Early experiment · Non-commercial · Open source
Open AI models on volunteers’ GPUs
AIs@home runs open-weight AI models on GPUs that volunteers contribute: AI that does not have to go through a handful of big companies.
Hardened, not confidential. The volunteer whose GPU answers you can technically read the conversation. How your data is handled
Science
Our GPUs prove theorems
7 proofs by Qwen3.8 27B accepted on prove2me so far (5 of 21 open leaves tried; 2 more closed by reusing a proof already there), where theorems from research papers are stated in Lean 4 and every proof is checked by machine. Trial 8 is running. See the trials
27 pull requests by Qwen3.8 27B merged into Tau Ceti, an open library of formal mathematics where every change must pass the project’s own review. Failures are counted too. A finished project: we stopped on 7 November 2026 at the maintainers’ request. See every pull request
All scienceDev log
- #4Stopping the Tau Ceti experiments
28 of Qwen's 34 Tau Ceti pull requests were merged before the maintainers asked us to stop: what they were, what they cost, and prove2me and a Windows node as next steps.
- #3AIs@home's contributions to Tau Ceti
Five of our twelve Tau Ceti pull requests are merged: what a refactor and a new lemma cost in tokens and GPU time, why the runner stays on the server, and prove2me as a next case.
- #2Cases for AI inference at home for science
Why selling tokens is hard, Stockfish and Lean proofs (Tau Ceti) as science cases for AI at home, where agentic work should run, the week’s costs, and the 24 GB limit.
Live
- Nodes online
- Requests served
- Tokens processed at home
- on consumer GPUs with open models, not OpenAI, Anthropic or Google
Updated UTC
How it works
You ask
Get help from AI the way you are used to: in the chat, through the API, or as a coding agent in opencode. The difference: open-weight models and open-source software instead of a company’s closed service.
A volunteer’s GPU answers
Not a big company’s data centre: volunteers’ own GPUs, at home, run the model (currently Qwen3.8 27B) and write the answer. Anyone with an NVIDIA GPU of 16 GB (a light desktop at most) or 24 GB and more can join; while the network is small, the maintainer adds a few of their own, some of them rented.
The answer comes back
Streamed back the same way. The gateway keeps usage numbers, never the content.
Use it, or give your GPU
Chat
Sign in with GitHub or Google and chat. Chats are temporary unless you save them.
API and opencode
An OpenAI-compatible API with a personal key and a daily limit.
Run a node
An NVIDIA GPU with 16 GB (a light desktop at most) or 24 GB, on Windows or Linux: one installer or one command. Voluntary and unpaid.
Running a node
No open ports
The node only connects outward to the gateway. No router changes.
Sandboxed
The model runs as its own unprivileged user: no access to /home, a read-only system, no core dumps, no swap.
Users stay anonymous
Your node never learns who sent a request: no name, no email, no account id.
Stop any time
One command pauses the node and frees the GPU. Uninstalling removes everything except the model file, unless you ask for that too.