AWS AI Updates: November 4, 2026

1. Amazon Bedrock Adds India-Only Inference for Kimi K3 Through a New Geographic Profile

Amazon Bedrock now offers in-country inference in India for Moonshot AI’s open-weight multimodal Kimi K3, alongside the existing global and US-based options. Calling the in.moonshotai.kimi-k3 cross-Region inference profile keeps processing within the Mumbai (ap-south-1) and Hyderabad (ap-south-2) Regions, while global.moonshotai.kimi-k3 stays the recommended default for lowest cost and broadest availability; IAM policies need to allow both India destination Regions, and organizations that restrict Regions through SCPs must keep the source Region permitted and allow routing to both. Kimi K3 is served through the Chat Completions, Responses (both OpenAI-compatible and recommended by AWS for this model), Converse and InvokeModel APIs, with tool calling, structured outputs through function calling or a JSON Schema output format, and image and PDF input. AWS notes that output quality is sometimes higher when images come before text, and suggests the high image detail setting for OCR at the cost of more input tokens. Prompt caching runs in an implicit mode by default or an explicit mode with prompt_cache_breakpoint markers, the flex service tier is 50% cheaper than standard, and CloudTrail data events record the Region that actually served each request in additionalEventData.inferenceRegion, while metrics and invocation logs stay in the source Region. The post includes a sample pipeline for OCR, JSON field extraction and Hinglish follow-up questions over Hindi documents, and a quote from Ultrahuman, which says prompt caching cut its LLM costs by 80% and made multi-turn chats 85% faster in its Jade assistant (more than 14,000 conversations a day). Source

2. AWS Transform for Mainframe Groups Business Functions by Data Field and Adds a Requirements Coverage Report

AWS Transform for mainframe updated its assess and reimagine workflow for z/OS COBOL and PL/I applications. Business functions are now grouped only when they affect the same data fields, rather than whenever they touch the same table or file, and they now include screens and queues and carry clearer names and descriptions, giving teams a more accurate catalog to modernize from. A new requirements coverage report classifies every line of source code by how it was handled, links every business rule to the exact lines it came from, and backs every requirement with the business rules that support it, so coverage can be reviewed from the full portfolio down to a single line before forward engineering in Kiro or other IDEs; previously, users had to build their own traceability reporting from AWS Transform artifacts. The changes apply in all Regions where AWS Transform for mainframe is offered, and the announcement gives no pricing change. Source