Availability and Capabilities

Moonshot AI has launched Kimi K3 on Amazon Bedrock, making the open-weight model available for coding and knowledge work. AWS describes the release as part of its continued expansion of open-weight options on Bedrock, which has added dozens of models from providers including DeepSeek, Google, MiniMax, Mistral AI, NVIDIA, OpenAI, and Qwen since 2025. Kimi K3 is positioned as a capable option for long-running workflows that require sustained context across repositories, documents, and images. AWS states that it is building for a future in which organizations adopt open-weight innovation with the reliability and security required for production.

Technical Specifications and Efficiency

According to Moonshot AI, Kimi K3 is its most capable model and the first open model to reach 2.8 trillion parameters. It combines native vision with a 1-million-token context window and delivers an approximate 2.5x improvement in scaling efficiency over Kimi K2. These characteristics are intended to support tasks that demand extended context and visual understanding within a single model call, such as analyzing large codebases or multi-page documents alongside images.

Prompt Caching and Inference Optimization

Kimi K3 is the first open-weight model on Amazon Bedrock to support explicit prompt caching, which allows users to mark reusable prompt prefixes after at least 1,024 tokens by adding a prompt_cache_breakpoint to supported input content. Cached tokens are billed at a higher rate but remain in cache for at least 30 minutes, while matching subsequent requests receive discounted input-token pricing and do not count against input-tokens-per-minute quotas. AWS added support for tool calling, structured output, reasoning, response streaming, and the Responses and Chat Completions APIs in 2026, and these platform capabilities apply to new open-weight models as they become available on Bedrock.

Security and Data Handling

AWS states that data processed through Kimi K3 stays within the AWS data boundary, is not shared with the model provider, and is not used to train the underlying model. Zero data retention is always enabled for inference requests, and zero operator access prevents AWS operators from viewing prompts and completions during inference. These protections are presented as standard for all open-weight models on Bedrock and are intended to let organizations use open-weight models without changing their security posture.

Access and Integration Options

Users can test Kimi K3 in the Bedrock console under Test > Playground or invoke it programmatically via the bedrock-runtime endpoint, which supports OpenAI-compatible Responses and Chat Completions APIs as well as the Amazon Bedrock Invoke and Converse APIs. A quick example uses the OpenAI SDK and the aws-bedrock-token-generator library for Python to generate short-term bearer tokens for authentication. Cross-Region inference is available through the global profile global.moonshotai.kimi-k3, which routes requests to any supported commercial AWS Region worldwide and costs approximately 10% less than a geographic profile. The US geographic profile us.moonshotai.kimi-k3 keeps processing within the US geography for data residency needs. IAM permissions for bedrock:InvokeModel, bedrock:InvokeModelWithResponseStream, and bedrock:CreateInference are required.

Ecosystem and Agent Support

Kimi K3 can be used through coding assistants and agentic frameworks that support Amazon Bedrock or OpenAI-compatible providers. OpenCode, an open-source model-agnostic coding agent, includes a native Amazon Bedrock provider that uses the Converse API; users can configure the provider in user-level or project-level opencode.json files and select models with the /models command. Hermes Agent, an open-source productivity assistant, also supports Bedrock models, though its documentation mislabels Amazon Bedrock as AWS Bedrock. Users may need to set the AWS_PROFILE environment variable when using named profiles, or switch to an API key. AWS provided a samples repository for further examples, and a demonstration video shows Kimi K3 building a single-file browser-based game in OpenCode.

Sources

• AWS Machine Learning BlogIntroducing Kimi K3 on Amazon Bedrock