ML Model and Foundation Model Development

Task and modality fit (text, image, multimodal, embeddings), quality on your own evaluation data, context window, latency, price per token, language support, and whether the model supports the customization you need.

The maximum number of tokens (input plus output) the model can handle in one request; long documents, long chats, or large RAG contexts need a larger window.

Amazon Nova (for example, Nova Micro, Lite, and Pro for understanding, plus Nova Canvas and Nova Reel for image and video generation).

Smaller models cost less per token and respond faster; if they meet the quality bar on your evaluation set (for example, classification or extraction), they are the better choice.

Build a representative evaluation dataset and run Bedrock model evaluation jobs (automatic metrics, LLM-as-a-judge, or human review) across the candidates, then weigh quality against cost and latency.

An embedding model, such as Amazon Titan Text Embeddings V2 or Cohere Embed, not a text-generation model.