AWS has announced in-country inference for Kimi K3 on Amazon Bedrock, so requests to the model can be processed entirely within India. Kimi K3 is a multi-modal, open-weight model from Moonshot AI, which reported frontier-level coding and general agentic performance at release.
Two profiles, two trade-offs
Kimi K3 is served through cross-Region inference profiles. You call the model from a source Region, and Bedrock routes the request to a destination Region. The profile ID you invoke decides which destinations are eligible:
- Global (
global.moonshotai.kimi-k3): AWS recommends it for most uses, as the lowest-cost and most widely available option. - India geographic (
in.moonshotai.kimi-k3): for workloads where processing has to stay in the Mumbai and Hyderabad Regions.
AWS says traffic between Regions is encrypted and stays on the AWS network, with no need to build multi-Region VPC infrastructure. The model card is the place to check which source Regions can call each profile.
What you need to change
The post uses the Mumbai Region as the source and shows an IAM policy that allows the model in both Indian Regions plus the India inference profile. Organisations that restrict AWS Regions need to keep the source Region permitted and allow routing to both Indian destinations. The post points to a profile-scoped exception where that fits. The model is reachable through OpenAI-compatible Chat Completions and Responses APIs, as well as Converse and InvokeModel.
One detail matters for audits: metrics and logs are stored in the source Region where the request was made, whichever destination handled it. CloudTrail data events can record the destination Region.
Document workloads as the example
AWS demonstrates the profile with a scanned Hindi electricity bill. The model reads the image directly, extracts fields as JSON under a schema, writes an English summary and answers follow-up questions in Hinglish, with no separate OCR or translation step. AWS advises verifying extraction quality for your own use case and adding validation or human review for critical workflows.
Cost levers
The post lists a flex service tier at a 50 percent discount on standard pricing for latency-tolerant work, and a priority tier at a higher price. Prompt caching, implicit or explicit, discounts repeated prompt prefixes. The post does not say whether the India profile is priced differently from the global one, so check the pricing page.
AWS quotes Ultrahuman, whose health assistant handles more than 14,000 conversations a day. The company reports an 80 percent cut in LLM costs and 85 percent faster multi-turn chats with prompt caching. These are customer-reported figures, not independent benchmarks.