AWS adds India-only Kimi K3 inference profile to Bedrock

AWS has added an India geographic inference profile for Kimi K3 on Amazon Bedrock. Customers can send requests from its Mumbai or Hyderabad regions while limiting model processing to those same two regions.

This is a routing update, not Kimi K3’s Bedrock debut. AWS made Kimi K3 generally available on Bedrock on September 18, offering it through cross-region inference. The India profile adds an in-country route for workloads with data-residency requirements.

The profile ID is in.moonshotai.kimi-k3. According to the Kimi K3 model card, the only eligible source Regions are Asia Pacific (Mumbai), ap-south-1, and Asia Pacific (Hyderabad), ap-south-2. A request from either source Region can be processed in either Mumbai or Hyderabad.

The boundary comes from the inference profile selected for the request, not simply from sending the request to an endpoint in India. AWS says the separate global.moonshotai.kimi-k3 profile can route to supported commercial AWS Regions worldwide, while the India profile restricts destinations to Mumbai and Hyderabad. The profile-based boundary distinguishes the launch from other region-specific deployments, including AWS’s controlled Bedrock route through GovCloud.

In Bedrock’s terminology, the source Region is where a request is made, while a destination Region is where the service may process it. AWS’s geographic cross-Region inference documentation says prompts and outputs may move outside the source Region during cross-Region inference, even when a geography-bound profile keeps processing inside the selected geography.

AWS also says access policies must cover the geographic profile, the model in the source Region and the model in every destination Region. Organizations that block unused Regions through service control policies must allow both India destinations or configure an inference-profile exception.

The Bedrock documentation history records the India Kimi K3 profile as added on October 6, nearly four weeks before AWS published its November 3 announcement. AWS advises customers to call GetInferenceProfile from each intended source Region to check the current routing configuration.