Choosing foundation models on Bedrock
Lesson 4 of 5 in Choosing a Modeling Approach.
When the task is generative or open-ended language/vision work, the build decision becomes a selection decision: Amazon Bedrock serves foundation models from multiple providers behind one API. The exam does not ask you to rank vendors — it asks whether you can walk the selection dimensions for a scenario: modality, context length, the quality/latency/cost tier, customization support, and availability/governance constraints.
| Dimension | The question to ask | Why it decides |
|---|---|---|
Modality | Text only? Images in or out? Embeddings? Speech? | A text-only model is disqualified the moment the scenario says "screenshots" or "product photos." |
Context length | How much must fit in one request — a tweet or a 200-page contract plus retrieved chunks? | Long documents and RAG payloads need a window that holds them; chunking only stretches so far. |
Quality vs latency vs cost tier | Does the task need frontier reasoning, or is it high-volume and simple? | Model families ship small/fast/cheap through large/slow/strong tiers — match the tier to the task. |
Customization support | Will you need fine-tuning, continued pre-training, or distillation on this model? | Customization options vary by model — a hard requirement narrows the menu. |
Availability & governance | Is the model in your Region? Do provider terms and data-handling fit your compliance needs? | A perfect model you cannot run where your data lives is not a candidate. |
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.