Skip to content

SageMaker model packages

Subscribe to a Software Sushi model on AWS Marketplace and deploy the regional package in your own SageMaker account. These products use model packages with network isolation; they do not use the AMI launch kit.

Published products

Product Version Guide
Brazilian Personal Data Redaction 1.0.1 Portuguese text redaction
Support Ticket Classifier 1.0.0 Ticket classification
LatAm Identifier Redaction 0.2.0 Identifiers and email; names/addresses excluded
E-Invoice Extractor 0.2.0 XML and PDF preview
Source-Grounded Business Summaries 0.1.0 Review-required summaries and actions
Multilingual Relevance Reranker 1.0.0 Bounded passage ranking
Assisted Business Document Type Suggestions 1.0.0 Two suggestions with manual choice, ES/PT/EN

Each guide names its supported inputs, instance types and measured scope. A research preview video is not a Marketplace product. Use the guide's product-specific notebook and limitations.

Subscribe and select the regional package

Open the product's Marketplace listing, review the current version, software price, trial, EULA and refund terms, and subscribe. Choose Continue to configuration, select a supported region and copy the subscribed model-package ARN. Do not substitute another product's package or a seller's private qualification ARN.

AWS hosting, batch, storage and transfer charges are separate from the software fee, including during any software trial. Confirm instance quotas and capacity before deployment.

Caller permissions and execution role

Your signed-in caller or notebook role needs the permissions required to subscribe and orchestrate the operations you choose: create models, endpoint configurations, endpoints or batch jobs, invoke an endpoint and clean up its resources.

The SageMaker execution role is the separate role SageMaker assumes to run the model and access the required private S3 objects. It must trust SageMaker and belong to your account. Follow your organization's least-privilege policies; an execution role does not automatically give the notebook caller subscription or control-plane permissions. Never paste access keys into a notebook or request.

Create the model

Use the regional subscribed package and your own execution role:

sm.create_model(
    ModelName="<your-model-name>",
    ExecutionRoleArn="<your-sagemaker-execution-role-arn>",
    EnableNetworkIsolation=True,
    PrimaryContainer={"ModelPackageName": "<regional-subscribed-model-package-arn>"},
)

The code assumes you already configured an authorized boto3 SageMaker client named sm. Follow the product guide for the exact instance, GPU driver/AMI settings, input format and startup timeout. A CPU classifier and a GPU summarizer have different resource requirements.

Invoke or run batch

A real-time endpoint keeps running until you delete it. A batch job processes its configured S3 input and stops when finished. Use private input/output locations, line splitting and the product's concurrency/payload settings. Inspect every per-record error; completed transport or a completed job does not establish that every record was accepted or correct.

Clean up deliberately

A notebook closing, waiter timeout or failed client call does not stop AWS compute. Inspect the exact resources you created. Delete temporary endpoints, endpoint configurations and models, and stop unfinished batch jobs when appropriate. Follow the notebook's ownership checks; do not delete unrelated or shared resources.

S3 evidence may be deliberately retained by an example. Apply your own retention policy to temporary objects and preserve shared buckets. Infrastructure charges remain governed by AWS and are separate from the software refund policy. Human review and workload-specific evaluation remain part of using each product.