prinet
Docs menuIntroduction

Resources

Getting started

Introduction

Private SI inference on GPUs people contribute, with privacy you can measure.

PRINET is a network for private SI inference. Open models run on GPUs that people contribute, and every prompt is split into secret shares so that no single machine can read it. How well that works is not a promise: it is measured, and the number is published.

0.21%

Prompt recovered from one PRINET share

0.20%

Chance baseline: what guessing gets you

2.42 MiB

Extra traffic per token, the cost of the bound

Why PRINET exists

Large models are too big for one consumer GPU, so sharded networks split a model into blocks of layers and run each block on a different machine. That solves the size problem, but not privacy. Every node still handles real activations, and because the model weights are public, those activations can be traced back to your words.

PRINET was built to measure that exposure and close it. Instead of passing real activations between nodes, it passes secret shares. One share on its own is noise.

Principles

  • Measured, not promised. Privacy is a number: how much of a prompt one node can recover. A run only counts as private if that number sits at chance.
  • Decentralized. Contributed GPUs are pooled into swarms. There is no datacenter.
  • Uncensored. No content-policy layer sits between you and the model.
  • No profile. Usage is billed in $PRINET tokens. There is no prompt history to mine.

Where to go next