Models - Aug 4, 2026

Deploy a SageMaker Custom Model: Bedrock vs. Endpoint

Quick answer

AWS documents two deployment choices from a SageMaker custom model’s details page:

  • SageMaker AI Inference endpoints;
  • Amazon Bedrock Custom Model Import.

Choose only after the model passes evaluation. The short deployment page establishes that both paths exist; it does not establish that every model, training method, artifact, or Region is compatible with both.

Decision checklist

DecisionSageMaker endpointBedrock import
Current model supportVerify SageMaker inference compatibilityVerify Bedrock Custom Model Import compatibility
API and clientConfirm endpoint contract and container behaviorConfirm Bedrock import and invocation contract
CapacityDefine serving configuration and scalingConfirm supported capacity and quota model
GovernanceConnect model package, lineage, IAM, and monitoringPreserve source lineage and Bedrock governance controls
Region and dataVerify endpoint and artifact locationVerify import, storage, and inference Region
CostModel total endpoint and traffic costModel import and inference cost from live pricing

This table is a review structure, not a product support promise. Resolve every row from current AWS documentation and your account.

Deployment gate

Before either path:

  1. compare the custom and base models on an immutable holdout;
  2. review safety, privacy, bias, security, and domain regressions;
  3. pin the model, dataset, job, evaluation, and artifact lineage;
  4. register the approved model version;
  5. define least-privilege invocation and administrative roles;
  6. test latency, throughput, errors, scaling, and cost;
  7. prepare canary, monitoring, rollback, and artifact retirement.

A registered model improves traceability; it does not prove deployment readiness. A successful import or endpoint creation does not prove output quality.

Avoid coupling training and serving decisions

Serverless customization manages training infrastructure. Serving is a separate lifecycle with its own availability, scaling, authentication, logging, retention, and billing. Do not describe the training job as “serverless inference” unless the selected serving product explicitly uses that term and configuration.

Use the data and evaluation guide before approval and the cost and operations guide for end-to-end economics.

Frequently asked questions

Where can a SageMaker customized open-weight model be deployed?

AWS documents two paths from the custom model details page: a SageMaker AI inference endpoint or Amazon Bedrock Custom Model Import.

Is every custom model deployable to both targets?

Do not assume so. Verify the exact model, artifact, technique, Region, quota, and current deployment documentation for each target before choosing it.

Should deployment start immediately after training completes?

No. Evaluate the logged model on an untouched holdout, review regressions and security, register the approved version, and define rollback and monitoring before serving traffic.

Official sources

Source check: August 4, 2026. Verify the exact model and artifact support, Regions, quotas, IAM, serving configuration, prices, monitoring, and import requirements before deployment.