Kimi K3 gains India-based inference on Amazon Bedrock
The new India profile keeps inference within Mumbai and Hyderabad.
AWS has introduced an India inference profile for Kimi K3. Its example extracts fields from a fictional Hindi electricity bill and answers follow-up questions mixing Hindi and English. AWS recommends verifying extraction results before important decisions; the demonstration does not guarantee accuracy for other documents. The walkthrough also flags an API-specific requirement: Responses needs default-project access in addition to the inference permissions discussed below.
The model documentation shows that requests can originate in either Mumbai or Hyderabad and be processed in either region. For application developers, AWS recommends Chat Completions or Responses instead of Converse. Converse rejects attached PDF and HTML documents and can fail when earlier reasoning blocks accompany a follow-up request. Those blocks must be removed for multi-turn use through Converse.
The routing guide explains the access requirements before calling a geographic profile. Identity permissions must cover the profile, the model in the requesting region and the model in every listed destination. Organizational policies restricting regions must permit those destinations or include an applicable inference-profile exception. That exception does not override restrictions on the originating call: the source region must remain permitted.