Jargonautica
Running It in Production

Model Router

Sending each request to the model that suits it, rather than sending everything to the largest one. Most traffic is simple and a small model handles it identically at a fraction of the price. The difficulty is deciding which is which without spending as much on the decision as you saved.

Technical view

Routing signals include a lightweight classifier over the request, task type, required capabilities such as tool use or long context, and cascade approaches that try a small model and escalate when confidence or a verifier check fails. Must be evaluated as a system, since routing errors concentrate on exactly the hard cases the small model handles worst. Also serves availability, by failing over across providers.


Part of AI Terminology Index, section: Running It in Production.

How it connects

Model Router links to 8 other terms. The faint lines are links between those terms. See it in the full graph.

Cost Per TaskCost Per TaskModel Confidence and CalibrationModel Confidence an…Small Language ModelSmall Language ModelAI GatewayAI GatewayFallback and FailoverFallback and Failov…Early ExitEarly ExitAI WrapperAI WrapperShadow DeploymentShadow DeploymentModel Router