OutYet reporting

Bedrock adds an India-only inference path for Kimi K3

AWS’s new geographic profile routes Kimi K3 processing to Mumbai and Hyderabad, giving teams with residency requirements a defined deployment option. It also comes with the usual cross-Region setup and model-output checks.

OutYet Editorial Desk

Amazon Bedrock now offers an India geographic inference profile for Moonshot AI’s Kimi K3. AWS identifies the profile as `in.moonshotai.kimi-k3` and says it keeps inference processing within the Mumbai and Hyderabad Regions. That gives teams with Indian data-residency requirements a concrete routing option for this model, rather than relying on the broader global profile.

The change is about where requests are processed, not a new model release. AWS describes Bedrock’s cross-Region inference profiles as routing requests from a source Region to eligible destination Regions. For the India profile, the documented destinations are Mumbai and Hyderabad. AWS says cross-Region traffic stays encrypted on its network and does not traverse the public internet; teams still need to configure their Region permissions and allow the destination routing described in its guidance.

AWS lists several ways to invoke Kimi K3 on Bedrock, including Chat Completions, Responses, Converse, and the model-specific InvokeModel API. Its example uses the OpenAI-compatible API from a Mumbai source Region, with the India profile controlling the inference destinations. AWS also provides an example for reading a synthetic Hindi electricity bill and answering follow-up questions in Hinglish. These are integration examples, not independent evaluations of Kimi K3’s accuracy or a guarantee that it will handle every Indic document well.

For technical teams, the useful detail is that the residency choice is attached to an explicit profile ID, while the source Region and allowed destination Regions still matter for IAM and service control policies. AWS says CloudWatch usage metrics are recorded in the source Region and that CloudTrail can record the inference destination. That distinction matters when teams build audit trails: the location where they submit a request and the location where inference runs are different parts of the routing record.

AWS also cautions that image detail settings can raise input-token use and that automated extraction should be checked for the target workflow. The post’s customer cost and latency figures are attributed to Ultrahuman, and its worked bill is synthetic, so neither establishes typical performance. The verified change is regional availability and the documented integration path; users still need to test extraction quality, review costs, and confirm their own compliance requirements.

Related models

Sources