Amazon SageMaker JumpStart
A model hub for deploying and fine-tuning pretrained models in your account.
❓ What is it?
A catalogue of hundreds of open-weight and proprietary foundation and task models (Llama, Mistral, Falcon, Stable Diffusion, plus classic vision/NLP models) that deploy to SageMaker endpoints in a few clicks, with notebooks and fine-tuning recipes included.
💡 Why does it exist?
Bedrock offers models as a shared API; some teams instead need the model INSIDE their own VPC — for data isolation, latency control, or weight-level customisation. JumpStart gives that control without hunting model zoos and writing deployment code from scratch.
⏱️ When should you use it?
Choose JumpStart when you want open-weight models hosted on infrastructure you control, or need to fine-tune with full access to training code. Choose Bedrock when serverless per-token pricing and zero infrastructure win.
🗺️ Where does it fit?
Inside SageMaker Studio: models deploy to real-time endpoints on instances in your account, so cost is per instance-hour, and network isolation / VPC placement is yours to configure.
🔌 How do you integrate it?
Open JumpStart in SageMaker Studio, pick a model card, deploy to an endpoint or launch the bundled fine-tuning job with your S3 dataset, then invoke the endpoint via the SageMaker runtime SDK.
🧩 Commonly integrated with
🎯 Exam angle (AIF-C01)
- JumpStart = models running on instances YOU pay for hourly; Bedrock = serverless token pricing. Cost-model questions hinge on this.
- It is the fastest path on the exam for "deploy an open-source foundation model with minimal effort".