what your machine needs to think locally
sol thinks about your journal on your own device by default — no cloud, no account, nothing to configure. that takes some real memory, and on linux a GPU. here's the honest bar, and what happens if your machine doesn't clear it.
the floor
- apple silicon (mac) — the local model needs about 13 GB of free memory, so in practice that means a 16 GB machine to have enough headroom day to day. sol checks your free memory before it loads anything, so a busy 16 GB machine can still come up short in the moment — that's expected, not a bug.
- intel mac — not supported. sol and the journal on macOS require Apple Silicon; use an Apple Silicon Mac, or run the journal on supported linux.
- linux — a supported GPU (AMD, NVIDIA, or Intel via Vulkan), with roughly 6–8 GB of GPU memory as a practical guide. there's no single hard cutoff — it depends on your specific card and backend — so sol checks your hardware and tells you plainly if it can't run local rather than guessing.
- windows — sol on windows pairs with a journal running on a mac or linux machine; the journal itself doesn't run on windows yet, so the floor above applies to wherever your journal lives, not to the windows PC.
sol checks all of this during setup, before anything downloads.
if your machine can't clear the bar
nothing breaks — you just pick a different way for sol to think:
- bring your own AI engine — a token from Claude, Gemini, or GPT, or your own OpenAI-compatible endpoint. your billing, your control; the key stays in your journal, and sol pbc is never in the path.
- wait — take it in now, think later. sol can finish setup with no thinking engine chosen at all: it keeps taking in your day and your journal keeps building, sol just doesn't process any of it yet. open the thinking app whenever you're ready to choose.
- confidential processing — coming, scouts first. think off your device on confidential hardware sol pbc runs — no content retained, no human review, nothing used to train, and your journal stays on your device either way. not open yet outside the scout program.
whatever you choose, sol never quietly reaches for a cloud model on its own — if nothing's chosen, it tells you rather than picking for you. see three ways sol thinks for the full picture.
still not sure?
open the thinking app (solstone's web ui, by default http://localhost:5015) — it tells you directly whether this machine can run local, and why not if it can't. or see solstone's memory use for the full memory picture.