solstone's memory use: local-first, and your cloud options

sol thinks about your journal — what you've seen and heard goes in as text and images, and audio becomes text sol can reason over. all of it runs locally, by default — on every device, everywhere. the bundled model runs right in your journal, so your thinking never leaves the machine. a cloud lane is there only as an option: for a machine that can't run a local model, or if you'd rather not spend your device's own power on it.

local is the default

out of the box, sol thinks on your own device:

thinking and screen-analysis — the reasoning over your journal, and describing what's on screen — run through a bundled local model in your journal. your thinking never leaves the machine.
transcription — turning audio into text — also runs locally, with a small on-device model: roughly ~0.9 GB.

sol will not quietly send your thinking off to a cloud model on its own. if a machine can't run a local model and you haven't chosen a cloud lane, sol simply has no model to think with yet — it tells you, rather than reaching for the cloud behind your back.

what it takes to run locally

the bundled thinking model is sized to the machine it runs on — a current consumer machine runs it. the practical floor is roughly 6–8 GB of GPU memory on a supported GPU, or 16 GB of unified memory on apple silicon (the local model itself needs about 13 GB free — sol checks before it loads anything, so a busy 16 GB machine can still fall short in the moment). on disk that's about 3.4 GB on linux, and about 10.5 GB on apple silicon, which runs a larger model — plus roughly 0.9 GB for the transcription model. solstone checks before it loads anything:

apple silicon — it checks your free memory first and won't activate the local model if it won't fit, pointing you to a cloud lane instead.
linux — the local model runs on a supported hardware GPU (AMD, NVIDIA, or Intel via Vulkan). without one, solstone won't activate a local model and points you to a cloud lane instead.

most modern machines clear this bar and run local without you doing anything. if yours is below it, that's what the cloud lanes are for.

choosing in the thinking app

open solstone's web ui (by default http://localhost:5015) and go to the thinking app. it lays out how sol can think as a few lanes, one active at a time:

local — the bundled model runs right in your journal, so your thinking never leaves. this is the default. the lane tells you whether this computer can run one, and flags when it's short on memory or missing a supported GPU before anything loads.
BYO — your own key from Claude, Gemini, or GPT, or your own endpoint URL. your billing, your control; the key stays in your journal. this is the cloud option available today if your machine can't run local.
confidential processing — think without using your device's own power. the parts of your journal sol is thinking about go to confidential hardware sol pbc runs that keeps nothing: no content retained, no human review, nothing used to train — and your journal still stays on your device. this isn't open yet; scouts get first access.

(scouting for solstone? within BYO, sol can provision a Gemini key on your behalf while you scout.)

if solstone is using too much memory

open the thinking app and confirm the local lane is active and your machine has the memory and a supported GPU for it — that's the intended footprint on a capable machine.
if your machine is below the local bar, switch to a cloud lane — your own BYO key.
make sure you're on a current version — recent releases refined these memory checks. see keeping solstone up to date.
still stuck? see getting help with solstone, or file a request and include your journal doctor output.