Engineering noteProposed topic — outline, not published findings
Small Language Models for Specialized Applications
When a smaller, adapted model beats a frontier model — on cost, latency, privacy, and control.
The case for small
Narrow tasks rarely need a general-purpose frontier model. A small model adapted to a domain can be faster, cheaper, easier to evaluate, and deployable where data can't leave.
Where adaptation helps
Domain vocabulary, fixed output formats, and repetitive structured tasks respond well to fine-tuning and quantization. The trade-off is breadth: small models fail more often outside their lane.
What we want to test
Head-to-head evaluations on real tasks: quality, latency, and cost per correct answer — including the cost of the adaptation itself.
