Amazon Bedrock offers Kimi K3 through an India geographic inference profile that routes requests between AWS Regions in Mumbai and Hyderabad. CIO Elets reported that AWS announced the option on November 3, 2026; AWS identifies the profile as in.moonshotai.kimi-k3 and lists Mumbai (ap-south-1) and Hyderabad (ap-south-2) as both eligible source Regions and inference destinations (AWS model card, CIO Elets).
That distinction matters when configuring a deployment: the AWS Region from which an application submits a request is not the same thing as the end user’s physical location. AWS’s India profile lists two eligible source Regions, so a team assessing it should check where its Bedrock requests originate as well as where inference can run.
For an India-geographic deployment, check the AWS request Region and logging settings alongside the inference destination.
India and Global profiles route differently
The India profile confines its listed source and destination Regions to Mumbai and Hyderabad. AWS’s Global profile routes among supported commercial AWS Regions worldwide, making it the broader geographic option (AWS model card).
| Kimi K3 profile | Routing boundary | Practical distinction |
India (in.moonshotai.kimi-k3) | Mumbai (ap-south-1) and Hyderabad (ap-south-2) as source and destination Regions | Use this profile when the deployment calls for the listed India-geographic routing. |
Global (global.moonshotai.kimi-k3) | Supported commercial AWS Regions worldwide | Offers a wider geographic routing scope. |
The profile describes inference routing; it does not by itself settle where every application record is stored. AWS says logs and metrics are stored in the AWS source Region where the request was submitted, regardless of the inference destination. Cross-Region traffic is encrypted in transit and stays on the AWS network rather than travelling over the public internet (AWS announcement).
Logging follows the request’s source Region
Model-invocation logging is optional and can include prompts and outputs. AWS also says CloudTrail data events can record the inference destination in additionalEventData.inferenceRegion (AWS announcement).
For a deployment review, that means checking the AWS source Region and the invocation-logging configuration—not just the India profile’s destination Regions. If an application enables content logging, its prompts and outputs may be included in those logs.
APIs and input limits to account for
AWS lists Chat Completions, Responses, Converse and InvokeModel for Kimi K3, and recommends the OpenAI-compatible APIs where possible. The model accepts text and image inputs and returns text; video input is unsupported. The model card lists a context window of 1 million tokens (AWS model card).
Converse has specific constraints: multi-turn requests that include earlier reasoning content can fail with an InternalServerException, and attached PDF and HTML inputs are rejected. On the bedrock-runtime endpoint, each output token consumes 10 tokens from the model quota. These details can affect API choice and quota planning.
AWS’s Hindi electricity-bill example uses fictional, synthetic documents and shows structured extraction and Hinglish responses. AWS advises validating extraction for the intended use before relying on it for important decisions (AWS announcement).