Systems Design · AI
Model Router
Sending simple chores to small bikes and heavy cargo to big trucks.
1. Not every question needs a supercomputer
Asking for a synonym does not need a giant model that costs five cents per query.
Why have different sizes?
Big models can solve complex math, but they are slow and expensive. Small models can format text or answer simple facts in a fraction of a second for almost nothing.
2. The traffic sorter
The router inspects the question and picks the cheapest model that will do the job well.
How does the router know?
A quick mathematical check or tiny classifier reads the structure of the prompt. If it spots multi-step code, it sends it to the expert.
3. The escalation ladder
Try the fast model first; if the answer fails automated checks, call the expert.
Why not always use the giant model?
Because most requests succeed on step 1. You get 90% of your answers ten times faster for a tiny fraction of the bill.
Longer version: Reference Entry on Model Router
Sources: Ong et al., RouteLLM (2024); Chen et al., FrugalGPT (2023).