Amazon Bedrock Foundations and Model Selection

Amazon Bedrock is the AWS-managed gateway to foundation models, and it is one of the most heavily tested services on the AWS Certified Machine Learning Engineer – Associate exam. Candidates who understand how Bedrock provides a unified API for invoking models from different providers, how an application selects the target model using the modelId parameter, how the Converse API abstracts model-specific details, and how to select the right model based on cost, latency, and quality trade-offs will be well positioned to answer a significant portion of the generative AI questions. This lesson walks through the Bedrock architecture, the available model families, the unified API layer, and a practical decision framework that maps directly to exam scenarios.

The lesson covers three areas. First, you will learn which foundation model families are available and what each excels at. Second, you will understand how the Converse API decouples your application from any single provider. Third, you will develop a structured approach to model selection trade-offs.

Attention: A common exam mistake is choosing SageMaker-based custom training or self-hosted inference when the scenario describes a standard generative AI task. Bedrock is typically the simplest and most cost-effective answer for standard generative AI tasks unless the question explicitly requires custom model training, unsupported model types, or full infrastructure control.

Foundation models available in Bedrock

Amazon Bedrock provides access to foundation models (FMs)Large-scale neural networks pretrained on broad datasets that can be adapted to a wide range of downstream tasks without training from scratch. from multiple providers, each optimized for different task categories. Understanding these families is the first step in any model selection decision because the exam expects you to map a business requirement to the correct model category before you evaluate cost or latency.

The key model families and their strengths break down as follows:

• Amazon Nova (text, chat, and reasoning): AWS’s flagship multimodal models. They are tiered by capability (for example, Micro and Lite for fast, low-cost tasks; Pro and Premier for complex reasoning, math, and agentic workflows).

• Amazon Nova (media and speech): Domain-specific models for rich media. This includes Nova Canvas (image generation), Nova Reel (video generation), and Nova Sonic (real-time speech recognition and generation).

• Amazon Embeddings ...