Kimi K3 can now run entirely inside India on Amazon Bedrock

Amazon Web Services has added in-country inference in India for Moonshot AI’s Kimi K3 on Amazon Bedrock. According to the official announcement, a new profile keeps processing inside the Mumbai and Hyderabad Regions.

One profile ID decides where requests run

Kimi K3 is a multi-modal, open-weight model. Amazon says Moonshot AI reported frontier-level coding and general agentic performance at release. On Bedrock it is served through cross-Region inference profiles, and the profile you call determines which destination Regions may handle the request.

There are now two options. The global profile, global.moonshotai.kimi-k3, is what AWS recommends for most uses because it is the cheapest and most widely available. The new India geographic profile, in.moonshotai.kimi-k3, is meant for workloads where processing must stay in Mumbai and Hyderabad. Per the post, US-based inference was already available as well.

AWS says cross-Region traffic is encrypted in transit and stays on its network, with no multi-Region VPC setup needed. One detail matters for compliance teams: logs and metrics are stored in the Region where the request was submitted, not where it was processed. Optional CloudTrail data events record the destination Region in an additionalEventData.inferenceRegion field.

Developers can call the model through OpenAI-compatible Chat Completions and Responses APIs, or through Converse and InvokeModel. AWS recommends the OpenAI-compatible ones. Its example uses ap-south-1 (Mumbai) as the source Region and short-lived tokens from the AWS token generator library instead of long-lived API keys. Organizations that restrict Regions must allow the source Region plus both Indian destinations in their IAM and service control policies.

A Hindi bill reader and a caching claim

The post demonstrates the model on a synthetic Hindi electricity bill. It reads the scanned image, extracts fields as JSON with dates normalized and Devanagari digits converted, and then answers follow-up questions in Hinglish. AWS suggests setting the image detail to high for OCR, at the cost of more input tokens, and putting images before text. It also cautions that automated extraction should be checked, with validation or human review for critical workflows.

Those figures come from Ultrahuman, quoted in the announcement. Aditya Patawari, the company’s VP of Engineering, said its Jade health assistant handles more than 14,000 conversations a day and that, with prompt caching, “we’ve cut LLM costs by 80 percent and made multi-turn chats 85 percent faster.” The numbers are the customer’s own and were not independently verified.

On price, the post says a flex tier gives a 50 percent discount on standard pricing for latency-tolerant jobs, while a priority tier costs more for faster queueing. It does not publish per-token rates or say whether the India profile is priced differently from the global one, so readers need the Bedrock pricing page.

Who gains from a Mumbai and Hyderabad boundary

The main beneficiaries are Indian banks, insurers and public sector teams that digitize KYC forms and bills, the use case AWS chose. They can now use a large open-weight model without sending data abroad. The source does not say which regulations drive the demand, so that link is context rather than part of the announcement.

The legal backdrop is mixed. Commentary on India’s Digital Personal Data Protection framework, such as this law firm analysis, describes a default that permits transfers abroad unless the government bars a destination, with no universal localization mandate. The same commentary says sector regulators such as the Reserve Bank of India can still impose their own rules, and that the cross-border provisions had not yet commenced, with one source expecting major provisions around May 2027. In practice, an in-country profile lets a compliance team avoid that question for now, rather than answering it. Those details come from secondary commentary and should be checked against the official notification.

The profile also has a narrower footprint than the global one. When AWS introduced Canada cross-Region inference for Claude Haiku 5.5 on October 27, it routed only between Canada Central and Calgary and warned that such profiles have limited model selection and may not get future models. The India post carries no such warning, but it points readers to the model card for current profile availability. Teams planning around a single country’s profile should check there before assuming new Kimi versions will follow.

The model itself is large. Press coverage at launch, such as PureAI’s report, described Kimi K3 as a mixture-of-experts system of about 2.8 trillion parameters with a 1-million-token context window and native vision. Moonshot’s performance claims against rival models are its own, and the AWS post offers no benchmarks of its own, so the quality of Indic-language results is shown only through one synthetic example.