All research notes

Model Development

Small Models, Specific Jobs

Jun 20, 2026 · 5 min read

Distillation and targeted fine-tuning have matured to the point where a small model, trained on well-curated traces from a strong teacher, matches the teacher on a narrow task band.

The engineering discipline is defining the band honestly. Small models fail hard and quietly outside their training distribution, so the deployment must include an out-of-distribution detector and a clean escalation path.

Our production pattern is a portfolio: several small specialists behind a router, with a frontier model as the fallback and as the ongoing teacher for new trace collection.

This is what diffusion of AI into real value looks like operationally — not one giant model everywhere, but many small ones where they are demonstrably correct.