LAPSUS is a peer-to-peer network of personal machines running local models. When your GPU sits idle you share it; when you need more, you borrow someone else's — and the prompt travels directly from your machine to theirs. A thin coordinator only introduces the two peers and keeps the credit score. It never sees your prompts.
Prompts and answers go direct & encrypted — they never touch a server.
Community-shared resources instead of a handful of cloud giants. Voluntary, non-commercial — a commons, not a marketplace.
You share resources and you can use others'. Requests go straight from client to client over a direct encrypted channel — no server in the path.
Share a text model; tap someone's image generation or text-to-speech in return. Together we stay independent and build a decentralized local-AI network.
Sharing your AI earns Compute-Credits; using the network spends them. The thin coordinator keeps a ledger of mutually-signed receipts — it never sees a prompt.
Every request you serve credits your account, with a receipt signed by both peers. The more you offer, the more you can use.
A request costs credits scaled by the model and the tokens in and out. You spend what you've earned — new peers get a small starter grant to try the network.
Before a peer spends compute on you, the estimated cost is reserved from your balance, then settled against the real usage — so a provider is never left unpaid, you can't overspend, and the unused estimate is refunded.
Sharing is bounded — all set by you, enforced on your own machine. Nothing runs unbounded.
The contribution % is a share of a full day's generation. Your machine measures its own speed (tokens/sec); the budget = speed × your % × a day's length (4, 8 or 12 h — your call). Hit it and the network stops routing to you until it resets.
A hard max output tokens per request (default 1024) protects your VRAM and time — no one can ask for an essay and tie up your GPU. And max concurrency (default 1) means one job at a time.
Your machine watches its own throughput: if a served request suddenly runs far slower than usual, it assumes you're using it and pauses sharing for a cooldown — so the community never gets in your way. No GPU drivers needed.
An AI model is just a text generator — text in, text out. There's no tool executor, no file access and no shell wiring it to your machine, so a remote prompt simply can't read files, fetch URLs or run code — the capability isn't there. Tool-calling fields are stripped from every request, by design.
Read more →On top of that, every served request is prefixed with a light safety prompt, enforced by the serving machine so the asker can't bypass it. Permissive by default — and fully open.
Read the prompt →You set how much you give — a daily budget, max output per request, concurrency — and sharing auto-pauses when you're using the machine. When your quota is hit the network simply stops routing to you until it resets — requests don't even arrive.