Amazon Bedrock Adds Moonshot AI's Kimi K3 with Native Explicit Prompt Caching

AWS has made Moonshot AI's Kimi K3 model generally available on Amazon Bedrock via US Geo and Global cross-Region inference profiles. Featuring 2.8 trillion parameters and native vision support, Kimi K3 handles up to a 1-million-token context window with a reported 2.5x scaling efficiency improvement over its predecessor. Crucially, Kimi K3 is the first open-weight model on Amazon Bedrock to support explicit prompt caching alongside standard platform capabilities such as tool calling, structured outputs, response streaming, and OpenAI-compatible Chat Completions and Responses APIs. The availability of Kimi K3 directly affects enterprise architects, platform engineers, and AI development leads who manage high-context generative workloads such as automated codebase migration, full-repository code reviews, and multi-document synthesis. Previously, deploying massive-context open models required specialized self-hosted GPU clusters or proprietary SaaS endpoints that raised data sovereignty concerns. Running Kimi K3 natively inside Bedrock provides zero operator access and zero data retention, ensuring enterprise prompts and completions remain strictly confined to the customer's cloud boundary without third-party exposure. This release reflects the broader cloud infrastructure trajectory toward standardizing open-weight foundational models on managed hyperscaler platforms. Rather than treating open models as second-class compute tasks requiring manual orchestration on containerized clusters, platforms like Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure are turning them into fully managed, serverless primitives. The inclusion of platform-level capabilities like explicit prompt caching and cross-region routing brings open-weight inference economics into direct competition with proprietary frontier models, intensifying commoditization at the inference layer. In practice, technical practitioners should evaluate Kimi K3 for batch and agentic workloads where repetitive system prompts, extensive schemas, or massive codebases previously inflated API expenses. Implementing explicit prompt caching checkpoints (with the standard 1,024-token threshold and 30-minute retention) can significantly reduce both time-to-first-token (TTFT) and input costs for repetitive multi-turn agent interactions. Teams should benchmark latency across US Geo versus Global cross-Region inference endpoints and verify their Bedrock IAM policies to take advantage of the Priority and Flex service tier pricing structures based on workload criticality.