A doctor in a rural clinic. A soldier in the field. A device with no signal. We build AI that is fast, accurate, and uses a fraction of the power, attacking the problem at three levels: the engineering around the model, the models themselves, and the compute underneath them.
The bet
Not one enormous model in a data centre answering every question.
Many small ones, each good at a single job, talking to each other, running where the work happens. That is the bet the whole company is built on. Everything below is us making it true.
Mission
We are on the way there, not at the end of it.
We attack the problem at three levels: the engineering around the model, the models themselves, and the compute underneath them. Two of those ship today, on off-the-shelf models. Our own models and our own substrate are not in them yet. When they are, we will say so here.
What we ship
Both do useful work on models anyone can buy. The research underneath is a bet on where this goes, and it is not in the products yet.
A tiny chat-style AI you train on your own product, drop in with one URL, and run for free, for every user, forever.
Agentic orchestration for enterprises: frontier-level agents that run inside your workflows, stop for a person at the steps that matter, and leave an audit trail for every action.
Research
Three directions, all pointed at the same end: units small enough to run where the work is, cheap enough to run everywhere, and good enough to trust. None of this is in the products yet.
The unit of model we are building toward: small enough to run where the data is, specialised enough to be right about one thing, and designed to be composed with others rather than used alone. In development.
Making a unit cheaper to store and move, and fine-tuning small models onto tasks that today need large ones. Quality measured and reported, including the results that go against us.
Computing by letting physical chaotic oscillators settle into an answer, rather than a GPU calculating one. One direction among several, aimed at a large reduction in power. Early research: we will say so until it isn't.
Lab to live
Small, specialised models doing real work in shipped products, not benchmarks in a paper.
Computer vision, on-device
A customer needed to recognise food waste on-device, in kitchens with no reliable connection. We turned a cloud-scale DINOv2 vision model into a 50 MB model that runs fully offline on commodity hardware, keeping roughly 80% of its top-5 accuracy.
Function calling, in production
An open-source WordPress plugin that puts an AI command bar in WooCommerce. Type "refund order 1042" and a small, fine-tuned Fern model, not a frontier LLM, turns it into the right action, shown in plain English first and logged, with a person confirming anything destructive.
Supported by
Work with us
Whether you're a hardware partner, an enterprise customer, a researcher, or an investor, we'd like to hear from you.
