AWS Introduces Kimi K3.1 Pro Highspeed Model with 1 Million Token Context

Amazon Web Services (AWS), the cloud division of Amazon, has added the Kimi K3.1 Pro Highspeed artificial intelligence model from Chinese company Moonshot AI to the Amazon Bedrock platform. The model supports a context window of up to 1 million tokens and is available through global processing profiles, as well as separate profiles for the US and India.

Kimi K3.1 Pro Highspeed is based on Kimi K3 with an architecture of class 3 trillion parameters and open weights. During additional training, developers used more computing resources, focusing on performing complex multi-step tasks and reducing the number of factual errors.

According to Moonshot AI, the token output generation speed of the Highspeed version is 2–3 times higher than that of the standard Kimi K3.1 Pro. Independent comparative testing results are not provided in the AWS publication.

The model supports working with images and external tools

Kimi K3.1 Pro Highspeed supports image processing, external tool invocation, structured response formation, and reasoning depth customization. These capabilities allow the model to be used for document analysis, financial modeling, and programming.

For developers already using Kimi K3 through Amazon Bedrock, there is an option to upgrade to the new version by changing the model identifier without significant reworking of the existing integration.

According to AWS documentation, requests are processed within the company's infrastructure and are not transmitted to Moonshot AI. Amazon also states that user data is not used for model training, and the content of requests and responses is not saved after processing.

AWS has provided several pricing modes

For Kimi K3.1 Pro Highspeed, Priority, Standard, and Flex modes are available, differing in cost and request processing conditions.

The Flex tariff offers a cost 50% lower than Standard but allows for longer response times. This mode can be used for tasks that do not require minimal latency, including document processing and automated computational processes.

Additionally, the model supports saving recurring parts of requests for reuse, which can reduce costs when working with large volumes of identical context.

Actual costs of using the model depend on the chosen region, pricing mode, and the number of tokens processed. AWS did not provide an independent comparison of the overall cost of task execution with competing models in the original publication.