[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"glossary-model-cascade::en":3,"gloss-cluster-model-cascade::en":26,"gloss-next-model-cascade::en":9},{"slug":4,"category":5,"name":6,"definition":7,"meta_desc":8,"faq":9,"schema_markup":9,"related":10},"model-cascade","core-ai","Model Cascade","A model cascade is an inference pattern in which a request is first handled by a small, fast, inexpensive model, and passed to a larger one only when the cheap attempt is judged inadequate. It rests on a distribution most workloads share: the majority of requests are easy, a minority are hard, and paying premium rates for all of them means overpaying for most of them.\n\nThe design question is entirely about the escalation signal — how the system decides that the first answer is not good enough. Options include the model's own expressed uncertainty, a validator that checks the output against a schema or a rule, a verifier model that scores the answer, agreement between two cheap attempts, or a heuristic on the input such as length, language or topic. Each is imperfect, and the quality of the cascade is bounded by how well this decision is made rather than by either model.\n\nTwo failure modes define the trade-off. Escalating too rarely lets weak answers through, which is worse than never having used the cheap model, because a wrong answer that looks confident costs more than a slow one. Escalating too often incurs the cost of both models on the same request, which can end up more expensive than sending everything to the large model directly. Measuring the escalation rate and the cost per request together is the only way to know which side you are on.\n\nLatency behaves differently from cost here. A cascade adds the cheap model's latency to every escalated request, so the tail of the distribution gets slower even as the average gets cheaper. For interactive features that matters, and it is often worth routing obviously hard requests straight to the large model on input signals rather than making them wait for a failed cheap attempt.\n\nCascades work best where correctness is checkable. Structured extraction, classification into a fixed set, code that must compile, and answers that must cite a retrieved source all provide a cheap objective test that can drive escalation. Open-ended generation, where quality is a matter of judgement, is much harder to cascade well, because the only reliable judge may be as expensive as the large model itself.\n\nOperationally, treat the cascade as one versioned configuration: both models, their versions, the escalation rule and its thresholds. Evaluate the whole thing as a unit against a fixed set of cases, and monitor the escalation rate in production, because it drifts as inputs change and is usually the first sign that either the cheap model or the threshold has stopped being appropriate.","A model cascade tries a cheap model first and escalates only when the answer falls short, buying most of the quality of a large model at a fraction of the cost.",null,[11,14,17,20,23],{"slug":12,"name":13},"confidence-threshold","Confidence Threshold",{"slug":15,"name":16},"fallback-model","Fallback Model",{"slug":18,"name":19},"inference-cost","Inference Cost",{"slug":21,"name":22},"model-router","Model Router",{"slug":24,"name":25},"small-language-model","Small Language Model (SLM)",[27,31,35,39,42,45,48,51,54,57,60,63],{"slug":28,"category":5,"name":29,"updated_at":30},"agentic","Agentic AI","2026-09-08T21:32:34+00:00",{"slug":32,"category":5,"name":33,"updated_at":34},"alignment-tax","Alignment Tax","2026-09-08T21:32:35+00:00",{"slug":36,"category":5,"name":37,"updated_at":38},"artificial-intelligence","Artificial Intelligence (AI)","2026-09-08T21:32:36+00:00",{"slug":40,"category":5,"name":41,"updated_at":30},"attention","Attention",{"slug":43,"category":5,"name":44,"updated_at":38},"beam-search","Beam Search",{"slug":46,"category":5,"name":47,"updated_at":34},"benchmark-contamination","Benchmark Contamination",{"slug":49,"category":5,"name":50,"updated_at":34},"catastrophic-forgetting","Catastrophic Forgetting",{"slug":52,"category":5,"name":53,"updated_at":38},"computer-vision","Computer Vision",{"slug":55,"category":5,"name":56,"updated_at":34},"constitutional-ai","Constitutional AI",{"slug":58,"category":5,"name":59,"updated_at":30},"context-window","Context Window",{"slug":61,"category":5,"name":62,"updated_at":38},"deep-learning","Deep Learning",{"slug":64,"category":5,"name":65,"updated_at":30},"diffusion-model","Diffusion Model"]