Your own assistant, on your own hardware
Every mainstream AI assistant is somebody else’s computer. You send your half-formed thoughts to a server you don’t control, on terms you don’t set, and pay by the token for the privilege.
Patch is the other option. It’s a chat interface wired to Qwen3 8B running locally on a laptop at home, reachable from anywhere through a password-gated URL. The model weights sit on the disk, the inference happens in the room, and the conversation history never crosses the network to anyone else.
Nothing leaves the house. That’s the whole feature.
Why 8B is the interesting size
An 8-billion-parameter model is the sweet spot for this kind of project. It’s small enough to run on consumer hardware at a genuinely usable speed, and large enough to be good company — it drafts, summarises, explains, and thinks through a problem without needing a datacentre behind it.
The honest trade-off is that it isn’t a frontier model, and it doesn’t pretend to be. Building on it means designing for a smaller model’s strengths: shorter context, clearer prompts, and an interface that makes it easy to steer rather than assuming the first answer is the final one.
What it’s built from
-
Local inference
Qwen3 8B runs on the machine itself. No API key, no rate limit, no usage meter — the only ceiling is the laptop.
-
Reachable anywhere
A public URL tunnels back to the home machine, so the assistant is available away from the desk without moving the model to the cloud.
-
Password gate
A single shared password sits in front of everything. It’s a private tool with a public address, and the gate is what keeps those two facts compatible.
-
Zero marginal cost
After the hardware, running it costs electricity. Long conversations and throwaway questions carry no per-token anxiety.
The constraint that shapes it
The laptop has to be on. That single fact defines the project more than any design decision: it’s a personal tool with personal uptime, not a service with an SLA. Accepting that is what makes the rest of it simple — no autoscaling, no bill, no account system, no data policy to write, because there is no one else’s data in it.