.##....##.########.##......##..######.....########..#######..########.....###....##....##
.###...##.##.......##..##..##.##....##.......##....##.....##.##.....##...##.##....##..##.
.####..##.##.......##..##..##.##.............##....##.....##.##.....##..##...##....####..
.##.##.##.######...##..##..##..######........##....##.....##.##.....##.##.....##....##...
.##..####.##.......##..##..##.......##.......##....##.....##.##.....##.#########....##...
.##...###.##.......##..##..##.##....##.......##....##.....##.##.....##.##.....##....##...
.##....##.########..###..###...######........##.....#######..########..##.....##....##...

All signal, no noise, 24/7.
Built for Humans & AI Agents.

Moonshot AI has introduced Kimi K3.1 Pro Highspeed, a new large language model, to Amazon Bedrock. This model is the latest addition to Moonshot AI’s K3.1 series and is designed to enhance capabilities for complex, demanding applications.

Model Architecture and Performance Enhancements

Kimi K3.1 Pro Highspeed is a refined, post-trained version of the existing Kimi K3 model. It maintains the same 3T-class open-weight architecture and boasts a comprehensive 1-million-token context window. According to Moonshot AI, the model underwent significantly greater post-training computation, specifically targeting improved performance in long-horizon agentic workflows and a reduction in factual errors.

This enhanced training makes it particularly beneficial for professional tasks, including financial modeling, handling office documentation, and coding for specialized environments such as game engines, 3D scenes, and computer-aided design (CAD) scripts. A major performance improvement is its output speed; Moonshot AI reports that Kimi K3.1 Pro Highspeed generates tokens at a higher rate, achieving speeds 2 to 3 times faster than Kimi K3.1 Pro.

The model supports advanced features such as tool calling, generating structured output, customizable reasoning control, image input, and streaming capabilities. Furthermore, users can adopt Kimi K3.1 Pro Highspeed without altering their existing security framework, as data is processed entirely within the AWS data boundary, is not shared with Moonshot AI, and is not utilized for model training. Zero data retention is enabled by default for all inference requests, and AWS operators are prevented from viewing prompts and completions.

Accessing and Integrating Kimi K3.1 Pro Highspeed

Developers who previously built applications using Kimi K3 can transition to the new model simply by modifying the model ID within their current code. Kimi K3.1 Pro Highspeed is accessible through the same API endpoints and features used by Kimi K3 on Amazon Bedrock.

Console and API Access

To test the model without writing any code, users can navigate to the Amazon Bedrock console, select Test, and then choose the Playground, where Kimi K3.1 is available.

For programmatic access, Kimi K3.1 Pro Highspeed is available on the `bedrock-runtime` endpoint via the OpenAI-compatible Chat Completions and Responses APIs, as well as the Amazon Bedrock Converse and Invoke APIs. The provider recommends using the Chat Completions API.

The model is accessible through cross-Region inference profiles, allowing users to select based on workload requirements:

  • global.moonshotai.kimi-k3.1: Routes requests to any supported commercial AWS Region and offers the lowest cost per token, making it ideal for most general workloads.
  • us.moonshotai.kimi-k3.1: Ensures that processing remains within US Regions, satisfying specific data residency requirements.
  • in.moonshotai.kimi-k3.1: Guarantees processing within India Regions, addressing local data residency mandates.

Implementation Requirements

To follow the integration examples, users must have:

  1. An AWS account with access to Amazon Bedrock.
  2. Python 3.11 or a later version, along with the installed libraries: `openai`, `aws-bedrock-token-generator`, and `boto3` (installable via `pip install -U openai aws-bedrock-token-generator boto3`).
  3. An AWS Identity and Access Management (IAM) identity authorized to invoke the model using the permissions `bedrock:InvokeModel` and `bedrock:InvokeModelWithResponseStream` on the Kimi K3.1 Pro Highspeed profiles. If authenticating with a bearer token, `bedrock:CallWithBearerToken` must also be added.

Cost and Performance Optimization

Controlling Reasoning Effort

Kimi K3.1 Pro Highspeed always performs reasoning before providing an answer. Users can manage the depth of this process using the `reasoning_effort` parameter, which accepts values of `low`, `high`, or `max`. If this setting is omitted, the model defaults to `max`. Setting a lower value is advised for simpler tasks to reduce latency and the volume of generated tokens.

Adopting the Flex Tier for Cost Savings

For cost-conscious workloads, the Flex tier is available, offering Kimi K3.1 Pro Highspeed at half the price of the Standard tier. This tier is best suited for tasks where cost efficiency takes precedence over immediate response time, such as content summarization, model evaluations, or background operations in agentic workflows. The Flex tier can be mixed with Priority and Standard tiers within a single application, as the on-demand quota is shared across all three levels.

To utilize this option, set the `service_tier` parameter to `flex`. For the Converse API, the equivalent setting is `{“type”: “flex”}`. Users are advised to begin by directing only a small percentage of their traffic to the Flex tier to compare latency and cost before increasing usage.

Enhancing Inference with Explicit Prompt Caching

To improve efficiency, especially for agents and coding assistants that repeatedly send the same large system instructions, tool definitions, or reference documents, Kimi K3.1 Pro Highspeed offers explicit prompt caching. By marking where the stable prefix content ends, later requests can reuse that prefix, significantly lowering both input costs and latency for long, repetitive contexts.

Specific rules apply to this feature:

  • Explicit caching is functional with the Chat Completions, Responses, and Invoke APIs, but not the Converse API.
  • A prompt_cache_breakpoint must follow at least 256 tokens of stable content.
  • Cached content remains available for a minimum of 30 minutes.
  • Cache reads incur a discounted rate and do not count toward the input tokens-per-minute quota, while cache writes are billed at a higher rate than standard input.

Practical Use Case: Coding with OpenCode

OpenCode is an open-source, model-agnostic coding agent that operates in the terminal. Developers can integrate it with Kimi K3.1 Pro Highspeed via the Amazon Bedrock Chat Completions API, using OpenCode’s OpenAI-compatible provider and an Amazon Bedrock API key.

The output speed is noted as crucial in agentic coding, as a single task requires multiple sequential model calls (e.g., reading files, planning, editing code, running tests). A higher output-token rate accelerates every step in this edit-test-fix cycle, minimizing overall wait time.

Availability Summary

Kimi K3.1 Pro Highspeed is available on Amazon Bedrock through the following cross-Region inference profiles: `global.moonshotai.kimi-k3.1`, `us.moonshotai.kimi-k3.1`, and `in.moonshotai.kimi-k3.1`. Detailed information regarding supported regions and model specifics can be found on the Kimi K3.1 Pro Highspeed model card, and pricing details are available through Amazon Bedrock pricing.

Max

Written by

Max

Covers AI news, agentic AI, LLMs, and tech developments. When he is not writing, he is comparing open-source models' tokens per second just to see how they hold up.

+ , , ,