Models
Kaer Reasoners
Small reasoning models designed to decide how long to think: to spend compute on hard questions and flag the ones they're likely to get wrong.
About Kaer-R1Independent AI lab · London
Reliable reasoning, measured in the open.
We're building small reasoning models that know when to stop, and ways to test how AI systems behave in everyday use. Everything we publish comes with its method, its cost and its limits.
NewsIntroducing Kaer Labs
What we do
We build small reasoning models, study how models behave in real use, and test them for the services that rely on them. Our own models get the same tests as everyone else's.
For institutions
Partner studies
Before a public service puts a model in front of people, we can test it on the service's own cases and report where it fails.
Start a study
Kaer-R1 · In development
A model that knows when to stop
Kaer-R1 is the first model we're building: a small reasoning model with a learned halting head. It decides, question by question, how long to think, and the same signal flags questions it's likely to get wrong, so they can go to a person or a larger model.
Benchmarks happen
in the lab
Behaviour happens
to people
Research
How to make models reason well on less compute, and how to tell whether a model is doing its job once people rely on it. Here is what we're working on now.
All research
- Can a model learn when more thinking will help?Training a halting signal that decides, question by question, whether to keep reasoning, and when to hand over instead.
- Which training data improves reasoning?Measuring each part of a training mix by what changes when it's taken out.
- How much does wording change an answer?Testing the same question in many phrasings, and reporting the spread rather than one score.
How we work
Five rules we apply to our own models before anyone else's.
Blind by default
When we test our own models alongside others, they go in under anonymous IDs. The people scoring them don't know which entries are ours.
Every score has a price
Accuracy is always reported next to the cost per solved task, measured at batch pricing.
Method before results
Protocols are published before the numbers. Test sets stay private so they can't leak into training data.
Failures in writing
Failed runs, harness bugs and known weaknesses are written up next to the results.
Judged by practitioners
Caseworkers, nurses, teachers and advisers score the answers, and they are paid for their time.
Good AI leaves people
more capable.
We measure whether a system strengthens the judgement of the people who use it, not only whether its answers are correct.
Read the researchQuestions
What does Kaer Labs do?
We're an independent AI lab in London. We're building small reasoning models, the Kaer Reasoners, and ways to test how AI models behave in everyday use, including for the public services that deploy them. We'll publish methods and results as the work matures.
Do you train your own models?
Yes. Our first model, Kaer-R1, is in development. It's a small reasoning model designed to decide how long to think about each question. We don't compete with frontier labs on scale; we're focused on getting reliable reasoning out of less compute.
You build models and evaluate them. Isn't that a conflict?
It would be if we marked our own work. When we evaluate our own models, they're entered blind, under anonymous IDs, and scored by people who don't know which entries are ours. We'll publish the protocol so anyone can check the result.
Can we use the models?
Not yet. Kaer-R1 is still in development. If you'd like early access, or want a model tested before you deploy it, tell us what you're building.
Putting a model in front of the public?
Talk to us before you launch. We'll test it on your real cases and tell you where it fails.
Start a conversation