Axom Network · Beta

Big-model intelligence. On any machine.

When your machine can't run the model, a node on the network runs it for you. Your memory, tools, and files never leave home.

How it works

Your machine thinks. A node does the lifting.

Axom runs on your machine exactly as it would locally. Only one thing changes: the model itself runs somewhere with a bigger GPU.

01 · You

Everything stays home

Memory, tools, files, ratings, and history live on your machine. Axom builds each turn locally, then sends only the prompt that turn needs.

02 · The relay

Matched, never exposed

The relay finds a node holding the exact same model and leaf — checked by hash — and passes sealed bytes between you. You and the node never learn each other's address.

03 · The node

The same Axom engine

A headless Axom engine streams the answer back. Nodes dial out to the relay and never open a port, so there's nothing on a node for anyone to connect to.

Private by construction

Sealed end to end. Proven, not promised.

A prompt is personal — it carries what Axom recalled about you. So the network is built so that it doesn't have to be trusted.

Sealed channel

The relay sees noise

Every turn is encrypted from your machine to the node with keys made fresh for each session. The relay only ever carries opaque bytes.

Pinned identity

No swaps

Every node has its own cryptographic identity. Your machine pins it, so nothing in the middle can quietly hand you to a different node.

Nothing kept

Sizes and timings only

Nodes and the relay log how big and how fast — never what was said. The relay's public status names no users and no addresses.

Node classes

Trust is proven by hardware. Never claimed.

Every node has a class, and your machine verifies it — a node can't simply say what it is. You choose the lowest class you'll accept, and the relay will never match you below it.

Verified

Sealed hardware

Confidential-computing GPUs with hardware attestation. The prompt is sealed inside the chip — not even the node's operator can read it.

Protected

An attested Axom node

Consumer hardware running a locked-down Axom node system, proven by the machine's own TPM chip. Every software path to your prompt is closed.

Open

Your own devices

Any machine with no hardware proof. Open nodes serve only their owner's own devices — never someone else's prompts.

The network is in beta. Today's nodes are Open. Protected and Verified are being built now — hardware attestation already verifies on a real TPM.

Measured

As fast as the node. Word for word.

1:1
Remote speed matches the node's own speed, from 0 to 200 ms of network latency.
+1round trip
Time to first word is the node's own, plus a single round trip.
~146bytes
Per token on the way back. A dial-up line could keep up.
~4µs
To seal and open each chunk of text. Encryption costs nothing you can feel.

Over the network, answers came back identical to the same model running locally, byte for byte.

Live

The network, right now.

Connecting to the relay…

—
Nodes online
—
Verified · Protected · Open
—
Models served
—
Open session slots

Read live from the relay's public status. It names no users, no prompts, and no addresses.

Run a node · Coming soon

Lend your machine. Keep your network closed.

A one-command node install is on the way. Here's what it will look like.

Outbound only

Nothing to open

A node dials the relay. No port forwarding, no inbound connections, no exposed address.

Its own identity

A key, not an account

Each node creates its own cryptographic identity on first start. That key is the node.

Class, honestly

Your hardware decides

The install checks what your machine can prove and tells you its class in plain words before it serves anything.

Axom Network

Every machine can run Axom.