Kimi K3 joins Amazon Bedrock

Amazon Bedrock now offers Moonshot AI’s Kimi K3 through managed Global and US cross-Region inference profiles. The Bedrock model card gives AWS customers access to K3’s one-million-token context window and native image input alongside IAM authorization, encryption options, audit logging, and consolidated billing.

The deployment targets large codebases, document collections, and long-running agent workflows. It also gives teams that cannot send data to Moonshot-hosted endpoints an AWS-managed route, subject to their data residency, security, and procurement requirements.

A sparse giant built for context

Kimi K3 is a roughly 2.8-trillion-parameter mixture-of-experts model. According to the Moonshot repository, it activates 16 of 896 experts for each token, allowing only a fraction of the network to run during inference. That sparse design makes a model of this size less expensive to serve than a dense model with the same total parameter count.

The architecture combines Kimi Delta Attention, Attention Residuals, and a Stable LatentMoE framework. Moonshot says Kimi Delta Attention reduces memory and compute requirements over long sequences, while Stable LatentMoE delivers a 2.5-fold scaling-efficiency improvement over Kimi K2 under its tests.

K3 accepts text and images within its one-million-token context window. Moonshot’s open weights also support video input, although the Bedrock deployment currently limits multimodal input to images.

At the time of the Bedrock launch, Artificial Analysis gave K3 a score of 60 on its Intelligence Index, tied with GLM-5.3 and three points behind Claude Opus 5. The index aggregates several evaluations and provides a directional comparison. Code, retrieval, tool-use, and latency tests still need to reflect the intended production workload.

Three APIs and a Converse bug

Bedrock exposes K3 through the Responses, Chat Completions, and Converse or Invoke APIs. AWS recommends the OpenAI-compatible Responses and Chat Completions interfaces, which reduce migration work for applications already built around the OpenAI SDK.

A current Converse failure mode affects multi-turn requests that include reasoning content from earlier assistant messages. Those requests can return an InternalServerException, which can break default LangChain and Strands Agents configurations. Applications using Converse should remove reasoning blocks from conversation history; applications using Chat Completions avoid that specific path.

• Prompt caching: Explicit caching requires at least 1,024 tokens per checkpoint and uses a minimum 30-minute time to live. Cache reads cost 10% of the applicable input-token rate.
• Image input: AWS recommends placing image blocks before text blocks for better results. Chat Completions honors the detail parameter with low and high settings, while Responses processes images at high detail.
• Video input: Bedrock does not expose the video capability available in Moonshot’s open weights.
• Regional routing: US Geo routes inference among supported US regions. Global routing can use supported commercial AWS regions listed by AWS, so residency-sensitive workloads should confirm the permitted destination regions.

Bedrock adds about 76%

Published Global Standard pricing is $3.00 per million input tokens and $15.00 per million output tokens. Cache reads cost $0.30 per million tokens, while a 30-minute cache write costs $3.75. The US Geo Standard profile adds 10% to those rates.

OpenRouter lists Kimi K3 at $1.70 per million input tokens, $8.50 per million output tokens, and $0.17 per million cache-read tokens. Bedrock Global Standard therefore costs about 1.76 times as much for uncached input and output. Bedrock includes direct integration with AWS identity, logging, routing, billing, and support controls.

Flex costs half the Global Standard rate and suits batch jobs that can tolerate lower scheduling priority. Priority costs 1.75 times the Standard rate for latency-sensitive traffic. The OpenRouter comparison covers list prices only because routing, retention, support, and provider configurations can differ.

Call K3 with the OpenAI SDK

The OpenAI-compatible Responses API requires a Bedrock base URL, a region-scoped token, and an inference-profile model ID:

from aws_bedrock_token_generator import provide_token
from openai import OpenAI

region = "us-west-2"

client = OpenAI(
    api_key=provide_token(region=region),
    base_url=f"https://bedrock-runtime.{region}.amazonaws.com/openai/v1",
)

response = client.responses.create(
    model="global.moonshotai.kimi-k3",
    input="Summarize this repository.",
)

print(response.output_text)

The inference-profile IDs are us.moonshotai.kimi-k3 and global.moonshotai.kimi-k3. The caller needs configured AWS credentials and permission to invoke the model, including bedrock:InvokeModel for standard requests and bedrock:InvokeModelWithResponseStream for streaming.

Good fits and hard limits

Bedrock already carries models from Anthropic, Meta, and Mistral; K3 adds a large open-weight model with an unusually long context window and native vision. Whole-repository review, large-document extraction, and extended agent histories are plausible uses, especially when repeated prompt prefixes qualify for cache discounts.

Long context can still lose relevant details when prompts contain large amounts of loosely structured material. Production evaluations should test retrieval quality, tool selection, output consistency, time to first token, and total generation latency at realistic context lengths.

Input-heavy workloads gain the most from Bedrock’s cache pricing. One million cached input tokens cost $0.30 on Global Standard, while one million generated tokens cost $15.00. Repeated analysis of stable codebases or document sets therefore has a different cost profile from workloads that generate long reports or sustained agent output.

Commercial thresholds in the license

Moonshot distributes K3’s weights under a custom license that falls outside standard open-source definitions. Companies with more than $20 million in annual revenue must negotiate with Moonshot before offering K3 to external customers as a service. A separate attribution requirement applies to companies with more than $20 million in monthly revenue or more than 100 million monthly active users.

AWS retains and serves the weights for Bedrock requests. Product obligations still depend on AWS service terms, Moonshot’s current license, and how the model is exposed to customers. Companies near the revenue or user thresholds should review those terms before launch.