Cloud concept · Updated

Foundation Models - Foundation models are models pretrained on broad datasets that can be adapted to many downstream tasks rather than trained separately for each task.

In short

Foundation models provide reusable capabilities for working with language, images, audio, and other data through prompting, retrieval, or additional training. Choosing between a foundation model and traditional task-specific machine learning depends on the workload, available data, explainability requirements, regulatory obligations, and operational constraints.

Cloud Concepts Hub →

6 key aspects of Foundation Models

Foundation models offer reusable capabilities, but effective adoption requires deliberate adaptation, workload-based selection, governance, evaluation, and operational controls.

01
Broad pretraining and reuse

A foundation model learns reusable patterns from broad training data and supports multiple downstream tasks. This can reduce the need to build a separate model for each application, but suitability still requires task-specific evaluation.

02
Multiple adaptation methods

Prompting, retrieval-augmented generation, and fine-tuning offer different ways to adapt model behavior. Use retrieval to supply relevant external information and consider fine-tuning when measured results justify changing task behavior or specialization.

03
Workload-based model selection

Foundation models often fit flexible language, image, and multimodal tasks, while traditional ML can be more efficient for narrowly defined structured-data predictions. Compare both against a representative test set and explicit business requirements.

04
Governance and explainability

Evaluate privacy, data location, licensing, auditability, and explanation requirements before selecting a model and hosting arrangement. Fluent explanations and cited sources do not automatically provide faithful reasoning or regulatory compliance.

05
Operational trade-offs

Model size, input length, output length, serving architecture, and traffic patterns affect cost and performance. Benchmark latency and throughput under realistic load, including the overhead of retrieval and output validation.

06
Probabilistic outputs and evaluation

Generated outputs may be incorrect, inconsistent, or unsafe even when they appear confident. Maintain evaluation datasets, test security boundaries, and require validation or human approval where errors have significant consequences.

Cloud services for Foundation Models

Amazon Bedrock provides managed foundation-model inference, while Amazon SageMaker AI supports workflows requiring more control over model training and deployment and enables comparison with traditional ML approaches. Amazon S3 can hold governed documents, datasets, and artifacts. In a retrieval-augmented application, an ingestion process creates searchable representations of documents in Amazon OpenSearch Service; the application retrieves relevant passages and includes them in a model request. These components are optional architectural choices rather than mandatory dependencies. Evaluate the complete workflow for retrieval quality, authorization, data handling, output accuracy, latency, and cost before production deployment.

Amazon Bedrock

Amazon Bedrock provides managed access to foundation models through APIs, with supported capabilities for customization, retrieval-based applications, and guardrails. It reduces model-serving infrastructure work, but teams must still evaluate model suitability, data handling, and application behavior.

Amazon S3Amazon S3

Amazon S3 can store source documents, training datasets, evaluation datasets, and model artifacts. Access policies, encryption, and lifecycle controls help govern these assets, but storing data in S3 does not automatically make it available to a foundation model.

Amazon OpenSearch ServiceAmazon OpenSearch Service

Amazon OpenSearch Service supports vector and keyword search for retrieval-augmented generation architectures. Applications can retrieve relevant document passages and supply them to a foundation model, while enforcing appropriate authorization over retrieved content.

Foundation Models exam questions

Real practice exam questions on Foundation Models, with the correct answer and its explanation. Answer them with every option in the exam simulator.

AIF-C01

A procurement company wants to help employees compare an original contract clause with a supplier's proposed revision. The application must explain changes in obligations in plain language, including changes expressed through paraphrasing rather than matching words. Clause wording and requested explanations vary widely. The company has no labeled examples and wants an initial solution without task-specific model training. Employees will verify every explanation before using it. Which approach is MOST appropriate?

Show answerCORRECT ANSWER

Prompt a pretrained language foundation model with both clause versions and instructions to explain substantive changes, and evaluate its explanations on representative examples.

A language foundation model is appropriate for interpreting varied text and generating flexible explanations without task-specific training. Representative evaluation and employee verification remain necessary because the model can misinterpret clauses or generate unsupported explanations.

MLA-C02 (beta)

A company is building an application that compares vendor proposals containing text, engineering diagrams, and photographs. The application will use an existing foundation model through Amazon Bedrock without fine-tuning or operating model-serving infrastructure.

The approved workflow requires the model to examine the original images and all proposal text together in one request. Converting images to text or splitting a proposal across requests is not permitted. The largest supported request contains 30,000 input tokens, including the selected model's image-token accounting, and must allow up to 4,096 generated output tokens.

All candidate models are available through on-demand inference in the required AWS Region. Their documented context windows include both input and output tokens. Which model deployment choice meets these requirements?

Show answerCORRECT ANSWER

Use a text-generation model that accepts text and image inputs, has a 65,536-token context window, and supports up to 8,192 output tokens. Set the maximum output length to 4,096 tokens.

This model natively accepts the required text and images. Its context window accommodates the 34,096-token combined budget, and its output limit permits the required 4,096 tokens. Bedrock on-demand inference also avoids operating model-serving infrastructure.

AIP-C01

An engineering company is building an agent that answers questions requiring evidence from both equipment manuals and contractual service obligations. In Amazon Bedrock evaluations, one embedding foundation model provides the best retrieval quality for technical terminology, while another provides the best retrieval quality for legal terminology. A separate text-generation foundation model synthesizes the retrieved evidence.

The two embedding models produce vectors with the same number of dimensions. However, they were trained independently, and their similarity scores are not calibrated against each other. Each model meets the company's retrieval targets when queries and documents are embedded with that same model. The company must retain both retrieval capabilities and combine their results without assuming that the models share a vector space or comparable similarity scores.

Which coordination design meets these requirements?

Show answerCORRECT ANSWER

Maintain separate document indexes using the respective embedding models. Embed each query with both models, search each corresponding index, and combine the result rankings using reciprocal rank fusion with parameters validated on representative questions. Supply the retrieved passages and source identifiers to the text-generation model.

Each retrieval path preserves the query-document embedding compatibility demonstrated during evaluation. Rank fusion combines retrieval results without treating independently trained vectors or their raw similarity scores as directly comparable. The generation model receives readable evidence and source identifiers rather than embeddings.

Practice Foundation Models in a full exam

Questions on Foundation Models in each practice exam, with every answer option and an explanation for each:

Start a free practice exam

What is Foundation Models

3 min read

A foundation model is a broadly pretrained model that provides a starting point for multiple applications. Pretraining often uses self-supervised learning on large datasets, allowing the model to learn reusable patterns without requiring manually labeled examples for every downstream task. Applications then use those capabilities directly or adapt them to a specific domain. Large language models are one category of foundation model, but foundation models also support vision, audio, and multimodal workloads. Not every foundation model is a conversational system or generates content; some produce representations such as embeddings.

Foundation models are useful when a workload involves diverse, unstructured inputs or flexible outputs. Examples include summarizing support conversations, drafting product descriptions, answering questions about technical documentation, and extracting information from documents with varied layouts. Prompting can establish an initial solution without training a separate model. However, having no task-specific training requirement does not eliminate the need for representative evaluation data, security controls, and production monitoring.

Traditional machine learning is often a better starting point for a narrowly defined prediction problem with structured data and a measurable target. For example, gradient-boosted trees may predict equipment failure from sensor measurements more economically than a large language model. A logistic regression model may suit a credit-risk workflow requiring understandable relationships between input variables and predictions, subject to validation and applicable rules. Traditional ML is not automatically explainable: complex ensembles and neural networks can also be difficult to interpret.

In the cloud, engineers commonly access foundation models through managed inference APIs or deploy model weights on managed or self-operated infrastructure. Managed APIs reduce infrastructure work, while self-hosting provides more control over model versions, runtime configuration, and deployment boundaries. That control also introduces responsibility for capacity planning, accelerator availability, patching, scaling, and model serving. Compare measured latency, throughput, availability, and total cost rather than assuming either approach is always cheaper.

Adaptation methods solve different problems. Prompt engineering supplies instructions and examples without changing model weights. Retrieval-augmented generation retrieves relevant information and adds it to the model's context, making it useful for questions about current internal documentation. Fine-tuning changes model parameters using additional training data and can improve task behavior, style, or domain performance. Fine-tuning is not a reliable substitute for retrieving frequently changing facts, and retrieval does not guarantee that every generated statement is correct.

Regulatory and explainability requirements can determine whether a foundation model is appropriate. Review data residency, processing locations, retention, provider data-use terms, access controls, licensing, and audit requirements before sending sensitive information to an endpoint. Neither managed hosting nor private deployment establishes compliance by itself. A model-generated explanation is not necessarily a faithful explanation of its internal reasoning, and a citation does not prove that the cited material supports the answer. High-impact decisions may require a more interpretable model, human review, or a different workflow.

Operational constraints also matter. Foundation-model applications can face variable response times, context-window limits, inference charges, hallucinations, and prompt-injection attacks. Evaluate realistic inputs, including adversarial cases, and validate outputs before allowing them to trigger actions. A practical hybrid design might use traditional ML to estimate customer churn from account metrics and a foundation model to draft an outreach message. The best choice is the simplest approach that meets accuracy, governance, latency, and cost requirements, not necessarily the largest available model.

Learn Foundation Models in lessons

7 short lessons on Foundation Models in 4 certification courses.

and 1 more lesson

Open the lessons