Where Models Come From

Before a team can use a model, it has to get one. It can use a ready-made model from a provider, start from an open-source pre-trained model, or train its own from scratch. The exam asks you to match each source to a situation, and to know which AWS service offers which kind of model.

Pre-trained Models

A pre-trained model has already been trained by someone else on a large dataset, so it can be used straight away or adapted. Large pre-trained models that handle many different tasks, such as writing, summarising or answering questions, are called foundation models (FMs). Training one from nothing takes huge amounts of data, specialist skill, and expensive computing time. Starting from a pre-trained model lets a business skip almost all of that cost.

Open-Source and Provider Models

Pre-trained models come in two broad kinds. Open-source (or open-weight) models are published so anyone can download them, inspect them, run them on their own infrastructure, and change them, within the terms of their licence. Provider models are offered by a company through its service, and you use them without seeing or hosting the model itself. Open-source models give more control and transparency. Provider models take less effort to run. Either way, check the licence: it decides whether commercial use is allowed.

Where AWS Offers Models

• Amazon Bedrock gives access to foundation models from Amazon and from other AI companies through a single, fully managed API. There are no servers to run.
• Amazon SageMaker JumpStart is a hub of pre-trained models, including many popular open-source ones, that you can deploy to your own SageMaker AI endpoint or fine-tune on your data with a few clicks.

Both let a team try several models quickly before committing to one. A sensible pattern is to start with a managed model in Bedrock to prove the idea, then move to JumpStart only if the project needs an open-source model, its own hosting, or deeper changes than Bedrock allows.

Training a Custom Model

Sometimes no existing model fits, for example a niche prediction on a company's own data, such as forecasting failures of one type of factory machine. Then a team trains a custom model using its own labelled data, for instance with SageMaker AI training jobs. This gives the most control, but it costs the most time, data and skill. For general language or image tasks, adapting a pre-trained model is usually far cheaper than building one.

On the Exam

If a scenario stresses "least effort" or "no infrastructure to manage", look for Bedrock. If it stresses running or modifying an open-source model in the company's own environment, look for JumpStart.