The future of AI isn't bigger. It's tiny.

We're a small team building small AI models — each one trained to do a single job really well. Nothing's live yet. We're heads-down making the first few, and they're getting good.

Restraint

Most AI gets bigger to do more. We'd rather go small, and do one thing properly.

Every model we make is aimed at a single, narrow job — and we keep training until it's quick, sharp, and small enough to run almost anywhere. Honestly, the hard part isn't adding more. It's knowing what to leave out.

What we're betting on

01

Cheaper to run

A focused model needs a fraction of the compute. Small enough to be cheap enough to actually use, everywhere.

02

Closer to you

Light models can run near the work — on your own hardware, at the edge — so there's a lot less waiting around.

03

Easy to build on

Small enough to live inside a product or a workflow, instead of sitting behind a far-away API you have to call.

04

Sharp where it counts

Pick one narrow job, train hard for it, and a little model can hold its own against the giants. That's the whole bet.

Vision

The first tiny computers, in the areas we know best.

  • 01

    Software building

    A model per stack, not one for all of coding — each tuned to a single kind of app and fluent in its own tools and conventions.

  • 02

    Infrastructure automation

    Turning intent into config — provisioning, deploys, and the dull glue work that quietly keeps systems running.

  • 03

    Enterprise tooling

    Wired into the software a company already uses — search, summarise, and act inside internal tools.

  • 04

    Domain-specific reasoning

    Narrow, deep expertise — a model that knows one field, like law or medicine, far better than a generalist.

  • 05

    Structured generation

    Output downstream systems can trust — JSON, schemas, and formats that come out well-formed every time.

  • 06

    Agentic systems

    Built to take steps, not just answer — planning, calling tools, and seeing a task through to the end.

Craft

No magic. Just careful work, repeated until the model is small and very good.

Focused datasets
Efficient fine-tuning
Reinforcement learning
Lightweight inference

One day, software won't lean on a single giant model — but a whole set of tiny ones, each good at its own thing. We're starting on that now.

We're not live yet — but we're close. Leave your email and we'll tell you the moment the first model is ready to try.

No spam — just one email when there's something real to show.