Moonshot AI's Kimi K3 lands on Amazon Bedrock with 1M-token context and native vision

Moonshot AI's Kimi K3 is now available on Amazon Bedrock, making the 2.8-trillion-parameter open-weight model accessible through AWS's managed inference service.

What's new

AWS announced the launch plainly: "Today, Kimi K3 from Moonshot AI is available on Amazon Bedrock, giving you a powerful new option for coding and knowledge work."

Key specs of the Bedrock deployment:

• 2.8 trillion parameters, with native vision support built in rather than bolted on.
• 1-million-token context window, aimed at workflows that span large codebases, long documents, or many images at once.
• Roughly 2.5x improvement in scaling efficiency over Moonshot's prior-generation Kimi K2, according to AWS.
• Kimi K3 is described as the first open-weight model on Bedrock with explicit prompt caching support, a feature previously more common among closed frontier models.
• Available today through US Geo and Global cross-Region inference profiles.
• AWS highlights its standard Bedrock data-handling guarantees for the model: data is processed within the AWS boundary, not shared with the model provider, not used for training, with zero data retention and zero operator access available.

AWS frames the target use case directly: "Kimi K3 is well suited to long-running coding and knowledge workflows that require sustained context across large repositories, documents, and images."

Context

Kimi K3 is the latest release in Moonshot AI's Kimi line of open-weight models, which has increasingly competed with frontier labs on raw capability while remaining open-weight. Bringing it to Bedrock follows a broader pattern of major cloud platforms adding open-weight Chinese-lab models — alongside offerings from DeepSeek, Alibaba's Qwen, and others — to their managed model catalogs, giving enterprise customers access to those models without having to self-host or manage inference infrastructure directly.

For Moonshot, Bedrock distribution extends Kimi K3's reach into AWS's enterprise customer base, which typically cares as much about data-boundary guarantees and managed inference as about raw benchmark performance — exactly the assurances AWS is emphasizing in this launch.

Why it matters

A 1-million-token context window paired with native vision and prompt caching is a meaningful combination for enterprise coding and document-heavy workflows, and having it available through Bedrock's managed, data-boundary-controlled infrastructure removes a common blocker — self-hosting or data-residency concerns — that otherwise keeps large enterprises away from open-weight models. The "first open-weight model on Bedrock with explicit prompt caching" claim also signals AWS extending capabilities to open models that were previously differentiators mostly for closed frontier APIs, narrowing the operational gap between open and closed model deployment on major clouds.