Distillation trains a small "student" model on the outputs of a large "teacher" model. Instead of learning from raw text alone, the student learns to reproduce how the teacher responds — and inherits a good deal of its behaviour at a fraction of the size and serving cost.
This is why model families exist. The fast, cheap tier of a lineup is typically not a separately developed product but a distilled descendant of the flagship, which is how a small model can feel disproportionately capable for its price. It is also why the gap between tiers narrows over time: each generation distils from a stronger teacher.
Why it matters at your desk. It is the reason "use the cheapest model that works" is better advice than it sounds. Distilled models often match their parent on well-defined, repetitive work — classification, extraction, summarising a known format — while costing far less and responding faster. Reserve the expensive tier for genuinely hard reasoning. Releases pitched on agent speed and cost are selling exactly this trade.
Where it breaks down is the long tail. Distillation transfers the teacher's common behaviour more faithfully than its rare behaviour, so students degrade first on unusual inputs, specialist vocabulary, and long multi-step reasoning. If your work is the long tail, test rather than assume.
What to watch for: distillation also transfers the teacher's mistakes. A student inherits its parent's blind spots and biases along with its strengths — a smaller model is not a more neutral one.