Run open weight models on Amazon Bedrock in AWS European Sovereign Cloud

What it claims

Gemma 4 is the first open-weight model family on Amazon Bedrock in the AWS European Sovereign Cloud. It is served only in-Region in eusc-de-east-1 (Brandenburg, Germany). Global cross-Region inference is not available in this cloud, so requests never route to commercial AWS Regions. The Sovereign Cloud has been generally available since January 2026. Customer metadata (roles, permissions, resource labels, configurations) also stays in the EU, not just content. Operations are run by EU residents, with a stated, ongoing transition to EU citizens only. During that transition the team is mixed.

How it works

• Endpoint: bedrock-mantle exposes OpenAI-compatible Responses and Chat Completions APIs. Existing OpenAI SDK code ports by changing only the base URL and API key. Both the Python and TypeScript SDKs are named.
• Stateful conversations: The Responses API keeps conversation state server-side, so you don't resend history. Stored responses are scoped to a Bedrock project, a logical boundary that also drives access control, cost tracking and usage monitoring.
• Retention: Zero data retention is the default. Setting store to false on a Responses request means Bedrock keeps neither request nor response. The exception: "for certain models, limited retention might apply for abuse detection." The page doesn't say whether Gemma 4 is one of them.
• Audit: Every endpoint call is logged in CloudTrail. CloudWatch can alarm on usage patterns.

How to use it

IAM actions to grant. The sample policy is scoped to account and project IDs:

• bedrock-mantle:CreateInference runs inference.
• bedrock-mantle:CallWithBearerToken allows authentication with a Bedrock API key.

Install. Quotes are normalized here; the page's version has a curly quote that will break a shell:

pip install "openai>=2.45.0" aws-bedrock-token-generator

Authentication options:

1. Short-term API key. Expires within 12 hours and inherits the permissions of the generating IAM principal. This is the recommended choice for production; use the auto-refreshing generator. Passing one via an environment variable gives a key that does not refresh.
2. Long-term API key. Lasts until its configured expiry. Intended for development; if you use it, store it in Secrets Manager.
3. AWS credentials with SigV4 signing. This is the alternative shown for calling Chat Completions directly.

Model. The walkthrough targets Gemma 4 31B. There are three instruction-tuned variants, all with built-in reasoning, native function calling and text+image input.

Pricing. Per token, no upfront commitment, and usage counts toward existing AWS commitments. No rates are given; the page points to the Bedrock pricing page.

Limits

• Missing content in the fetched text:
• The variant-selection table is absent, so only the 31B variant is named.
• The code samples and the IAM policy JSON are absent, including the base URL and model ID strings.
• Single Region: One Region only, and there is no cross-Region failover path.
• Abuse-detection retention: It may apply depending on model. Check the abuse detection docs before assuming zero retention for Gemma 4.
• Fine-tuning: The page cites customization as an advantage of open weights. It does not say Bedrock in this cloud offers fine-tuning for Gemma 4.