Exam-style
A media company has settled on a specific open-weight summarisation model. The published weights are good enough as they are, so nothing needs training or fine-tuning. What the team does need is the model running inside its own VPC, on GPU instance types it chooses and can resize as demand becomes clearer. Which of these gets them there?
Reveal the answer
B. Deploy the model from SageMaker JumpStart onto an Amazon SageMaker AI endpoint in the company's own account
Three requirements sit in that scenario: a particular set of published weights, a VPC those weights have to run inside, and an instance type the team names. Nothing asks for training. SageMaker JumpStart is a catalogue of pre-trained models, and it deploys the one you choose onto an Amazon SageMaker AI endpoint in your own account. You set the instance type, the instance count and the VPC on that endpoint, so all three requirements are covered. Invoking the same model on demand through the Amazon Bedrock API gets the model right and the infrastructure wrong. That path is serverless: no instance type to set, no VPC placement, and no endpoint of yours for the traffic to reach. Building the model in Amazon SageMaker AI answers a requirement nobody has, since it runs a training job to arrive at weights the company already holds. Amazon Quick is a finished AI assistant for people asking questions of company content, not a place to host a model you have chosen. Kiro is an AI development environment, so it helps write the service around the model; it does not run the model either.
Q. A named open-weight model, no training needed, and it has to run in your VPC on instances you choose. Which of the five?
A. SageMaker JumpStart, deploying the pre-trained model onto an Amazon SageMaker AI endpoint in your own account, where the instance type and the network placement are yours.
Why? Three names, three jobs. Amazon Bedrock when you want an API and no infrastructure of your own. SageMaker JumpStart onto Amazon SageMaker AI when you want the weights on infrastructure you choose. Amazon SageMaker AI on its own when you are training or building the model yourself. Amazon Bedrock Marketplace is the exception to that first line. It deploys a catalogue model onto a SageMaker AI endpoint where you set the instance type, the count and the VPC. The on-demand Bedrock API does not. Amazon Quick and Kiro sit outside that line entirely: one is a finished assistant over company content, the other helps developers write code, and neither hosts a model you have picked.