There's no single best generative AI model, only the best fit for a job. Choosing one means weighing the model type against the task, then performance and capabilities, then constraints such as cost, latency, model complexity and compliance. The exam gives you a scenario with one deciding constraint and asks which model or approach fits, so learn what each factor changes.
Start From the Task
The first filter is model type, also called modality: the kind of data a model takes in and produces.
- Text models (LLMs) read and write language: summaries, answers, drafts, code.
- Image models, such as diffusion models, create or edit images from text descriptions.
- Embedding models turn text or images into lists of numbers called vectors, used for similarity search and recommendations. They don't write answers.
- Multimodal models accept more than one type of input, such as an image plus a question about it.
A model that can't handle the task's data type is ruled out before anything else.
Performance and Capabilities
Next, check whether the model can do the job well enough. Useful questions are: how good are its answers on your own examples, which languages does it handle, and how large is its context window, the amount of text it can consider in one request? A long contract needs a large context window. You should also check whether it can be customized, for example by fine-tuning, if you expect to need that.
Amazon Bedrock's model evaluation feature lets you compare candidate models on your own prompts, using automatic metrics, human reviewers or another model acting as a judge.
Cost, Latency and Complexity
Larger, more complex models, those with more parameters, tend to give richer answers but cost more per request and respond more slowly. Smaller models are cheaper and faster, and often good enough for simple, high-volume tasks. Latency, the time to respond, matters most where a person is waiting, such as a voice assistant. Pick the smallest model that meets your quality bar, not the biggest one available.
Compliance and Data
Regulated businesses must also check:
- Regional availability: is the model offered in the AWS Region where the data must stay?
- Data use: with Amazon Bedrock, your prompts and responses aren't used to improve the base models and aren't shared with model providers.
- Licensing and terms: each model has its provider's terms of use.
A Quick Selection Guide
| Requirement in the scenario | Points toward |
|---|---|
| Similarity search or recommendations | An embedding model |
| Very long documents in one request | A large context window |
| Instant replies while users wait | A smaller, faster model |
| Tight budget, simple high-volume task | The smallest model that meets quality |
| Data must stay in one country | A model available in that Region |
| Unsure which model is best | Compare candidates with model evaluation |
On the Exam
The trap is assuming bigger is always better. If a scenario stresses cost, speed or simple tasks, the answer is usually a smaller model that meets the requirement. If it stresses compliance, look for Region availability and data handling, not model size.