Small models win only when routing is trustworthy
Most production requests do not need the most capable model. Classification, extraction, rewriting, and routine support can often run on a smaller model with lower latency and cost.
The hard part is routing. If the system sends a diffic...
