Liminality MCP is in open beta. The first 3 solves on any key are free. Connect the MCP →

Three models, one idea.

The model you use should be yours. Tuned to your work, running where you can see it, not rented by the token. We train for that, three different ways.

Sia is the small one that runs on your machine and learns you. Noesis is the full reasoning model underneath. Liminal is a model trained on one piece of software until it knows that tool cold. Same belief, different jobs.

Side by side

What each one is for, where it runs, and where it is today.

Model What it's for Where it runs Status
Sia A personal assistant that adapts to how you work, over time. On your own machine. Offline when you want it. Private beta
Noesis Hard reasoning, the full architecture, not a distilled cut. A workstation. Built for the heavier loop. Not yet released
Liminal One piece of software, learned end to end. Alongside the tool it was trained on. By request

Sia

A model that's yours to keep.

A personal model that runs on your machine and keeps adapting to your work. No cloud dependency. No token meter.

Most models stop learning the day they ship. Sia doesn't. You work in your own tools; it privately reviews what mattered, adjusts, and the next time around it knows you a little better. This isn't a memory note pasted into a prompt. The model itself changes.

Day one, Sia is generally capable. Six months in, it writes closer to your voice, knows your project history, and stops over-explaining the things you already know. You see one assistant; underneath, each step routes to whichever specialist was tuned for it.

Runs
On your hardware, today
Offline
One toggle seals every external call
Beta
By August 2026

Noesis

The full one.

Sia is distilled for the machine on your desk. Noesis is the model it was distilled from: the full architecture, multiple specialists each on their own learning loop, built for the reasoning that doesn't fit on a laptop. It isn't released, and we're not going to put a price on it before it's real.

If you want to know when it opens, we'll tell you. No list-building, no drip.

Get notified about Noesis

Liminal

A model that knows your software.

Pick one tool: the design app, the ERP, the in-house system nobody outside the building understands. Liminal is a model trained on that one piece of software until it knows it the way someone who's used it for years does. The menus, the quirks, the order you actually have to do things in.

It's the same training stance as Sia, pointed at a tool instead of a person. Generic assistants guess at unfamiliar software from screenshots and docs. A model trained on the real thing doesn't guess. We build these one at a time, by request, for teams whose software is too specific for anything off the shelf to help with.

Tell us your software

Local or hosted?

The same model can live on your machine or on a box we run for you. Neither is the "real" version. They trade different things. Here's the honest cut.

On your machine

  • Your data never leaves the building. Pull the network cable and it still works.
  • No token meter. You paid once; running it costs you electricity, not API calls.
  • You decide when it updates. Nothing changes under you overnight.
  • The catch: it's bounded by the hardware you own, and you keep the box running.

Hosted by us

  • Nothing to set up or maintain. It's running when you sign in.
  • Headroom for the heavy models without buying a workstation.
  • Reach it from any device, not just the one it's installed on.
  • The catch: your work passes through our infrastructure, and you're depending on us to keep it up.

Not sure which fits?

Tell us what you're trying to do and we'll point you at the right one, or tell you none of them fit yet. We answer our own email.