Skip to main content

Operate distributed inference without hiding the route

Mycellios coordinates heterogeneous computers so an AI model can use memory and compute beyond a single machine. The network plans a short route, assigns model stages to compatible nodes, keeps each stage's KV cache local, and exposes a familiar inference endpoint.

Early access

Mycellios is an experimental runtime under active physical validation. The software runs real local and distributed stage pipelines, but the public evidence does not yet prove a large model across two residential GPU hosts.

Choose your path

  1. Use the application: install Mycellios and connect it to the public network.
  2. Contribute compute: decide what the machine may share and pause contribution at any time.
  3. Run an isolated network: create a local coordinator and connect trusted machines.
  4. Integrate an application: send requests through the OpenAI-compatible endpoint.
  5. Work on the runtime: use the source checkout workflow.

What exists today

CapabilityCurrent state
Desktop applicationInstallable builds for Windows, macOS, and Linux
Browser workerWebGPU probe with CPU fallback while the page remains active
CoordinatorWorker registration, model discovery, routing, streaming, deadlines, and cancellation
Physical runtimeSelective model stages, persistent transport, local KV cache, batching, and recovery
APIOpenAI-compatible models and chat-completions routes
Multi-host GPU proofPhysical gate exists; passing two-computer evidence is still pending

Mental model

Mycellios does not try to place hundreds of residential computers in one long pipeline. Nearby devices form cells. The scheduler then selects a small number of useful cells or stages for each route. A slow or unsuitable machine can be left out of that request without leaving the network.

The coordinator plans and observes the route. Model activations travel between the stage runtimes selected for the request rather than through a central data relay.

Operating principle

The default path aims to preserve the reference model result. Compression or graph-changing optimizations belong to an explicit approximate mode and must be measured and selected deliberately. See Exact and approximate modes.