Getting started

How it works

A request enters through an OpenAI-compatible endpoint, is dispatched to contributed GPUs, runs as a sharded pipeline over a sealed wire, and returns with receipts that a skeptic can re-check.

The shape of the network

NoviQ has four moving parts: the client that sends a prompt, the orchestrator that dispatches and meters work, the worker swarm that actually runs the model, and a public data layer that publishes aggregates with no prompt content in them.

Client to orchestrator to worker swarm. Only token counts flow to the public data layer - never prompt or response content.

Life of a request

Submit. A client calls the OpenAI-compatible API with a prompt. No account profile is attached to the content - the unit of account is the token, not your identity.
Dispatch. The orchestrator selects workers, taking trust tier and locality into account so the leaky boundary layers land on trusted nodes.
Run. The model runs as a pipeline: the trusted head embeds tokens, untrusted workers forward hidden states over the sealed wire, and the trusted tail produces logits and samples the next token.
Receipt. The run emits a correctness receipt (and, as the privacy stack lands, a privacy receipt) that anyone can re-check.
Settle. Workers are paid per job, but only against a valid receipt. Metering records token counts only.

Inside the pipeline

A transformer is a stack of layers. Split the stack into contiguous blocks, put one block on each machine, and stream activations through them in order. Each node holds only its block in VRAM and never sees the whole model.

The sharded pipeline. The head and tail are pinned to trusted nodes; the middle blocks run on untrusted, contributed GPUs.

Why round-trips, not FLOPs

Over the open internet the scarce resource is the network round-trip, not raw compute. NoviQ leans on speculative decoding, asynchronous pipelining, a direct-return ring, and in-region clustering to keep decode fast despite the hops. See Sharded inference.

Two receipts, not one

Correctness and privacy are separate, independently checkable claims:

Correctness receipt. Proves the split run produced the same output as the whole model would have - distinct nodes, real WAN latency, and an output hash. See Correctness receipts.
Privacy receipt. Records the measured worst-case token recovery across untrusted nodes and whether it passed the declared threshold. See Obfuscation and receipts.

What runs today

Live

Single-node workers (Ollama and in-browser WebGPU) metered by the orchestrator

Proven

Correctness (Gate A) and privacy obfuscation (Gate D) on a CPU proof harness

Multi-machine WAN pipeline and the permissionless worker economy

Honest scope

The proof gates run today on small models on CPU in seconds. The production swarm engine (WAN pipeline, KV cache, permissionless join) is being built in the open along the sequencing rule: correctness first, then real WAN, then cache, then the privacy proof.