NoviQNOVIQDocs
Launch app

Getting started

Private inference, actually proven

NoviQ is an uncensored, decentralized inference network powered by GPUs that people contribute, not rent - and it is the first network built to prove your prompts stay private, per run, to a skeptic.

What is NoviQ

NoviQ runs large language models across many machines that nobody centrally owns. A model too big for any single GPU is split into contiguous layer blocks, and each machine in the swarm runs only its block. Activations stream from one node to the next over a sealed wire until a token comes out the other end.

That much - fast, correct, decentralized inference over the open internet - was already proven by prior work (Shard). NoviQ inherits it and takes on the problem Shard left open: a node that runs your layer must decrypt to compute, so it sees the activations it processes and can reconstruct a large fraction of your prompt. NoviQ turns that leak into a measured number, drives it down by construction, and puts it in a receipt.

SEALED WIRE - ACTIVATIONS ONLY, NO NODE HOLDS THE WHOLE MODELPrompttokensEmbed / headlayers 0-kWorker Alayers k-2kWorker Blayers 2k-3kTailfinal + logitsTokensampledTrusted node (pinned)Untrusted node (sees only rotated activations)
A prompt is embedded on a trusted head, streamed through untrusted worker blocks, and finalized on a trusted tail - no node holds the whole model.

The one-line thesis

Thesis

Shard proved a frontier model can be served across machines nobody owns, fast and correctly, with a receipt a skeptic can check. It also showed a node running your layer can reconstruct a large fraction of your tokens. NoviQ's job is to turn that leak into a number, drive the number down by construction, and put it in a receipt. Privacy earns its word phase by phase, never on day one.

Why it matters

Every hosted AI provider asks you to trust a promise: that they will not log, train on, or hand over your prompts. Decentralizing the compute does not automatically fix this - it can make it worse, because now strangers run your layers. The industry answer so far has been the word "private" with nothing measurable behind it.

NoviQ treats privacy the way good systems treat correctness: as a measured, adversarial, per-run property with a reproducible number, not a marketing claim.

~35-59%

Prompt tokens a malicious node could recover in prior work - the number we exist to drive down

0.3%

Recovery on an untrusted node with per-request obfuscation, output bit-identical

per run

A privacy receipt proves a specific request, not a blanket promise

The three pillars

Proven private

A reconstruction adversary measures how much of your prompt an untrusted node can recover. Defenses drive that number below a declared threshold, and a privacy receipt records it per run.

Decentralized

Inference is sharded across contributed GPUs over the open internet. No node holds the whole model, and no single operator holds the network.

Uncensored

No policy gate sits between you and the model. You are billed in a token, not monitored by a profile - there is no prompt history to mine.

Receipt-backed

Both correctness and privacy ship as receipts a third party can re-check: distinct nodes, real WAN latency, an output hash, and the worst-case token recovery across untrusted nodes.

The honesty rule

Non-negotiable

"Private" earns its word phase by phase, never on day one. We publish the reconstruction number even while it is bad, and show it falling. A receipt proves a specific run was private to a measured degree - it is not a blanket guarantee, and we never present it as one.

Where to go next

NextHow it works