3.1.3. Foundation Models: Amazon Bedrock and SageMaker AI JumpStart
💡 First Principle: Foundation models encode broad knowledge from massive datasets. Instead of training from scratch (expensive, data-hungry), you start with a foundation model and adapt it to your specific task—like hiring an experienced chef and teaching them your restaurant's recipes instead of training someone from culinary school.
Amazon Bedrock provides API access to foundation models from many providers (Amazon, Anthropic, Meta, Mistral AI, Cohere, OpenAI and others) without managing infrastructure. You call models directly for text generation, summarization, question answering, code generation and multimodal tasks, and you can swap one model for another without rewriting application code. When prompting alone is not enough, Bedrock offers retrieval (Knowledge Bases) and model customization — supervised fine-tuning, distillation and reinforcement fine-tuning. How to choose among these is covered in 3.4.
Amazon SageMaker AI JumpStart provides pre-trained open-weight models (and solution templates) that you deploy or fine-tune inside SageMaker AI Studio. Unlike Bedrock, JumpStart models run on SageMaker AI instances and endpoints that you choose, size and operate. JumpStart fine-tuning supports domain adaptation (continued pre-training on unlabeled domain text) and instruction-based fine-tuning (prompt and response examples), including parameter-efficient LoRA settings.
| Feature | Amazon Bedrock | SageMaker AI JumpStart |
|---|---|---|
| Hosting | Serverless API managed by AWS | SageMaker AI endpoint you size and operate |
| Models | Curated models from many providers | Open-weight and AWS models, plus solution templates |
| Customization | Supervised fine-tuning, distillation, reinforcement fine-tuning | Domain adaptation (continued pre-training), instruction fine-tuning, LoRA |
| Cost model | Pay per token (on-demand), or Provisioned Throughput | Pay for instance time while the endpoint runs |
| Best for | Fast access to generative AI with no infrastructure | Full control of hosting, instance choice or model internals |
Choosing between Bedrock and JumpStart:
If the question says "add text generation with no infrastructure management," the answer is Bedrock. If it says "deploy an open-weight model with full control over the instance type and endpoint configuration," the answer is JumpStart. The distinguishing signals are who operates the infrastructure and how much control the team needs. A model customized outside Bedrock — for example an open-weight model fine-tuned on SageMaker AI — can also be brought into Bedrock with Custom Model Import, so the two are not mutually exclusive.
⚠️ Exam Trap: Bedrock and JumpStart are not interchangeable. Bedrock is the serverless path for generative AI workloads (text, conversation, code, multimodal). JumpStart is the path when you need SageMaker AI's training and hosting control, including for non-generative models such as vision classifiers. A question about classifying medical images with a model the team controls points to JumpStart or SageMaker AI training, not a Bedrock text model.
Reflection Question: A company wants a customer support assistant that answers from its product documentation with no servers to manage. Should it use Bedrock or JumpStart? What changes if the security team requires the model to run on instances the company sizes and patches itself?
