Kimi K3 gets Indian regional processing on Amazon Bedrock

Processing can remain in Indian regions

According to an AWS announcement dated November 3, 2026, Kimi K3 is now available on Amazon Bedrock with Indian geographic inference alongside global and US processing. Kimi K3 is a multimodal, open-weight foundation model that its developer, Moonshot AI, said achieved leading-edge coding and general agentic intelligence when it was released.

When the India profile is used, inference processing remains in the Mumbai and Hyderabad AWS Regions. The India geographic inference profile uses the in.moonshotai.kimi-k3 identifier, while the global profile is available under global.moonshotai.kimi-k3. AWS recommends the global option for most use cases because it offers the lowest cost and broadest availability.

With cross-Region inference, the customer initiates a request from a source Region, and Amazon Bedrock processes it in a potential destination Region. AWS says data is encrypted while moving between Regions within the AWS network and does not pass over the public internet.

Ultrahuman: cheaper and faster conversations

According to AWS, Ultrahuman is already using Kimi K3 in production. The company’s Jade health and wellness assistant responds in natural language based on users’ own Ring and glucose data, handling more than 14,000 conversations per day.

Aditya Patawari, Ultrahuman’s vice president of engineering, said the company was able to integrate Kimi K3 through the API it was already using, while keeping health data within AWS and having Mumbai available as a backup Region. The company said prompt caching reduced large language model costs by 80 percent and made multi-turn conversations 85 percent faster.

Can answer Hinglish questions from documents

AWS’s demonstration describes an Indian document-processing use case. Kimi K3 can read a scanned Hindi-language document, extract structured data from it, and then answer mixed Hindi and English, or so-called Hinglish, questions.

AWS links this to financial and public-service processes that require the digitization of customer identification forms, utility bills and similar documents. The model can process images, PDFs and several other file types directly, meaning AWS’s example shows it can interpret multilingual text without a separate optical character recognition or translation service.

In the example, the model had to answer how much was due on an artificially generated Hindi-language electricity bill, when the payment deadline was, and how much would be owed after a late payment. The model’s response said the bill’s total was 2,217.20 Indian rupees, the deadline was October 20, 2026, and the amount in case of late payment would be 2,261.54 rupees.

APIs and structured outputs

Kimi K3 can be accessed through multiple APIs at the Amazon Bedrock Runtime endpoint. These include the OpenAI-compatible Chat Completions and Responses APIs, the Converse API, which is unified across supported models, and the InvokeModel API, which uses a model-specific JSON format. In its demonstration, AWS uses the OpenAI-compatible APIs and the openai Python library.

The model also supports streaming responses, tool use and structured outputs. The latter can be used with function calling or a JSON Schema-based output format, which AWS says enables consistent field extraction in document-processing workflows.

Use requires an AWS account, Python 3.11 or later, and the AWS CLI installed and configured with the necessary permissions. AWS has also published the complete source code for the example and recommends using short-lived credentials to invoke the model wherever possible.