SageMaker JumpStart adds model hub and preconfigured endpoint deployment

SageMaker JumpStart made two improvements in 2026. First, the Model Hub area of the SageMaker AI console lets you browse the JumpStart catalog, view sample notebooks, and open them in SageMaker Studio. Second, it offers preconfigured endpoint deployment options with named configurations prioritized for low cost, low latency, or general-purpose workloads, covering many generative AI models including Neuron-optimized configurations. A new deep-link integration also lets you discover a model on Hugging Face and jump straight to deploying it in SageMaker AI. The article presents a reusable three-step evaluation flow based on SageMaker Python SDK v3: filter models by task or name with list_jumpstart_models, deploy the recommended instance type with ModelBuilder and JumpStartConfig, then use retrieve_all_examples to get official sample payloads, call the endpoint per record while measuring latency and per-request generation rate, and finally delete the endpoint. In testing, after deploying huggingface-vlm-qwen3-5-27b, a request asking "what is deep learning" returned about 153 words in 9.07 seconds, roughly 16.9 words per second per request. This update directly affects ML engineers and algorithm teams evaluating and shipping generative models on AWS: preconfigured endpoints cut the work of choosing instances and tuning configs, and one-click deployment from Hugging Face lowers the cost of moving models across platforms. Note that payload formats, required instance types, end-user license agreements, and response structures can still differ between models, so evaluation scripts need per-model adaptation; deployment time also varies with model, instance type, quota, and region.