Moonshot AI’s Kimi K3 is now available through Amazon Bedrock for coding and knowledge-work applications. AWS says the open-weight model supports images, accepts up to one million tokens of context and is the first open model on Bedrock with explicit prompt caching.
Prompt caching lets developers mark a stable prefix of at least 1,024 tokens, such as repository instructions or tool definitions. Writing that prefix to cache costs more initially, but matching requests can reuse it for at least 30 minutes at a lower input-token price and without consuming the normal input-token-per-minute quota. The feature targets long-running workflows that repeatedly send the same large context.
Kimi K3 can be called through Bedrock’s native Invoke and Converse interfaces or OpenAI-compatible Responses and Chat Completions APIs. A global inference profile can route requests among supported commercial AWS regions and costs about 10 percent less than a geographic profile; a separate US profile keeps processing within the United States. AWS says request data remains inside its boundary, is not shared with Moonshot AI and is not used to train the model. Availability begins through those cross-region profiles, so developers with stricter regional requirements should check the supported-region list and pricing before adopting it.