RAM: the spec that decides which models you can run

On-device AI lives or dies by RAM. How much memory 0.5B to 12B models actually need, and how to check what your Android phone can comfortably handle.

Why RAM is the deciding spec

A language model has to sit in memory while it thinks - the whole model, loaded and ready. Storage decides what you can download, but RAM decides what you can actually run. That is why two phones with the same chip can behave very differently: the one with more free memory can hold a bigger, smarter model. When people say a phone 'can't run AI,' they usually mean it ran out of RAM long before it ran out of ideas.

The rough ladder

As a rule of thumb, quantized models need about a gigabyte of room per billion parameters, both in storage and in memory while running. Models around 0.5B to 2B run comfortably on almost any modern phone; the 3B to 4B range suits mid-range devices; 7B to 8B wants a recent high-end phone; and 12B models are flagship territory - possible, but only with plenty of memory free and nothing else competing for it.

Checking what you have to work with

Android reports total RAM in Settings under About Phone, but what matters is free RAM at the moment you chat. A quick experiment beats a spec sheet: install Local AI, download a small model first - it will run on practically anything - then step up a size and watch whether replies stay smooth. If the app stutters or reloads when you switch away, you have found your phone's comfortable ceiling, and smaller models live below it happily.

Squeezing out headroom

You do not always need more hardware - sometimes just fewer roommates. Close heavy apps before a long session, restart the phone occasionally to clear stale memory, and prefer the smaller model for quick questions so the big one loads only when it earns its keep. And keep perspective: a well-prompted 3B model answers everyday questions beautifully. Bigger RAM buys headroom for bigger models, not a requirement to enjoy AI on a phone at all.