This is Ari.
She runs on the CPU beside the chain and, by design, only talks.
- Parameters
- 7B
- Written from scratch
- NumPy
- Custom CPU runtime
- INT4
- Telemetry, cloud calls
- None
How she’s built
Seven billion parameters, written from first principles.
No framework underneath
NumPy only. No PyTorch, third-party API, upstream policy or deprecation.
Trained from scratch
No base model. We own the weights, evaluation suite and training pipeline.
Xeris Inference Engine
Ours. Whole model on the CPU; bandwidth-bound, so tuned for dual-channel DDR5.
On the Xeris Computer
Factory-installed beside a full Xeris node, served over local HTTP/RPC.
What she does
Her only function is to hold a conversation.
She does
- Answer questions.
- Run on your own hardware, offline if you like.
- Keep every prompt on your machine.
She does not
- Hold keys or sign anything.
- Move funds or touch the chain.
- Browse the web, call tools or run code.
Beside the chain
The chain already has agent rules.
For now
- Our focus
- The Xeris L1, all we build
- Ari’s chain role
- None; she only talks
- Shared machine
- Xeris Computer: full node + Ari, 65 W
- Agent rules
- In the design from the start, before Ari
- ARI protocol
- Registers an agent under an owner
- Any future role
- Bounded, through the ARI protocol
Enforced on any agent
- Spend limit
- per transaction
- Daily cap
- measured in 21,600-slot windows
- Whitelists
- of contracts and instruction types
- Expiry
- a slot after which the delegation is void
- Kill switch
- the owner revokes everything at once