Switchyard routing

NVIDIA NeMo Switchyard is an opt-in routing strategy in Ramp Router. It uses recent conversation signals to choose between a capable frontier model and a configurable efficient model for each eligible turn.

Enable Switchyard and choose its efficient model and switching sensitivity on the Strategies page. The model in the request remains the capable tier.

What Router chooses

Router scores signals such as errors in recent tool output, context compaction, tool-call patterns, passing tests, and turn depth. Simpler turns can route to the efficient model, while complex or error-heavy turns stay on the capable model. Switching sensitivity controls which model handles ambiguous turns.

The requested model passes through unchanged when Switchyard is disabled, the request includes image input, the requested model is already in the efficient tier, or an eligible efficient model is unavailable. It also passes through when its catalog output-token price is less than twice the efficient model's, or when either model is missing catalog pricing.

Combine with cost-efficient routing

Switchyard and cost-efficient routing can run together. Switchyard first chooses the capable or efficient model tier, then cost-efficient routing orders the eligible standard and Flex service-tier variants within that tier.

This composition applies when cost-efficient routing is enabled for the account, at least one model tier is Flex-eligible, and the request leaves allow_flex_tier unset and service_tier omitted, null, or "auto". An explicit allow_flex_tier setting or non-neutral service_tier preserves the request's service-tier intent instead.

See where a request routed

The request detail in Logs shows which model and service tier handled each request. Supported CLI status lines also indicate when Switchyard is enabled.