Every headline goes to the largest frontier model. But look at what actually ships in products and you'll find small models doing the work — running on a laptop, a phone, or a cheap server, answering in milliseconds.
What changed
Small models used to be visibly worse. Three shifts closed much of the gap:
- Distillation. Train a small "student" on the outputs — and increasingly the reasoning traces — of a large "teacher." The student inherits capability far beyond what its size alone would predict.
- Data quality. Carefully curated, deduplicated, high-signal data lifts small models more than raw scale does.
- Task focus. A small model tuned for your task beats a giant generalist on that task, at a fraction of the cost.
The right question for most teams is no longer "which is the best model?" but "how small can we go before quality drops below the bar?"
The economics
Cost, latency, and privacy all point the same way. A small model you can run yourself is cheaper per call, faster, and keeps data on your own hardware. For high-volume, well-scoped work, that combination is decisive.
The pattern that wins
The emerging default is a cascade: a small model handles the common cases instantly, and escalates only the genuinely hard ones to a larger, slower, pricier model. You get frontier quality where it matters and small-model economics everywhere else.
