This is Ari.

A seven-billion-parameter language model we wrote from scratch in NumPy, built to run on the CPU of the same machine that runs the chain. She holds a conversation. By design, that is all she does.

How we built herRun her on a Xeris Computer
Parameters
7B
Written from scratch in
NumPy
Runtime on the CPU
INT4
Telemetry or cloud calls
None

How we built her

Seven billion parameters, written from first principles.

We wrote the model in NumPy and the engine that runs it, and we trained the weights from scratch. There is no third-party API or base model under her.

  1. No framework underneath

    The architecture is written in NumPy and does not depend on PyTorch or a hosted API. No upstream policy or deprecation applies to her.

  2. Seven billion parameters

    She is a 7B-parameter model we trained ourselves, from scratch. We also own the weights, the evaluation suite and the training pipeline.

  3. An inference engine for the CPU

    The Xeris Inference Engine runs the whole model on the CPU with a custom INT4-optimized matrix runtime. Inference is bound by memory bandwidth, so the engine is tuned for dual-channel DDR5.

  4. Installed on the computer

    She ships factory-installed on the Xeris Computer beside a full Xeris node and serves inference over local HTTP/RPC. Prompts stay on the machine.

What she does

Her only function is to hold a conversation.

Ari is a conversation-only model. You ask a question and she answers it. By design she has no tools, keys or authority, on the chain or anywhere else.

She does

  • Answer questions and hold a conversation.
  • Run on your own hardware, offline if you like.
  • Keep every prompt on the machine it was typed into.

She does not

  • Hold keys or sign anything.
  • Move funds or touch the chain.
  • Browse the web, call tools or run code.

Ari and the chain

The chain already has the rules for an agent with authority.

The Xeris L1 is all we build. Ari is our model, and for now she runs beside the chain with no role on it. The chain’s agent rules were written before she was ready for them.

Today the two only share a machine. The Xeris Computer runs a full node and serves Ari on the same 65 W machine, and she has no standing on the chain.

Bounded agents were in the chain’s design from the start. The ARI protocol in the runtime registers an agent under an owner with a per-transaction spend limit, a daily cap, whitelists of the contracts and instructions it may touch, and a kill switch the owner can pull at any time. An agent can delegate downward only within tighter limits than its own.

She already runs where the chain runs, and any authority she is given would pass through those primitives, with caps, whitelists and a kill switch. If Ari is ever given a role on the chain, it will be a bounded one, and until then she only talks.

Read the ARI protocol in the whitepaper

What the runtime enforces on any agent

Spend limit
per transaction
Daily cap
measured in 21,600-slot windows
Whitelists
of contracts and instruction types
Kill switch
the owner revokes everything at once
Sub-delegation
bounded depth, tighter limits at each tier

Talk to her on your own hardware.

Ari ships on the Xeris Computer, pre-installed and served locally, next to a full node of the chain.

The Xeris ComputerRead the whitepaper