lili AI Router

Model fit

Which model fits your device?

Running a language model on a phone is mostly a memory question. Choose your device and the context length you want, and this page runs the same estimate the app runs before a download — weights, KV cache and overhead against what the system will actually let a model have.

Loading the catalogue…

How the estimate works

Three things have to sit in memory at once: the weights, the KV cache that grows with context length, and a compute buffer. The app adds about 10% on top for runtime overhead. When the shape of a model is not published — which is the normal case for a file you have not downloaded yet — it falls back to the file size multiplied by 1.2, and so does this page.

The result is compared against a budget, not against total RAM. On iPhone and iPad that budget is the physical memory the system reports. On an Apple silicon Mac the weights live in GPU-wired unified memory, which Metal caps at roughly 74% of RAM — a 16 GB Mac has about 11.8 GB to give a model, and anything above that swaps and crawls.

Under 70% of the budget is comfortable, up to 90% is tight, and beyond that the app will tell you the file is too large before you spend the bandwidth.

A note on honesty. Every file size and every piece of model metadata on this page comes from the Hugging Face Hub API, not from us — the same source the app reads. Catalogue last refreshed —. Real-world memory use varies with the runtime and what else is open; treat these as good estimates, which is exactly how the app presents them.

Then what?

Install the app, open Models → Local Models → Search Hugging Face, and download the file you picked. Everything after that happens on the device — including with the network off.

Get the app free How on-device works