people tend to download several to compare them. Add image generation tools, Python environments, Docker images and app caches, and a 256 GB or 512 GB drive fills up faster than expected. A nearly full SSD slows everything, because macOS has less room for swap and temporary files.
Cloud assistants cause lighter versions of the same trouble: tabs that never close and helper processes that start at login. The useful tools fall into three groups: ones that show what is happening, ones that show where the disk space went, and ones that remove clutter.
Utilities worth installing
Activity Monitor
It’s already on your Mac (Applications > Utilities), and it’s still where you’d look first. Open the Memory tab and find the Memory Pressure graph at the bottom. If it’s red, you’ve run out of memory. Yellow means macOS is straining, and green means you’re fine. Click the Memory column to sort, and whatever is eating RAM floats to the top, usually a model runner or a browser with forty tabs open.
The limitation: you have to open it, so it tells you nothing while you work in another app.
Stats
Stats is a free, MIT‑licensed monitor that lives in the menu bar. It shows CPU, GPU, memory, disk and network activity, plus battery and sensor readings such as temperature. That makes it handy during a long inference run: you can see the GPU load climb and the memory number creep up without switching windows. It supports macOS 12 and newer.
The tradeoff is that it only reports. It won't close anything for you, and its fan control module is marked as legacy and no longer maintained, so don't count on it for cooling.
DaisyDisk
When storage is the problem, a visual map beats a list of folder sizes. DaisyDisk scans a drive and draws it as a sunburst chart, so a 40 GB models folder or a dusty old set of virtual machine images stands out right away. It can also scan as administrator to reach restricted folders, and it breaks down the vague "System Data" category. The price is a one‑time $9.99.
Deletion is on you, though. There is no automatic cleaning, and what you remove is gone, so look twice before dragging a folder to the bin.
A dedicated cleaner for caches and leftovers
Analyzers show you the space. Cleaners go after the pile of small things: app caches, leftover files from apps you uninstalled, old logs, login items that launch at startup. If you try a new AI app every other week, those leftovers add up.
Options range from free tools that need Terminal knowledge to paid suites with one‑click scans. A roundup of CleanMyMac alternatives is a reasonable place to compare them, since it lines up prices, free versus paid tools, and what to check before you trust any cleaner: how deep it scans, whether it previews what it deletes, and what data it collects. Free tools tend to cost you more manual work; paid suites cost money and sometimes push upsells.
Ollama's own unload controls
If you run models through Ollama, you already have one useful utility. By default a model stays in memory for five minutes after its last request. Run ollama stop <model> to unload it immediately, or set the OLLAMA_KEEP_ALIVE environment variable to change the default for every model. The Ollama FAQ lists the accepted values, including 0 to unload right after a response. Models are stored in ~/.ollama/models, which is also the first folder to check when disk space disappears.
Don’t skip System Settings
macOS already covers a couple of the basics, so check it before you download anything. Under General > Storage you can turn on iCloud offloading, trim Apple TV and Mail downloads, and let the Trash empty itself after 30 days. Login Items & Extensions, in the same General section, shows every helper that starts with the Mac, and many of them do nothing you’d ever notice.
A short checklist before a heavy AI session
- Open Activity Monitor and check Memory Pressure before you launch anything. If it’s already yellow, close a few apps first, because loading a model on top will push you into swap.
- Quit the browsers and apps you aren’t using. Reopening Safari takes a few seconds, and swapping drags on for as long as the model runs.
- Leave 15‑20% of the disk empty. That’s folklore rather than an Apple spec, but a nearly full SSD really does drag.
- Eject models when you’re finished. Ollama will hold one in RAM for five minutes unless you tell it otherwise.
- Plug in. On battery, a long inference run will throttle.
Start with Stats, since it’s free. Add DaisyDisk if the drive is small, and a cleaner only once the leftovers pile up. On an 8 GB machine, nothing here will make a large local model comfortable. A smaller model or a cloud assistant is the realistic fix. On anything with more memory, watching it and tidying up now and then gets you most of the way.