Built at the CRECIMIENTO hackathon · Aleph, Buenos Aires 2026 · Pears track
One binary. Your machine runs a local LLM, announces itself signed on the P2P network, and serves requests through an OpenAI-compatible gateway — with no company in the middle keeping the margin or seeing every prompt.
The problem
A handful of companies, in a handful of datacenters, set the price, the rate limits and the terms — and change them without notice. If one goes down or rewrites the rules, there is no plan B.
Your prompt travels to a third party that can log it or train on it. You pay per token, and the seller controls how many tokens get generated.
You can’t verify which model answered you, or audit its quality against what you were promised.
Meanwhile, the bar to join is low — a GPU, 32GB of RAM, a little free disk space — and most machines that already clear it just sit there, idle.
What it is
Pyrus is the pear tree. A Pear is one machine on it — any server capable of running AI, idle or not, on your own network. Join the orchard and it becomes a live node, running inference from minute one.
There is no PyrusLLM server in the middle. Every node on the network runs a local LLM with QVAC, settles payments through WDK, signs a manifest with Ed25519 and announces itself. Each one is both host and guest — it can serve requests, share files with the rest of the network over Hyperdrive, and send requests of its own.
A gateway compatible with the OpenAI API routes each request to the node it belongs to — yours, or another operator’s — and meters what got consumed so it can be settled in batches. Reach it from your terminal, Telegram, WhatsApp, Open WebUI, or anything that already speaks OpenAI.
Payments run on WDK. The economics, end to end: → fiuidao.web.app — quotas, caps, receipts & x402
The app
Your node starts switched off, and the first thing you see is a single button. Launching it puts your server on the network — and that is what gets you access to everyone else’s. Point a browser at it from anywhere on your LAN; there is nothing to install on the machine you are looking from.
Ask, and every answer names the machine that produced it — who served it, time to first token, tokens per second. Or tick local only and nothing leaves the server.
Everyone else. Price, tags and availability per node, and Connect to use any of them from your own OpenAI client.
Your server as a provider: what it publishes, what it charges, how loaded it is, which model it serves.
What the node earns, in one place. Balance, send, deposit, and every settlement it has collected — on Plasma, real chain, real receipts.
What we can and can’t promise
Whoever matches a request to a machine sees both ends. Whoever meters consumption controls the money. Today those tasks have an owner — us — and the plan is for them to stop having one: they move inside an enclave, where we couldn’t look even if we wanted to.
What you can take as true today: your prompt doesn’t pass through any corporation that aggregates it at scale. What you can’t: the node that answers you reads it in plaintext. To infer, you have to read. Closing that needs confidential inference, which already has measured overhead —4–8%— and known limits: memory and the interconnect stay unencrypted, and timing still leaks structure.
Model substitution is detectable. Quantization is not: that’s covered with a signed declaration, a stake, and arbitration. Economic, not technical, and we say so.
Measured overhead: arXiv:2509.18886 · Quantization detection: arXiv:2504.04715 · Optimistic verification: arXiv:2501.14784
Get started
Pick the one that fits your setup — all three give you the same pyrusllm node. It opens in your browser and waits there; no model is downloaded until you send the first prompt.
Rather run a plain binary with nothing to install first? → every build is on the releases page, each with its .sha256.
npm i -g pyrusllm pyrusllm
Pulls the Bare runtime for your platform automatically. Upgrade with npm update -g pyrusllm. No OTA on this channel.
npm i -g pear pear install pear://8f789f6hsf4ghymku5eqwqmbqiubiigp8xpy6boiymunnbyznpny
The installed copy updates itself over the P2P network in ~10 s, with no user action.
git clone https://github.com/gazzimon/qvac_node.git cd qvac_node npm install npx bare bin.mjs
Runs straight from the repo — also the way to hack on it. Needs npm ≥ 10.