Language model
Ari
Your prompts never leave home.
The model
We trained her ourselves.
- parameters
- 7B
- the whole architecture
- NumPy
- base models or outside APIs under her
- 0
Xeris Inference Engine
Made for the CPU.
- matrix runtime, written for her
- INT4
- dual-channel memory
- DDR5
- HTTP and RPC, on the machine
- Local
What she does
She only talks.
She does
- Answer questions.
- Run offline, on your hardware.
- Keep every prompt on the machine.
She does not
- Hold keys or sign anything.
- Move funds or touch the chain.
- Browse, call tools or run code.
ARI protocol
The chain sets the limits.
Ari has no role on the chain. Agents registered on it run under these rules.
Limits on every registered agent
- Spend limit
- per transaction
- Daily cap
- 21,600-slot windows
- Whitelists
- contracts and instruction types
- Expiry
- void after a slot the owner sets
- Kill switch
- the owner revokes everything at once
Tech specs
Ari
Model
- Parameters
- 7B
- Written in
- NumPy
- Weights
- Trained from scratch
- Base model
- None
Engine
- Runtime
- Xeris Inference Engine
- Precision
- INT4
- Runs on
- CPU
- Tuned for
- Dual-channel DDR5
Privacy
- Serves
- Local HTTP/RPC
- Telemetry
- None
- Cloud calls
- None
- Prompts
- Stay on the machine
Scope
- Conversation
- Yes
- Tools
- None
- Keys
- None
- Role on the chain
- None
Where she runs
- Computer
- Xeris Computer
- Installed
- At the factory
- Beside her
- A full Xeris node