AWS Adds 2.8T-Parameter Kimi K3 to Bedrock with Prompt Caching Support

Moonshot AI's Kimi K3 model is now generally available on Amazon Bedrock through US Geo or Global cross-Region inference profiles, as AWS does not list a single-Region option. [4] The launch introduces a 2.8-trillion-parameter mixture-of-experts model that activates 104 billion parameters during inference. [4] According to Moonshot AI, this model is its most capable offering to date. [2] Kimi K3 supports native vision capabilities and a one-million-token context window, processing image and text inputs to return text outputs. [1][4] The developer reports that Kimi K3 delivers approximately 2.5 times better scaling efficiency than its predecessor, Kimi K2. [1][4]

The model is the first open-weight offering on Amazon Bedrock to support explicit prompt caching. [1] Developers can set explicit cache controls using the Responses and Chat Completions APIs, requiring a minimum of 1,024 tokens for each cache checkpoint. [4] Cached content remains available for at least 30 minutes, which AWS notes can improve cache-hit rates while reducing latency and costs when reusing long prompt prefixes. [4] For Global cross-Region inference on the Standard tier, AWS prices the model at $3 per million input tokens and $15 per million output tokens, while cache reads cost $0.30 per million tokens and cache writes cost $3.75 per million tokens. [4]

AWS asserts that Kimi K3 inference data remains within the AWS data boundary and is not shared with the model provider or used to train the underlying model. [2][4] The deployment utilizes zero data retention and zero operator access policies, preventing even AWS staff from accessing prompts and completions. [2][4] The integration builds on a broader expansion of open-weight models on Amazon Bedrock, which has added models from DeepSeek, Google, MiniMax, Mistral AI, NVIDIA, OpenAI, and Qwen since 2025. [2] Kimi K3 developers can also utilize platform-level inference capabilities added to Bedrock in 2026, including tool calling, structured output, reasoning, and response streaming. [2]

With its massive context window and native vision support, Kimi K3 offers a compelling enterprise alternative to proprietary systems like OpenAI o3 and Claude 3.5 Sonnet for intensive document processing tasks. The introduction of explicit prompt caching gives enterprise developers a crucial mechanism to mitigate the steep costs typically associated with large-context workflows. Furthermore, by utilizing cross-Region inference, AWS ensures the massive compute demands of a 2.8-trillion-parameter model do not bottleneck single availability zones. What the sources don't address: Whether Kimi K3's real-world cache-hit rates and latency improvements validate the added engineering complexity of managing explicit cache checkpoints.