AWS adds India-only inference for Kimi K3 AWS adds India-only inference for Kimi K3

The profile sets the boundary

The India profile joins the existing global route. The destination set follows the profile selected in the request. A source endpoint and a processing destination are separate parts of the configuration.

The model was already available

Kimi K3 reached general availability on Bedrock on September 18. That release brought Moonshot’s open-weight model into a managed service with native vision and a one-million-token context window. AWS described coding and knowledge work as target uses. The model can work with visual material alongside text, making document-heavy applications relevant to its deployment footprint. The September release also identified Kimi K3 as Bedrock’s first open-weight model with explicit prompt caching. Those capabilities predate the India profile. The November change concerns where requests run rather than a new model version or a larger context window.

Routing is a managed service

Bedrock introduced cross-region inference in August 2024 to handle traffic bursts and capacity constraints. Instead of an application distributing requests itself, the service can use capacity within a predefined group of regions. Its launch announcement described prioritizing the source region when possible and routing elsewhere when needed. That design separates the application’s request destination from the infrastructure that completes the work. It also puts the eligible region group alongside model selection as an operational choice. A team can keep a familiar API integration while changing the geographic scope of the capacity available behind it.

Logs expose the second location

AWS’s cross-region implementation guide explains how an inference profile groups model resources behind one identifier. Applications pass that identifier instead of an individual model ID. Requests routed between regions travel over AWS’s backbone with encryption. The guide distinguishes this processing path from logging: CloudWatch and CloudTrail records stay in the source region. A rerouted CloudTrail event includes an inferenceRegion field identifying where processing occurred. Model invocation logs can also include that field when enabled. These records give operators a way to distinguish the endpoint they called from the region that actually served a request. The guide also warns that data can be processed and transmitted across the regions configured in a profile.

Caching is a separate control

The September Kimi K3 launch guide described explicit caching for repeated prompt prefixes. Stable instructions, shared context and tool definitions can be reused across requests rather than processed afresh each time. Cache writes and reads have different billing treatment, so the benefit depends on reuse. AWS also described access through its Bedrock Runtime APIs, including OpenAI-compatible interfaces. These are separate decisions from geographic routing: an application chooses how to send a request, what context can repeat and which profile may serve it. Keeping those choices distinct makes a deployment easier to inspect.