How small can a model be and still reason well?
Distillation, data and architecture choices that let compact models punch above their size.
Intelligence,
compressed.
Cerebellum Labs is a research and product lab building small AI models that work alongside people.
Bigger isn't
the only way.
The field is betting everything on scale. Most real work doesn't need a giant. It needs a sharp, reliable partner.
Same work.
A fraction of the cost.
When capable models are cheap enough to run anywhere, intelligence stops being a luxury and starts being everywhere.
Scattered signals, pulled into one small, precise system.
We're early and we work in the open. Experiments, notes and code will be published as they happen.
Capable AI that's nearly free to run.
Quantization, speculative decoding and caching, measured in real cost per task.
Knowing when to act and when to hand back.
Models built as collaborators that ask at the right moment and keep people in control.
Small experts versus giant generalists.
Compact models tuned for specific workflows, tested head to head on real tasks.
Small parts, working together.
Every experiment adds a node. Signals travel further as the network grows, and we publish each step as we go.
Share of brain volume ~10%
Share of neurons ~80%
The cerebellum is about a tenth of the brain by volume, yet it holds most of its neurons and quietly runs the skilled, coordinated work you do every day. Small, dense, always working alongside you. That's what we're building.
Big ideas, made small.
Take the noise of a giant model and keep only what matters. What's left is small, sharp and cheap to run.
Our first experiments and open-source releases are on the way.