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

What it is

Amazon Bedrock now serves Gemma 4 (Apache 2.0, three instruction-tuned variants with reasoning, function calling, and text/image multimodal input) through a new bedrock-mantle endpoint inside the AWS European Sovereign Cloud's first and only Region, eusc-de-east-1 (Brandenburg, Germany, GA since January 2026).

How it works

bedrock-mantle exposes OpenAI-compatible Responses and Chat Completions APIs, so existing OpenAI SDK code can point at it by changing only the base URL and API key. The Responses API can manage conversation state server-side (avoiding re-sending history), scoped to a Bedrock "project" boundary used for access control, cost tracking, and usage monitoring — but this storage is opt-in: setting store to false means Bedrock retains nothing. Authentication uses either AWS SigV4 (for Chat Completions with AWS credentials) or Bedrock API keys: short-term keys expire automatically within 12 hours and inherit the issuing IAM principal's permissions (recommended for production), while long-term keys persist until a configured expiration and are meant only for development. Global cross-Region inference — Bedrock's normal routing across worldwide commercial Regions — is explicitly disabled in this sovereign cloud, so every Gemma 4 request stays confined to eusc-de-east-1.

Preconditions / Affected

Requires an AWS account with access to the AWS European Sovereign Cloud and an IAM principal granted the bedrock-mantle:CreateInference and bedrock-mantle:CallWithBearerToken actions. Requires pip install "openai>=2.45.0" aws-bedrock-token-generator. Applies only within eusc-de-east-1; not relevant to standard commercial Bedrock Regions.

What to do

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

• Create and attach an IAM policy granting only bedrock-mantle:CreateInference and bedrock-mantle:CallWithBearerToken, scoped to your account and project resources.
• Prefer short-term Bedrock API keys (≤12-hour expiry, inherit caller's IAM permissions) over long-term keys for any production use; store keys in AWS Secrets Manager rather than passing them via environment variables if they need to persist.
• Set store: false on Responses API calls where you don't want Bedrock to retain request/response data.
• Enable CloudTrail logging on the endpoint and set CloudWatch alarms for anomalous request-volume patterns to audit who invoked inference and when.

Caveats

The post notes that for certain models, Bedrock may retain limited data for abuse detection despite the zero-data-retention default — the post points to separate Bedrock abuse-detection documentation rather than detailing that exception itself. No persistent resources are created by the walkthrough, so no cleanup beyond letting short-term keys expire.