About
An AI chat your group runs together.
ClusterChat runs a large open AI model, Qwen3.8 27B, on the graphics cards of the people using it. There's no data centre behind it. Everyone in a cluster lends part of their GPU, and together their browsers run one model that each of them can chat with.
- Everyone gets their own chat. Conversations are separate, and several can run at the same time.
- Nobody has to stay. There's no host. Each piece of the model can be held by more than one computer, so when someone leaves, the others carry on and pick up what they held.
- Nothing to install. It runs in a browser tab.
How it works
The model is about 27 GB, too big for most single graphics cards. So it's cut into pieces, and each computer that lends its GPU downloads some of them.
When you send a message, it passes from browser to browser. Each one runs its pieces of the model and hands the result on, until the last one picks the next word of the reply. That repeats for every word. It's like a torrent, except people share the work of running a model instead of the pieces of a file.
People come and go, so the cluster rearranges itself. A new computer starts with a few pieces and takes on more while it's idle, up to the limit its owner chose. You can also join just to chat, without lending anything.
Private clusters and the public demo
A private cluster is for a group: friends, a lab, a club. Only people with its link can join, and only their computers run its chats. Start one from the home page and share the link. Up to 64 people can join, and whoever starts it chooses how many replies can run at once.
The public demo is a cluster anyone can join. It's there to try ClusterChat, so don't send it anything personal.
The technology
- WebGPU lets a web page use your graphics card. ClusterChat's model code is written from scratch for it: no AI libraries, no plugins, nothing to install.
- WebRTC lets browsers talk to each other. It carries the model's working data from one browser to the next, encrypted, through a relay, so nobody learns anyone else's IP address.
- Two small servers help out. One introduces browsers to each other, and the other relays their encrypted traffic. Neither runs the AI or can read your messages.
All of it is open source. The code is on GitHub.
Why it exists
ClusterChat is an experiment that explores a question: can large AI models be run by ordinary people, on consumer hardware, freely and privately?
In a world where AI is dominated by large, private data centres, or really expensive GPUs, ClusterChat aims to bring control back and make private AI more affordable.
- A more democratic AI. No company gatekeeps AI, deciding who can use what and how. The model is run by the people using it, the code is open source, and anybody can host their own private instance.
- A more accessible AI. It's free, with no account or subscription. A library, a friend group or a community can pool a few graphics cards and run their own model.
- Your own everyday hardware, put to use. It takes advantage of the GPUs that ordinary people, mostly gamers, already own. It puts those computers to work rather than relying on more data centres.
- Your own data. No accounts, and no companies collecting your data to sell you more things. In a private cluster, only the people you invite could see anything you send.
What you need
About 27 GB of GPU memory across the group. Any mix that adds up works: three 10 GB gaming PCs, six 5 GB laptops, or a 24 GB desktop and a 4 GB laptop. A little extra on top means anyone can leave without stopping the chats.
| Browser | Can lend its GPU on |
|---|---|
| Chrome, Edge 113+ | Windows, macOS, ChromeOS, Android. Linux may need WebGPU switched on in chrome://flags. |
| Safari 26+ | macOS, iPhone, iPad |
| Firefox | Windows (141+), Apple silicon Macs (145+) |
Any modern browser can join to chat without lending.
What to expect
- It's slower than commercial chatbots: about 6 tokens a second on two gaming GPUs, and faster on repetitive text like lists and code. It also needs enough GPU memory between its members to run at all.
- It can be wrong. Like any AI, it makes mistakes. Check anything important.
Questions
Can others in my cluster see my chats?
Not in the app. Your chats show up only in your browser, and they're saved there. But the computers in your cluster run the model on what you send, so someone who changed their copy of ClusterChat could read it. Your chat is separate, not secret.
What happens if the person who started it leaves?
Nothing special. There's no host, so the cluster keeps running as long as the people still in it hold the whole model between them. When everyone has left, the cluster closes. Opening the link again starts it back up, and your saved chats are still in your browser.
Do I need a GPU?
No. You can join to chat without lending anything. The cluster as a whole needs about 27 GB of GPU memory to run the model.
How is this different from sharing one chat room?
Everyone has their own conversations, with their own history. Several people can get replies at the same time, and nobody sees anyone else's messages in the app.
A note from me
Yes, as you can probably tell, this project used a lot of help from LLMs. It allowed me to build it, and to be honest even attempt it, way faster than I otherwise could have. The idea, architecture and design choices were all mine, and AI focused on implementation and domain-specific knowledge.
Thanks for checking out my project :)