AWS AI Updates: July 18, 2026

1. xAI’s Grok 4.3 Is Now Generally Available on Amazon Bedrock

AWS. xAI’s Grok 4.3 is now generally available on Amazon Bedrock, adding xAI as a new model provider on the platform. The model offers a 1 million token context window, configurable reasoning effort (none, low, medium, high) per request, tool calling, structured JSON output with strict schema validation, and image understanding, accessed through Bedrock’s Standard, Priority, and Flex inference tiers. At launch it runs with in-Region inference only, for example in us-west-2, giving teams an OpenAI-compatible option for long-document reasoning, agentic workflows, and contract or financial document analysis. Source

2. Amazon Bedrock Managed Knowledge Base Adds Enterprise Search for Agents

AWS. Amazon Bedrock Managed Knowledge Base is a fully managed retrieval service that removes the infrastructure work from enterprise RAG pipelines, providing native connectors to S3, SharePoint, Confluence, Google Drive, OneDrive, and web sources with real-time ACL checks enforced at query time. It automatically parses multimodal content including PDFs, presentations, audio, and video up to 10GB, and offers two retrieval paths: a basic Retrieve API that returns ranked chunks and an Agentic Retrieval option that decomposes complex queries into sub-queries for multi-hop and comparative reasoning. Built-in CloudWatch observability and AgentCore Gateway integration expose knowledge bases as tools for MCP-compatible agent frameworks, letting teams skip the days or weeks typically spent assembling comparable pipelines by hand. Source

3. Amazon SageMaker HyperPod Supports Partition-Level Topology for Slurm Clusters

AWS. Amazon SageMaker HyperPod now supports network topology configuration at the partition level for Slurm orchestrated clusters, letting a single cluster run different topologies per partition based on the underlying hardware. Partitions with UltraServer instances such as ml.p6e-gb200.36xlarge use block topology, while hierarchical-interconnect types like ml.p5.48xlarge, ml.p5e.48xlarge, and ml.p5en.48xlarge use tree topology, and HyperPod maintains these settings automatically through scaling events and node replacements. Aligning job placement with each instance type’s interconnect characteristics improves GPU-to-GPU communication and NCCL collective performance, raising training throughput for large model workloads; it is enabled by default on Slurm 25.11 and later. Source