Kimi K3 on Amazon Bedrock is now live in India: data stays in the country

If you are building a product for Indian fintech or the public sector, you know the pain: customer data cannot leave the country, while the best models usually run somewhere in the United States. AWS has closed that gap for Moonshot AI's Kimi K3. Amazon Bedrock now offers an inference profile that processes requests only in the Mumbai and Hyderabad regions.

How to choose where a request is processed

Kimi K3 on Bedrock works through cross-Region inference profiles. You invoke the model from your own region, and Bedrock chooses where to run it. The profile ID defines the set of permitted regions.

AWS recommends the global.moonshotai.kimi-k3 global profile for most workloads: it is the cheapest and most widely available. If you need data residency, use in.moonshotai.kimi-k3: processing remains in Mumbai and Hyderabad. Inter-Region traffic is encrypted and does not traverse the public internet, and there is no need to configure your own multi-region VPC.

In practice, permissions matter. The IAM policy must allow calls to both destination regions in India, not just the region from which you send the request. If your organization uses regional SCP restrictions, you will need to update them or requests will fail. Logs and metrics remain in the source region, while CloudTrail can include the inferenceRegion field so you can verify where each request was actually processed.

Example: reading a Hindi bill without OCR

AWS demonstrates a typical workflow for KYC forms and utility bills. You send a base64-encoded image of an electricity bill, and the model reads the Devanagari text directly. You can then ask questions in Hinglish—Hindi written in the Latin alphabet and mixed with English. For example: “Bill kitne ka hai aur last date kab hai?” The model responds with the amount of 2,217.20 rupees and the payment deadline.

For structured output, use function calling or JSON Schema. The example pipeline does four things: transcribes the document in its original script, extracts fields into JSON, normalizes dates to ISO 8601, converts Devanagari numerals into standard digits, and transliterates names. It then writes a short English summary and answers follow-up questions.

The article offers a couple of practical tips. For fine details, set detail: high; it is more accurate but consumes more tokens. Placing the image before the text can sometimes improve quality. You can connect through the standard openai library because the API is OpenAI-compatible, with a token generated from standard AWS credentials.

Cost: tiers and caching

The article's headline figure comes from a customer. Ultrahuman uses Kimi K3 in Jade, an assistant that answers questions about smart-ring and glucose data across more than 14,000 conversations a day. According to the company, prompt caching reduced costs by 80% and sped up multi-turn chats by 85%. This is the company's own claim in an AWS partner blog post and has not been independently verified.

To reproduce that effect, place the static portion—the system prompt and tool descriptions—at the beginning of the request, and dynamic content such as the image and question at the end. There is an implicit caching mode that works automatically and an explicit mode in which you set breakpoints to achieve more cache hits. For workloads without strict latency requirements, the flex tier offers a 50% discount; urgent workloads can use the priority tier for an additional charge.

What to do with it

If you have Indian customers who require their data to remain in the country, you can now use a powerful open model without hosting it yourself. Start by checking IAM and SCP settings, then build a field-extraction prototype using 20–30 real documents. AWS itself advises against blindly trusting automated extraction: add validation and human review for critical processes. Claims that the model delivers “frontier” performance come from Moonshot AI, so benchmark it against your current model using your own data.