One giant model — or five hundred small ones?
Instead of a single enormous model in a datacentre, fogswarm runs small models inside ordinary browser tabs. Your tab downloads a model, runs it on your GPU, and joins a shared pool. A coordinator learns what your particular model is actually good at — and what it is hopeless at — then routes work to whoever can do it.
Nothing you type is stored on the server, and the model runs entirely on your machine. Joining costs you a one-off download and some GPU time.
1 Pick a model and join
checking whether your browser can run models…
The model downloads once and is cached in this browser. The first join can take a few minutes on a slow connection.
2 Finding out what your model can do
Every joining agent sits a short exam — 24 questions with known answers, across 8 kinds of task. This is how the swarm learns to trust you with some jobs and not others.
idle
3 Your agent's profile
Each score is the swarm's estimate of how often you get that kind of task right. It keeps moving as you answer more work.
4 Give the swarm something to do
Ask anything — the coordinator picks whichever agent it thinks can handle it.
the swarm right now
Routing only gets interesting with more than one agent. Send this page to someone: