Talk to us

Small enough
to get to zero.

CatsAI builds small LLMs made for one job only. Because they are small, they run on ordinary CPUs. And that means you no longer need these three things.

0GPUs

GPUs required

Runs on the CPUs in the servers and PCs you already have.

0dollars

Usage-based API fees

Your costs don't grow with the number of calls.

0bytes

Data sent outside

The model runs in-house or on the device.

For reference: the Lynx model file is 0.67 GB (0.8B, Q4_K_M).

The blockers,
rewritten.

You want to bring in AI, but the project stalls. We hear three reasons again and again. With a small model, here is how each one changes.

  • API fees
    can't be predicted.

    API fees
    don't exist.

    It runs on your own CPUs, so costs don't grow with the number of calls.

  • Internal documents
    can't leave the company.

    Internal documents
    never leave the company.

    The model runs in-house or on the device. Your documents stay inside.

  • A GPU server
    isn't in the budget.

    A GPU server
    isn't needed.

    We build to a size that runs on the servers, PCs and phones you already have.

Open weights.
Models to try right now.

We publish the weights for Lynx, Noir, Serval and Ocelot. Download them and try them as they are on your own PC or phone. If you like what you see, we will build a model for your use case the same way.

* The animations above are samples played back from a script (no model is running on this page). For the real thing, download the weights and try them. The sample texts and names are fictional.

Type something.
It turns black.

Type in the cells on the left. Your text is copied to the cells on the right, and only the cells with personal information turn black. It is a sample of the job Noir does. To try the real Noir, download the weights and run it yourself.

Noir output (sample)

* This box is a sample of the behavior. It is reproduced with simple rules; no actual model is running. Nothing you type is sent anywhere. All names, phone numbers and addresses are fictional.

An honest look
at the alternatives.

Cloud APIYour own GPU serverCatsAI small model
CostGrows with usageGPU purchase and upkeepA fixed price set up front provisional
Where data goesOutside the companyIn-houseIn-house or on the device
Hardware neededNone (connection required)GPU serverThe CPUs and phones you have
If the connection dropsStopsKeeps runningKeeps running
What it can answerBroad. Ask anythingDepends on the modelOnly the chosen use
Time to get startedRight after signupProcurement and setupBuild time needed

The bottom two rows (the outlined cells) are where CatsAI is at a disadvantage. If you need a general-purpose AI right away that you can ask anything, a cloud API is the better fit. Small models come into their own when the use is decided and you want to keep costs and data in your own hands.

From first talk to integration,
4 steps.

  1. 1

    Talk

    We ask what you want to use it for and what data you have on hand.

    From you
    Your use case and a picture of your data
    From us
    Our read on what can and can't be done, plus a quote
  2. 2

    Prototype

    We build small with a little data and check the accuracy in numbers.

    From you
    A small set of sample data
    From us
    A prototype model and its measured accuracy
  3. 3

    Build

    We train the model for your use case and slim it down until it runs on a CPU or a phone.

    From you
    Training data and the specs of the target hardware
    From us
    The model itself, built for your use case
  4. 4

    Integrate

    We also build the business system that runs the model, all in one.

    From you
    Details of the systems to connect
    From us
    A complete system running in the field

Cost and schedule are set and fixed in the first conversation. contract terms are provisional

We don't build a model
and walk away.

CatsAI is a company with years of business-system development behind it. We have only just started building small LLMs, and we don't yet have customer deployments we can make public.

What we won't do is hand over a model and call it done. We connect it to the systems you use today and stay on until it is running in the field. And we show you the accuracy in numbers at the prototype stage.

XX yearsof system development provisional
XXsystems delivered provisional

Frequently asked questions.

Do you have a track record of deployments?

We have only just started building small LLMs, and we don't yet have customer deployments we can make public. Instead, we publish the weights of the models we have built. Run them yourself first and check the accuracy and speed. On a custom build, too, we show you the accuracy in numbers at the prototype stage.

Can I use the released models freely?

Anyone can download the weights for Lynx, Noir and Ocelot and try them. For the terms of business use, please check each model's license. licenses are provisional

Do we have to hand over our internal data?

How training data is handled is decided in our first conversation. If your data can't leave the company, we will propose a way to proceed. wording is provisional

Who owns the finished model?

The delivered model runs in your environment, and you are the one running it. contract terms are provisional

What kind of CPU does it run on?

We decide on the target hardware first, to match your use case and the speed you need, and then build for it. measured figures coming

Can you handle languages other than English?

Serval, our translation model, supports 10 languages, including Japanese, English, Chinese and Korean. Other languages can be added with fine-tuning (Croatian is one we added later).

Start with a conversation.

Tell us what you thought of the released models, or just ask "could this work for our use case?" Nothing is charged until you agree to a quote. provisional

Talk

Talk to us →

(This is a demo; sending is a mock)

DEMO | Placeholder figures