Independent AI research
Intelligence,
before words.
We build models for the moment
understanding becomes a
decision.
A thought doesn’t
have to be
a sentence.
Language is one expression of intelligence. Judgment is another.
We study models that can read a situation, weigh the possibilities, and return a useful signal. Compact enough to work inside software. Explicit enough to be measured.
Our first modelDo the useful work.
Match the model to the decision. Spend computation where it changes the result.
Keep the possibilities.
Return a distribution. Let the application choose what happens next.
Make it examinable.
Work toward releases with weights, evaluation methods, and documented limits.
A model for
the moment
of decision.
Give Hunch a state, a question, and possible answers. Get a probability for every answer in one forward pass.
Explore the interface
State in.
Probabilities out.
Three primitives.
One programmable interface.
“I’ve been charged twice for my last order. Can you help?”
QUESTIONWhere should this ticket go?
A walkthrough of the interface, not live inference or an evaluation result.
Judgment you
can inspect.
Hunch is in research preview. Its interface is simple; understanding where it works takes careful evaluation.
Read the model briefThe application stays in control.
Choose candidates, set thresholds, and decide when to ask for human review.
A probability is a claim to test.
Calibration can change across tasks. Evaluate the model on the situations where you will use it.
The limits are part of the work.
The preview explores routing and triage. Release evaluation and work on unfamiliar task families are ongoing.
Farther out.
Closer to understanding.
Antares Labs is an independent research lab exploring efficient, useful machine intelligence.
We begin with judgment: how a model represents possibilities, how it expresses uncertainty, and how those signals become part of real software.
Hunch is our first model.