Moonshot AI's Kimi K3 is now available on Amazon Bedrock, and according to the developer it is the first open model to reach 2.8 trillion parameters. The model pairs native vision with a 1-million-token context window and, per Moonshot AI, delivers roughly 2.5x better scaling efficiency than Kimi K2. For companies running coding and knowledge workflows over large repositories, documents and images, that combination sets a new baseline for what an open-weight model is expected to handle in production.
What the model brings to Bedrock
Kimi K3 is positioned by Moonshot AI as its most capable model to date, and the parameter count is the headline figure: 2.8 trillion, a scale previously associated with closed systems. The 1-million-token context window is what makes long-running tasks practical — a repository, a document set or a batch of images can stay in context across a session instead of being re-fed in fragments. Native vision means image input is handled by the model itself rather than by a separate pipeline. Kimi K3 is also the first open-weight model on Amazon Bedrock to support explicit prompt caching, which cuts latency and input token costs when the same context is reused across calls.
Explicit prompt caching works by letting the developer mark reusable prompt prefixes — repository instructions, tool definitions, reference documents — so that later requests matching a cached prefix are served from cache. On the infrastructure side, the model is invoked through a cross-Region inference profile. The global profile, global. moonshotai. kimi-k3, routes each request to any supported commercial AWS Region and costs approximately 10% less than a geographic profile; the US profile, us. moonshotai. kimi-k3, keeps processing inside the US for data residency requirements. Calls go through the bedrock-runtime endpoint, which supports the OpenAI-compatible Responses and Chat Completions APIs as well as the Amazon Bedrock Invoke and Converse APIs.
The launch sits inside a broader build-out of open-weight support on Bedrock. Since 2025 the platform has added dozens of open-weight models from DeepSeek, Google, MiniMax, Mistral AI, Moonshot AI, NVIDIA, OpenAI and Qwen. In 2026 Bedrock added tool calling, structured output, reasoning, response streaming and the Responses and Chat Completions APIs. Because these are platform capabilities rather than per-model integrations, each new open-weight model inherits them on arrival — which is why the value of a catalogue like this grows with its size rather than with any single release.
What this means for business
For teams choosing a model, the practical change is that a 1-million-token context and vision no longer require a closed provider. Coding assistants, personal agents and agentic frameworks that support Bedrock or OpenAI-compatible providers can point at Kimi K3 — OpenCode, for instance, ships a native amazon-bedrock provider built on the Converse API, and Hermes Agent supports Bedrock models for research and task automation. A small team gets frontier-scale context without negotiating a separate vendor contract; a larger organisation gets it inside an AWS account it already governs, with model selection handled by configuration rather than procurement.
The data-handling terms are the part to verify before committing. AWS states that data is processed 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 even AWS operators from seeing prompts and completions during inference. What this does not settle is cost at volume — prompt caching reduces input token costs on reused context, but pricing should be checked against actual traffic patterns. Questions worth putting to the vendor: which Regions the chosen profile covers, how caching behaves with your prompt structure, and whether your residency rules allow the global profile or require the US one.
The marker to watch is adoption of the global profile against the US geographic one. If the 10% cost difference pushes most production traffic to global. moonshotai. kimi-k3, it will show that buyers treat data residency as a per-workload decision rather than a blanket policy — and that open-weight models at this scale have moved from evaluation into routine deployment.
