September 22, 2026
Kimi K3 is available through Amazon Bedrock: $3 per million input tokens
Kimi K3 has been available in Amazon Bedrock since September 18, 2026: a million input tokens cost $3, while output tokens cost $15.
Kimi.ai
@kimi_moonshot
Kimi K3 is now on Amazon Bedrock! Work with code, analyze documents, and run long agentic workflows with Bedrock's access controls, encryption, and auditing. Explicit prompt caching is supported. Start building with K3 on AWS 👉 https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-moonshot-ai-kimi-k3.html
· 293.5K views
Kimi K3 became the first open-weights model in Amazon Bedrock where cache checkpoints are set manually.
Before September 18, 2026, you had to run such a model yourself or use third-party hosts. Now its 2,8 trillion parameters and 1 million-token context can be called with the same key as other Bedrock models, along with encryption and invocation logs. The model accepts text and images as input, and returns text only.
How to start. Create a long-term API key in the Bedrock console, install `pip install -U openai aws-bedrock-token-generator`, and call Responses or Chat Completions at the base_url `https://bedrock-runtime.{region}.amazonaws.com/openai/v1`. boto3 is enough for Invoke and Converse.
In the catalog, the model is listed under the id `moonshotai.kimi-k3`, but it responds only through the cross-region profiles `us.moonshotai.kimi-k3` and `global.moonshotai.kimi-k3`. In-Region does not work in any region.
You must set the cache manually. A checkpoint is accepted from 1 024 tokens and lives for at least 30 minutes. Cache reads cost $0.30 per million versus $3 for regular input, and this works only in Responses and Chat Completions.
The model also responds through Converse, but AWS warns of InternalServerException when reasoning from previous turns is included in a multi-turn request. This breaks the default configurations of LangChain and Strands Agents, while Converse discards embedded PDF and HTML files. The vendor recommends the OpenAI-compatible Responses and Chat Completions.
Pricing tiers affect the bill more than geography: Flex halves the rate, while Priority multiplies it by 1,75. Both tiers are available only in Responses and Chat Completions. Invoke and Converse remain on Standard.
The US profile covers five regions, while Global works in a few dozen, including Frankfurt, Tokyo, and Singapore. According to the AWS blog, the global profile is about 10% cheaper.
In July 2026, Moonshot rewrote its open-weights license: hosts with annual revenue above $20 million must pay. AWS clears that threshold comfortably; according to July 2026 data, Amazon's AI business was running at $25 billion annually.
Kimi K3's code figures are still vendor-reported: Terminal-Bench 2.1 88,3 and DeepSWE 67,5 await independent runs.
