Network

The swarm

NoviQ is a small control plane wrapped around a large, contributed pool of GPUs. The control plane dispatches and meters; the swarm does the work; a public data layer keeps everyone honest without exposing content.

Four parts

Client, orchestrator, worker swarm, and data layer. Prompts flow right; only token counts flow down to the public aggregates.

Client

Anything that speaks the OpenAI-compatible API - an app, a script, or a browser. It sends a prompt and gets a completion plus receipts.

Orchestrator

The control plane: authentication, worker tokens, job dispatch, and metering. It decides which workers serve a job.

Worker swarm

Contributed GPUs that run the model - full model per node today, layer blocks in swarm mode next.

Data layer

A public dashboard of aggregates - inference volume, contributors, settlement - with no prompt or response content in it.

The orchestrator

The orchestrator is a WebSocket control plane. Workers connect to it and register a class; clients submit jobs; the orchestrator matches them, tracks token counts for billing, and records earnings per job. It is where a node's placement is decided - and, as the privacy stack lands, where trust-tier placement for the boundary layers is enforced.

Metering, not monitoring

The orchestrator records how many tokens a job used - not what the prompt said. Earnings are computed from job metadata, not content.

The workers

A worker is a lightweight client that holds a persistent connection to the orchestrator and serves jobs. There are two classes today and a third on the way:

Native worker - wraps a local Ollama install and advertises a larger model. Highest throughput and earnings.
Browser worker - runs a WebGPU model (WebLLM) with zero install, straight from a tab.
Swarm block - the v2 mode: a worker loads only an assigned layer block, receives activation tensors from the previous node, and forwards to the next. This activates once the swarm engine ships.

Setup and payouts are covered in Workers and earning.

The data layer

Trust in a private network is easier to give when the network's activity is public in aggregate. The data layer publishes volume, contributor counts, and settlement or treasury metrics - the shape of the network's health - while deliberately holding no prompt or response content.

Ephemeral by design

Nothing to retain

Jobs are ephemeral: once a job completes, the worker retains no prompt or response. There is no history to leak because there is no history kept. This is the operational half of the same promise the privacy receipt makes cryptographically.