New updates on AWS GovCloud (US) | Issue May 16 - May 31

GenAI availability in AWS GovCloud (US) is accelerating rapidly, with AgentCore, Kiro, and Mantle all launching in recent months, and the momentum continues. OpenAI GPT-OSS (20B and 120B), OpenAI GPT-5.4, and NVIDIA Nemotron (Nano 9B v2, Nano 12B v2, Nano 30B, and Super 120B) models are now available on Amazon Bedrock, joining existing AWS GovCloud (US) models from Amazon, Anthropic, and Meta.

For leaders evaluating AI strategies, this means reduced vendor lock-in risk, faster time-to-production for AI workloads, and the ability to benchmark multiple model families against mission requirements without re-architecting applications. OpenAI GPT-OSS models deliver powerful language understanding and generation with tunable reasoning depth. OpenAI GPT-5.4 brings native computer-use capabilities, and deep reasoning across coding, documents, and multi-step agentic tasks. The Nemotron models offer SLM and LLM capabilities with high compute efficiency for specialized agentic AI. To learn more and get started, visit the Amazon Bedrock console or visit the Bedrock service documentation.

From new foundation models to agentic AI infrastructure and AI-powered development tools, AWS GovCloud (US) is delivering GenAI capabilities at a pace and breadth that matches the urgency of government modernization. As agencies move from AI experimentation to production-scale deployment, the expanding model portfolio in AWS GovCloud (US) positions customers to adopt a multi-model strategy rather than betting on a single provider. Future issues will cover AgentCore, Kiro, Mantle, and the growing set of services enabling mission-critical AI in the most secure AWS environment.

Here's what practitioners need to know to get started with the new models.

GPT-OSS and Nemotron are accessible through both the Converse API on bedrock-runtime and the Chat Completions endpoint on bedrock-mantle, giving teams flexibility in how they integrate. For agentic architectures, GPT-OSS and the larger Nemotron models (Nano 30B, Super 120B) support native tool calling, while GPT-OSS and Nemotron models expose a reasoning_effort parameter (low, medium, high) to tune chain-of-thought depth against latency and cost. The Nemotron Nano 12B v2 variant supports multimodal input (text + images) suitable for document understanding, visual inspection, and image-captioning workloads in regulated environments.

OpenAI GPT-5.4 uses the Responses API, a newer interface that introduces server-side conversation chaining, response storage, and native support for parallel tool execution. It supports vision, structured JSON output, streaming, and configurable reasoning effort. Migration from existing Chat Completions code is straightforward: the Responses API accepts simplified input formats that closely mirror existing message arrays, making it easy to bring existing OpenAI workloads into AWS GovCloud (US).

All models are served on FIPS-enabled endpoints in us-gov-west-1 with Nitro isolation and Zero Operator Access, callable over HTTPS from any network location via SigV4 or Bedrock API key.