Skip to main content
In this lesson we cover how to choose a suitable foundation model from Amazon Bedrock’s model catalog. We follow a problem → solution approach: identify selection challenges, determine what you need from a model, use the Bedrock catalog to evaluate candidates, weigh trade-offs (cost, latency, capability), and finish with practical next steps. Many models appear in each AWS Region’s catalog. They differ in capability (reasoning depth, supported modalities), training data, and cost. Models are exposed by different providers and are not always available in every Region, which makes selecting the right model non-trivial.
A presentation slide titled "Problem: Organizations Need to Choose From Multiple Models Available" that lists four challenges: models vary in capabilities/performance/cost, providers expose models via different platforms, not all models are available in every region, and selecting the right model is challenging. Each point is shown with a blue icon on a dark background.
Common pitfalls
  • Teams often default to a popular large language model because it’s well-known rather than because it’s the best fit for the task.
  • Without a centralized catalog, each model requires separate integration and evaluation — slow and error-prone.
Start by being explicit about what you want the model to do. Different tasks require different input/output modalities and capabilities. Modalities, inputs, outputs, and common use cases
A presentation slide titled "Problem: You Need to Know What You Want From the Model" showing a table that compares model categories (Text/Chat, Image, Video, Multimodal) with what you send, what you get back, and example use cases. It highlights inputs like text or text+image and outputs such as generated text, images, and video for tasks like chatbots, marketing graphics, and ad generation.
Match the model to the modalities your application requires, then evaluate cost vs. performance. Model size and typical trade-offs Design with intent: different components of the same application can use different models. For example, route routine classification to a smaller model and reserve a larger model for complex reasoning tasks.
An infographic titled "Solution: Model Selection" showing two columns: Smaller Models (for tasks like sentiment analysis, summarization, Q&A, and text classification) and Larger Models (for legal argument, technical architecture, and long‑context reasoning). It also labels example models like Meta Llama 3.1 8B and 70B Instruct.
Practical considerations when selecting a model
  • Regions and availability: Choose an AWS Region near your users to reduce latency, but verify Bedrock and the specific model are available in that Region. If your preferred image or multimodal model (for example, Stable Diffusion) is not available locally, decide whether to accept extra cross-region latency or pick an available alternative.
  • Provider variety: Bedrock hosts models from multiple vendors (Meta, Anthropic, NVIDIA, Amazon, Stability AI, and others). Bedrock abstracts away infrastructure and gives you consistent access to these pre-trained foundation models.
  • Billing and throughput: Serverless (on-demand) pricing commonly bills by processed tokens for text/chat models and may bill per image or per-generation unit for image/multimodal models. Larger models usually cost more per token or per request. If you need predictable performance and throughput, consider Bedrock’s provisioned throughput options.
Real-world example: Meta’s Llama family includes smaller models (e.g., 8B Instruct) and larger variants (e.g., 70B Instruct). Training required massive GPU-hours; using Bedrock lets you leverage those trained models without managing the training pipeline. The model catalog: discovery and evaluation Bedrock’s model catalog is the primary tool to discover and evaluate models available in the selected Region. From the console you can inspect a model’s:
  • Overview and recommended use cases
  • Supported input/output modalities
  • Supported languages
  • Deployment options (serverless, provisioned throughput, Marketplace)
  • Max tokens / context window
  • Programmatic model ID — the exact identifier you supply in API requests
The model ID matters because it is the string you pass to Bedrock to select the foundation model for inference.
A presentation slide titled "Workflow: Model Catalog" showing a screenshot of a Llama 3.1 70B Instruct model page on the left and a callout on the right that reads "Determine suitability for our use case" with bullets about modality and max tokens. The slide summarizes checking a model's input/output data types and token limits.
Token and context window considerations
Tokens measure the size of data you send to and receive from a model. A model’s context window (max tokens) constrains the combined size of prompt input plus expected output. For long-document summarization or long-context reasoning, confirm the model’s max tokens and ensure input + output fit within that limit.
Use the model catalog’s pricing view to estimate per-request cost and run cost vs. performance comparisons. Select models that satisfy modality, language, and reasoning requirements while fitting your latency and cost constraints. Summary — key takeaway
A presentation slide titled "Key takeaway" with a short point: "Amazon Bedrock provides a centralized catalog of foundation models from multiple providers, allowing developers to easily discover and choose the right model for their applications."
Amazon Bedrock provides a centralized, Region-scoped model catalog so developers can discover model properties (modalities, languages, context window, deployment options, and model IDs). Use the catalog to intentionally choose the right model for each part of your application—don’t default to the largest model just because it’s popular. Next steps
  • Open the Bedrock console in the AWS Management Console and inspect models in your preferred Region.
  • Evaluate models by modality, max tokens, and pricing.
  • Prototype a routing pattern: small model for bulk tasks, large model for complex analysis.
  • If you require predictable latency or high throughput, explore Bedrock’s provisioned throughput options.
Further reading and references
  • Amazon Bedrock documentation: https://docs.aws.amazon.com/bedrock
  • Model provider pages (e.g., Meta Llama 3, Anthropic, Stability AI)
  • Best practices for prompt design and token budgeting
That concludes this lesson on selecting foundation models from the Bedrock model catalog. A follow-up demonstration shows how to access and use the Bedrock console via the AWS Management Console.

Watch Video