PRINET runs open models on GPUs that people contribute, and splits your prompt into secret shares so no single machine can read it. Uncensored, decentralized, and private by a number you can check.
Send the same prompt through
How a private run works
01 / Split
Your words never travel whole.
The model's activations are cut into two secret shares. Each node computes on its own share and never holds the real tensor. One share on its own is noise.
Share A
alone: noise
Share B
alone: noise
both shares together - the only way back to the prompt
02 / Route
Run by people, not a cloud.
Contributed GPUs are pooled into swarms that serve open models - no datacenter, and no content-policy layer in between.
03 / Measure
Privacy with a number on it.
We score one node's view for how much of the prompt comes back. A run only counts as private if that number sits at chance.
Why this exists
The leak
The model is public. On a sharded network, your prompt can be too.
Spreading layers across strangers' GPUs solved the size problem. But every node still handles real activations, and with the weights public, those can be traced back to your words. Entry and exit see the most. PRINET was built to measure that exposure, and close it.
The evidence
Five ways to run a model, one question: given a single node's view, how much of the prompt comes back? Each is scored against live model shares, not estimated.
Anything that doesn't split the data leaks almost all of it. Privacy has a price, and we print it: 2.42 MiB of extra traffic per token.
See the methodYour prompt arrives as plain text, and can be logged.
Raw activations at every layer.
Any transform you apply, an attacker with the weights can undo.
Every hop handles real activation tensors.
At chance: no better than guessing.
Get involved
Lend your GPU in the background through Ollama or vLLM. Serve the largest models and earn the top rate.
Top rate / token
npx @nuroaixyz/worker --token YOUR_TOKENContribute straight from a browser tab with WebGPU. Nothing to install, no terminal.
Entry rate / token
Check your browserChange one base URL and keep your code. Chat completions with no logging, billed per token.
Per token
POST /v1/chat/completionsPayments
Every run is billed in tokens, not tracked in a profile. The chain records how much compute was used - never what was asked.
Tokens are the billing unit. There is no prompt history and no profile behind them.
30% of every settlement underwrites the research that keeps the privacy bound honest.
If a run fails, it isn't charged.
Questions
A network for private SI inference. Open models run on GPUs that people contribute, and your prompt is split into secret shares so no single node can read it. How well that works is measured and published in the Paper.
Not from its own view. A single share recovers 0.21% of a prompt - the same as guessing, which sits at 0.2%. The entry and exit nodes see the most, and the Paper says so plainly.
There is no per-request receipt yet. Today the guarantee is a measured gate: a run that doesn't meet the bound doesn't count as private.
A hosted API receives your prompt as plain text and can log it. Obfuscating open weights or simply sharding the model doesn't fix that - in our measurements they leak 94–100% of the prompt.
Any open model the network serves, with no content-policy layer in the way.
About 2.42 MiB of extra traffic per token to keep activations secret-shared. That's the price of the bound, and we publish it rather than hide it.
No. Bring a GPU - native or in a browser tab - and get paid per token you serve. Users pay for inference; workers don't front anything.
Inference is metered in $PRINET. 30% of every settlement funds the privacy research, and failed jobs aren't charged.
Only the one shown on this site. Any other address claiming to be $PRINET isn't ours.
Open Chat for a private run, Earn to attach a GPU, or the Paper for the full measurement. Questions go to PRINET on X.
Start here
Talk to uncensored open models without handing your prompt to anyone. Or put an idle GPU to work and get paid for every token it serves.