Skip to content

SilentSwarm documentation

SilentSwarm is a community compute pool: members share their GPUs so the pool can run and train models no single machine in it could hold. Under that, it is a leader/fellow runtime for distributed transformer experiments with explicit control-plane orchestration, tensor data-plane transport, and cluster-backed experiment runs.

  • Get started


    Bring up a leader, add a GPU machine with one command, and chat with a model. About ten minutes.

    Get started

  • How it works


    The leader/fellow split, how a model is cut into stages, and what actually crosses the wire between them.

    Architecture overview · Runtime model

  • Run it


    Bootstrap workers, deploy behind a domain, manage accounts, and understand the public web surface.

    Operations · Deployment

  • Look something up


    Exact routes, contracts and commands, for when you already know what you are looking for.

    Protocol and API · CLI cheat sheet

  • Is this the right tool?


    Where Petals, Exo, GPUStack, Ray or a plain single-node stack is the better answer - and what SilentSwarm has not built yet.

    How SilentSwarm compares

  • The numbers


    What we intend to measure, the rules we hold ourselves to, and why the results table is still empty.

    Benchmarks

A reading path

If you would rather be led than choose:

  1. Start a local cluster and run one smoke sweep - Get started.
  2. Read the architecture overview and the runtime model.
  3. Use the experiments and operations pages for day-to-day work.
  4. Reach for the reference pages when you need exact commands, routes or contracts.

Writing docs rather than reading them? The conventions and the local build are in Contributing.