That works, and it means the interesting part of your data has left the building. It also means the feature stops existing the moment the connection does — which, for an app people use in basements and on mountainsides, is not a detail.
cronoWiz takes the other route. The model is loaded into the browser through WebLLM and executed on the GPU with WebGPU. The weights live in OPFS and survive updates, so the roughly 700 MB is a one-time cost and the engine is back in seconds afterwards. Generation happens on the device; the prompt is assembled there and consumed there.
It is worth being precise about the trade-off. A model this size is not a frontier model: it is good at explaining your own numbers back to you and poor at general knowledge. That is the job it was given, and the reason a small local model is enough for it.
-
WebLLM in the browser
no native runtime, no server round-trip, no API key.
-
Weights kept locally
stored in OPFS, surviving app updates: downloaded once, not per session.
-
Warm-up you control
load the engine at app start, at session start, or lazily on first use.
-
An honest limit
a 1–2B model explains your data well and is not a general-purpose assistant.
Free. Private. Yours.
No account. No server. No cookies. Just you, your phone, and an app that minds its own business. cronoWiz is a free PWA that runs entirely on your device — your data never leaves. Ever.
The questions people actually ask
What is WebGPU and do I have it?
It is the browser API that gives a page access to the GPU. Recent Chrome and Edge have it on desktop and on most current Android devices; support elsewhere varies. If it is missing, the coach stays off and nothing else changes.
How long does the first load take?
The download is roughly 700 MB, so it depends on your connection. After that the engine typically returns in a few seconds, and you can have it warm up when a session starts so it is ready by the time you ask.
Is the conversation stored?
The thread stays on the device and travels inside the encrypted backup if you make one. The model weights do not — they are re-downloadable, so there is no reason to carry them.