Private inference network

Anyone can run the model.
Nobody gets the whole prompt.

Pallium weaves every prompt into two sets of threads before it leaves your tab. Each set on its own is pure noise, so the stranger running your layers is holding a tangle of yarn with no picture in it.

CA soon
Woven field · move to ripple · click to shake the cloth
01 / The problem

Sharing a GPU is easy. Sharing a secret isn't.

Split a model across strangers' machines and every one of them still watches your words walk by. Enough to read the plan, the password, and the name you typed at 3 a.m. and instantly regretted.

02 / How it leaks

Your prompt walks past four strangers.

ENTRYlayers 0–15EXPOSED
NODE 2layers 16–31EXPOSED
NODE 3layers 32–47EXPOSED
EXITlayers 48–63EXPOSED

Every node handles your real activations, and every one of them can read them back. Flip the switch to see what Pallium hands them instead.

The measured one

Private

A raw hop reads your prompt back word for word. One Pallium thread set reads back as static: 0.39%, the odds of guessing a byte blind. The math fits on a napkin, and the napkin is in the Paper.

RAW
recover 100%
WARP
0.39% · chance
RAW HOP100%
SHUFFLED100%
WARP ONLY0.39%

Decentralized

Anyone's GPU, thrown into one pile. No datacenter, no landlord, no single machine that ever holds both the warp and the weft.

POLICY QWENLLAMAANYMODEL

Uncensored

Any open model. No content-policy middleman reading over your shoulder, because there is nothing readable to read.

Privacy isn't a toggle you ship in v2. It's the whole cloth, and Pallium is woven so no stranger ever holds all of it.

03 / Open vs private

Open models. Closed prompts.

API×
hosted inferencelogged forever
SHARD×
layers splitevery hop sees tensors
CLOCK 1.0×
——
WEIGHTS×
run it locallyyour GPU, your problem
SHUFFLE×
secret rotationundone in one line of algebra
PALLIUM×
private run
one thread set is noise0.39% · chance
RECOVERY×
APIhosted100%logged
WEIGHTSlocal100%raw
SHUFFLEopen100%invertible
SHARDmulti94–100%activations
PALLIUM2-set0.39%chance
LOOM 1.0×
weaving, row 01
NOTE.TXT×
Measured, not marketed.
Paper →
Drag the windows
04 / Try it

Weave your prompt into two.

Type anything you would never tell a stranger. It never leaves this page: the split runs on your machine with pallium.js, the same module the docs import.

Warp node 2
Weft node 3
// press Split. Then try to read the warp the way a curious node would.
05 / Product

Three ways onto the loom

ZERO INSTALL

Split in the tab

The splitter above, running on your own CPU. One click, no terminal, nothing sent anywhere.

Free · forever
Open the splitter
RECOMMENDED

pallium.js

One small module, no dependencies. Import it straight from this site and split prompts in your own app.

Ships as one file
import { split } from "SITE/lib/pallium.js"
Quick start
FOR SKEPTICS

Privacy receipt

Every split can print a receipt: hashes of both thread sets and the measured score of the attack. Re-run it yourself.

Per run
const r = await receipt("my prompt")
Read the receipt spec
Private prompts, public ticker

The $PALLIUM token

What goes on-chain is the ticker. What never leaves your tab is the prompt.

01

Counted, not watched

The chain sees buys and sells like it always does. It has no idea what you asked the model last night, and neither do we.

02

Two threads, one coin

A cloak with only the warp is a fishing net. A wallet with one $PALLIUM is a believer. We'd like more of both on the loom.

We keep a FAQ

pallium

Private, decentralized inference. Two thread sets, two strangers, no picture showing.

© 2026 Pallium