Skip to lesson

AWSAWS Certified Generative AI Developer - Professional2026-100 Prerequisites

Classify models by the required output

Completed by the knowledge check

Learning objective

Distinguish foundation models, LLMs, and multimodal models.

Core idea

For an architect choosing a model, the first question is what transformation the application requires. A foundation model (FM), a large language model (LLM), and a multimodal model are overlapping descriptions, not three mutually exclusive products. Foundation model describes a broadly trained model you can build on. Language model describes its language capability. Multimodal describes its ability to work across types of content.

A model can therefore be both a language model and a multimodal foundation model. It might accept an image alongside a question and return text. That does not mean it can create an image. Conversely, an image-generation foundation model is not a substitute for a conversational language model. Classify the required input and output separately before comparing model quality. This lesson establishes the categories; see 1.2.1 for the full capability shortlist.

How it works

Foundation models describe the reusable base

Amazon Bedrock defines a foundation model as a model with many parameters trained on a large, diverse dataset. Its glossary includes text generation, image generation, and conversion of inputs into embeddings among foundation-model capabilities. A single model need not support all those transformations. The category describes the broad base from which applications use specific capabilities.

Bedrock provides access to foundation models through managed APIs. The service is the access and application platform; it is not one model with one capability set. Selecting a Bedrock model still means selecting a particular provider, model, and version. You cannot infer that every model supports an interface or modality because another model on the same platform does.

Also distinguish inference from customization. Inference produces an output from an input. Customization changes model parameter values through a training process. Changing an inference setting, such as temperature, does not create a custom model. You usually start by checking whether an existing base model can perform the task before deciding that training is necessary.

Language models transform language

An LLM supports language tasks such as producing an answer, drafting text, summarizing a passage, or generating code. These tasks share a language interface, but success on one does not establish success on another. A fluent summary is not evidence that a generated program passes its tests. Evaluate the particular transformation the application needs.

Amazon Nova Micro provides a concrete text-only example: its documented input and output modalities are text. For a workflow that receives already extracted text and returns a prose summary, a text-only model can be an eligible candidate. If the source includes a diagram whose layout carries meaning, passing only a text transcription changes the task. The model cannot inspect an image that the application never supplies through a supported interface.

The boundary also separates language generation from embedding generation. Amazon Titan Text Embeddings V2 takes text and returns a numerical vector. It is a foundation model used for retrieval and similarity tasks, rather than an answer writer. A search application can use it alongside an LLM, with each serving a different stage. See 0.4 for the vector representation.

Multimodal models have directional capabilities

A modality is a type of information, such as text, image, video, or speech. Multimodal capability must be described as an input-to-output contract. Amazon Nova Lite, Pro, and Premier support text, image, and video inputs while producing text outputs. Amazon Nova Canvas takes text and image inputs for image generation. Amazon Nova Reel takes text and image prompts for video generation. These are materially different transformations even though all involve multiple modalities.

A question about a recorded demonstration and a request to create a demonstration video require different model roles. For the first, the application needs video understanding and text output. For the second, it needs video generation. Adding “generate a video” to a prompt cannot provide an output modality that an understanding model lacks.

Use the AWS model documentation to check the exact contract. Bedrock’s FoundationModelDetails exposes inputModalities, outputModalities, and responseStreamingSupported for model discovery. Treat these as useful metadata, then consult the selected model’s documentation for richer capabilities and constraints. A high-level modality label does not specify accepted file formats, attachment rules, or every supported task.

Key distinctions and trade-offs

Model category versus model capability: A category narrows the search. The version’s documented capability determines eligibility. “Foundation model” does not tell you whether the output is prose, a vector, an image, or a video.

Multimodal input versus multimodal output: A model that reads images and answers in text can serve visual question answering. It cannot replace an image-generation model solely because both are multimodal.

One model versus a composed application: A workflow can combine an embedding model, a language model, and a creative model. This gives each stage the needed output but creates separate interfaces and validation responsibilities. Choose composition because the stages require different transformations, not because more models imply better quality.

Model versus agent: Bedrock describes an agent as an application that orchestrates foundation models and other resources. The model generates outputs; the application decides how to use them. An agent label does not add a missing model modality or prove that a requested business action completed.

Architecture / decision scenario

A product team proposes one “AI model” for a support portal with three functions: answer questions about manuals, explain a photo of damaged equipment, and create an illustration for a training article.

Separate the contracts. The manual search needs text embeddings to locate related passages, then a language model to compose an answer from those passages. Titan Text Embeddings V2 can be evaluated for the vector stage, but its vectors are not customer-facing answers. The equipment function needs image input and text output; a Nova multimodal understanding model is a candidate. The illustration function needs image output; Nova Canvas is a different candidate.

The team records required input and output modalities for each stage and checks the exact model documentation before evaluating sample outputs. It rejects the proposal to use Nova Micro for photo interpretation and the proposal to use a text-output understanding model for illustration generation. It then applies the quality and operating gates in 1.2.1 rather than treating category eligibility as a final selection.

Common exam traps

  • Treating FM, LLM, and multimodal model as disjoint categories.
  • Assuming every foundation model generates conversational text.
  • Confusing image understanding with image generation.
  • Treating an embedding vector as a generated answer.
  • Assuming all models on Bedrock share the same modalities and API features.
  • Treating an agent label as proof that a business action completed.

What to remember

Start with the required transformation. Foundation model describes a broadly trained base; language and multimodal labels describe capabilities that can overlap. Verify input and output modalities separately for the exact model version. Compose models when the workflow requires different representations, and evaluate each stage against its own acceptance criteria.

Knowledge check

References