Notes on Small Models: Why Compact Reasoning Beats Raw Scale in the Classroom

The prevailing assumption in AI is that capability scales with size: bigger models, more parameters, better answers. In the classroom, that assumption breaks down against three hard constraints — cost, connectivity, and latency — that a data centre never has to feel.

A model that reasons well but requires a permanent connection to a hyperscaler is, in a school without reliable internet, a model that reasons not at all. A model that costs a few cents per query at demo scale becomes a budget line no ministry can sustain across ten million learners. The question worth engineering against is therefore not "how large can the model be?" but "how small can it be while still being right?"

That reframing changes what we build. Compact reasoning models tuned to a curriculum, running on local hardware, can deliver the pedagogically important behaviours — staying on the syllabus, showing its working, adapting to a learner — without the dependence and the cost. The intelligence lives where the teaching happens.

Scale still matters. But in education, the scale that counts is the number of classrooms a system can reach, not the number of parameters it carries. We optimise for the former.