Skip to main content
Foundation models are large neural networks trained on extremely large and diverse datasets. They provide a reusable base that can be adapted to many downstream tasks. The high-level workflow for building and applying foundation models:
  • Collect massive amounts of raw data from web crawls, books, code repositories, images, and other sources.
  • Clean and preprocess this data to produce high-quality training corpora.
  • Pre-train models at scale using high-performance accelerators (GPUs or TPUs).
  • Adapt the pre-trained model for specific tasks via fine-tuning, prompt engineering, or few-/zero-shot techniques.
  • Deploy the adapted model to production for inference in real-world applications.
The image is a diagram titled "Foundation Models – Introduction" showing a sequential process involving massive data collection, preprocessing, and pretraining on GPUs/TPUs. It illustrates the initial steps of building foundation models.
Common real-world applications powered by foundation models:
  • Conversational AI and chatbots that generate human-like responses.
  • Text summarization to extract key insights from long documents.
  • Computer vision tasks: image classification, object detection, and multimodal understanding.
  • Semantic search that captures intent and context rather than keyword matches.
  • Developer tools for code generation, completion, and debugging assistance.
Key characteristics of foundation models:
  • Fine-tunable: Can be specialized for tasks such as summarization, translation, and image captioning.
  • Large scale: Trained on data sources like Common Crawl, book corpora, code, and images; model sizes range from billions to trillions of parameters.
  • High compute requirements: Training and many inference workloads benefit from GPU/TPU acceleration.
  • Transfer learning: Support few-shot and zero-shot learning to accelerate adaptation to new tasks.
  • Multimodal capabilities: Some models support text, image, audio, and video modalities enabling cross-modal generation and understanding.
The image lists characteristics of foundation models, highlighting features such as their massive scale, pretraining and fine-tuning capabilities, multimodal support, and role as a base for downstream models.
Popular foundation-model families (proprietary and open-weight): Many of these models are available through managed platforms such as Amazon Bedrock and integrated services such as SageMaker JumpStart.

​
SageMaker JumpStart overview

SageMaker JumpStart is a curated model and solution hub inside Amazon SageMaker that accelerates prototyping by providing pre-trained models, solution templates, and example notebooks. JumpStart supports both open-source and commercial models and lets you evaluate and deploy models quickly using the SageMaker SDK or Studio UI.
The image shows the SageMaker JumpStart interface displaying a selection of AI model providers like Introduction to OpenAI, Meta, Hugging Face, and AWS, each with options to view models.
Primary benefits of JumpStart:
  • Low-friction experimentation: no heavy setup required to start evaluating models.
  • Seamless SageMaker Studio integration for interactive development.
  • Simplified endpoint deployment for real-time inference.
  • Fine-tuning support for customizing models to your data and tasks.
  • Ready-made templates and solutions for common ML problems (fraud detection, churn prediction, sentiment analysis, etc.).
The image outlines the key benefits of JumpStart: no setup required, hosted in SageMaker Studio, easy model deployment as endpoints, and support for fine-tuning select models.

​
Browsing, selecting, and deploying models

From the JumpStart dashboard you can browse foundation-model listings, inspect model cards (which describe inputs/outputs, cost, and resource needs), and use example notebooks. When you choose a model you can deploy it to a SageMaker inference endpoint directly from Studio. The UI typically exposes settings such as endpoint name, instance type, and inference configuration (UI details may change over time).
The image shows a screenshot from SageMaker Studio, depicting the process of deploying a model to an endpoint, with settings for endpoint name, instance type, and inference type displayed.

​
Example: load, deploy, and invoke a JumpStart model

A typical workflow using the SageMaker JumpStart SDK looks like this:
Notes about the example:
  • Ensure the model_id exists in your AWS region and account — JumpStart model identifiers vary by region.
  • deploy() supports deployment options such as instance_type and initial_instance_count.
  • The predict() request schema can vary by model family (text-generation vs. text-embedding vs. vision models); always consult the model card.
Before deploying a model, verify available model_ids in your account/region and read the JumpStart model card to confirm input/output schema, latency expectations, and required compute resources.

​
SageMaker JumpStart vs Amazon Bedrock

Use cases and trade-offs between the two services:

​
Best practices and cost considerations

  • Choose appropriate instance types and initial instance counts based on expected throughput and latency.
  • Monitor endpoint usage and set autoscaling policies where applicable.
  • Delete unused endpoints and snapshot large models when idle to avoid ongoing costs.
  • Review model cards for memory and GPU requirements; some models need large GPUs (e.g., p4d, p5) or multiple GPUs.
Large foundation models can incur significant compute and storage costs during deployment and inference. Monitor usage, choose cost-appropriate instance types, and remove unused endpoints to prevent unexpected charges.

​
References and further reading

Use these links and the model cards within JumpStart to evaluate models, understand input/output formats, and plan resource requirements before deploying to production.

Watch Video