Training on hardware you own
The model trains on one machine that sits on a desk or in your home. We build for the computer any consumer can buy.
Why we work this way
If a model needs a data center to learn anything, only the people with data centers get to own one. Everyone else rents, and what you rent can change under you.
So we build for the machine a small company can run in an office. The model stays small, it reads raw bytes, the arithmetic stays cheap, and what it knows keeps growing without starting the whole thing over.
Working inside that limit is what makes the models good. Every gain has to come from the design, there is no bigger machine to fall back on.
- One machine, no cluster and nothing rented
- A few evenings of work while nobody is using it
- The same box trains the model and answers with it
- Every version it makes along the way stays yours
What that means for you
The model learns on your machine, so whatever you give it never leaves the room. There is nothing to upload and nobody else's servers in the middle of it.
No waiting on capacity, no contract to renew, no queue behind somebody else's job. If the machine is free tonight, the model learns tonight.
What it runs on
Any ordinary computer can hold the model and answer with it, no graphics card needed.
It learns on the same machine and on the same processor, a graphics card only makes the night shorter.
The engine and the training scripts are published, so you can download them and train a small byte-level language model on a computer you already own. It is one binary written in C, there is no CUDA, no framework, and nothing else to install, and the same code serves the model once it has learned something.
- Any computer, it runs on the processor
- A free stretch overnight while nobody is using it
- Disk space for the versions it saves
Thinking about running a model in-house?
Tell us what you want it to do and we will tell you what it runs on.