One machine
Everything the product does today happens on one computer: the hardware scan, the model profile, the runtime, the documents and the index. The largest model it can run is bounded by that machine’s memory, storage and acceleration.
Private Rack
Several machines you own, one private network
A private AI rack is several trusted computers you own, on one private network, working as one AI system for a team. It is how AI Stays Local is designed to grow past a single machine without sending anything to a cloud.
Planned. The private rack is designed and not implemented. No multi-node execution has been run through AI Stays Local, and nothing here is offered today.
Several machines
A rack would add capacity by spreading work across machines: more people served at once, and model profiles too large for any single machine. That spreading is explicit work done by a coordinator. It is not something the hardware does on its own.
Not one shared memory pool
Each machine keeps its own memory. Four machines with 512 GB each do not become one 2 TB computer, and no design here pretends they do. Distributing a model across machines means deciding which parts run where and moving data between them over the network, and that costs time on every token.
Multi-node execution is designed and not implemented. Each machine keeps its own memory. Nodes would provide aggregate resources through explicit distributed execution; they do not form one shared memory pool.
One machine to a private rack
Begin with the hardware you own. Add private capacity when your work requires it.
Laptop
A compatible 16 GB machine is the entry point, for an efficient local assistant. Not measured yet.
Desktop
More memory brings more capable profiles and larger document collections.
Workstation
High-memory systems reach frontier-class open models, subject to validation.
Private rack
Several trusted machines on your own network, for team access and very large profiles. Planned.
The design
How a rack is meant to work.
Coordinator and trusted workers
One machine coordinates; the others run the work it assigns. Each worker is a computer you own, on your network.
Node identity, approved by hand
A machine joins only after it has identified itself and an administrator has approved it. Nothing joins because it happened to be on the network.
Network
The machines talk over your private network. Its speed decides how well a model split across machines performs, which is why a rack is planned around the network as much as the computers.
Storage
Large model profiles need fast local storage on the machines that run them, and room for more than one version so a failed update never replaces a working model.
Maturity
What exists, and what does not.
| Part | State |
|---|---|
| Single-machine runtime | Private beta for selected testers |
| Multi-machine coordinator | Designed, not implemented |
| Node approval and identity | Designed, not implemented |
| Distributed model execution | Designed, not implemented. Never run. |
| Rack documentation | Not published. Blocked until the design is built. |
Planning a rack?
Tell us how many machines, what they are, and what the team needs. We will say plainly what works on one machine today and what would have to wait for the rack.