Moonshot AI’s Kimi K3 Arrives on Amazon Bedrock With 1M-Token Context

Moonshot AI’s Kimi K3, an open-weight model its developer describes as the first open model to reach 2.8 trillion parameters, became available on Amazon Bedrock on September 18, 2026, adding a new option for coding and knowledge work.

Amazon Web Services announced the launch on the AWS Machine Learning Blog. According to Moonshot AI, Kimi K3 is its most capable model, combining native vision capabilities with a 1-million-token context window and delivering an approximate 2.5x improvement in scaling efficiency over Kimi K2.

In Moonshot AI’s Kimi K3 technical post, the company says the model is built on two architectural updates it calls Kimi Delta Attention and Attention Residuals, designed to improve how information flows across sequence length and model depth. Kimi K3 uses a mixture-of-experts design the company calls Stable LatentMoE, effectively activating 16 of 896 experts, and applies quantization-aware training from the supervised fine-tuning stage onward, using MXFP4 weights with MXFP8 activations. The company attributes the roughly 2.5x scaling-efficiency gain over its earlier model to these structural changes together with refined training and data recipes.

Moonshot AI states that Kimi K3’s overall performance still trails the proprietary Claude Fable 5 and GPT 5.6 Sol models, while reporting that Kimi K3 delivered frontier-level results across its evaluation suite and consistently outperformed other tested models. The company also lists limitations: the model was trained in a preserved thinking history mode, so generation quality can become unstable if an agent harness fails to pass back historical thinking content, and its emphasis on long-horizon tasks can lead it to make unexpected decisions on a user’s behalf when instructions are ambiguous, which the company says may require more explicit behavioral constraints in the system prompt.

Moonshot AI originally introduced Kimi K3 in July 2026, stating at the time that the full model weights would be released by July 27, 2026. At that introduction, the model was offered through Moonshot AI’s own channels: the Kimi app, the Kimi Work desktop application, the Kimi Code terminal tool, and the Kimi API.

Open-Weight Models and Data Handling on Amazon Bedrock

AWS said the launch reflects sustained investment in open-weight models on the service. Since 2025, Bedrock has added dozens of open-weight models from providers including DeepSeek, Google, MiniMax, Mistral AI, Moonshot AI, NVIDIA, OpenAI, and Qwen. In 2026, Bedrock added support for tool calling, structured output, reasoning, response streaming, and the Responses and Chat Completions APIs as platform capabilities rather than per-model integrations, so new open-weight models can use them as they become available.

AWS states that, as with all open-weight models on Amazon Bedrock, customer 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 AWS operators from accessing prompts and completions during inference.

Amazon Bedrock documentation lists Kimi K3 as one of three Moonshot AI models in the service, alongside Kimi K2.5, a multimodal model with improved reasoning, coding, and multilingual capabilities, and Kimi K2 Thinking, a reasoning model with chain-of-thought capabilities for complex problem solving in math, coding, and logic.

Console Access, APIs, and Inference Profiles

Users can try Kimi K3 in the Amazon Bedrock console under Test > Playground, or call it programmatically through the bedrock-runtime endpoint. The endpoint supports the OpenAI-compatible Responses and Chat Completions APIs as well as the Amazon Bedrock Invoke and Converse APIs.

The model is invoked through cross-Region inference profiles. For workloads without regional restrictions, AWS recommends the global profile, global.moonshotai.kimi-k3, which routes each request to any supported commercial AWS Region worldwide and, according to AWS, costs approximately 10% less than a geographic profile. The US geographic profile, us.moonshotai.kimi-k3, keeps processing within the US geography for data residency requirements.

Prerequisites listed by AWS include an active AWS account with Amazon Bedrock access, Python 3.10 or later, and AWS Identity and Access Management permissions for bedrock:InvokeModel, bedrock:InvokeModelWithResponseStream, and bedrock:CreateInference.

Explicit Prompt Caching and Developer Tooling

According to AWS, Kimi K3 is the first open-weight model on Amazon Bedrock to support explicit prompt caching, which targets long-running coding and knowledge workflows that repeatedly resend stable context such as repository instructions, tool definitions, or reference documents. A promptcachebreakpoint marker can designate the exact end of a reusable prompt prefix after at least 1,024 tokens. In explicit mode, tokens written to cache are billed at a higher rate but are kept in cache for at least 30 minutes. Subsequent requests that match a cached prefix are billed at a discounted input rate and do not count against input-tokens-per-minute quotas.

Beyond direct API use, the model works through tooling that supports Amazon Bedrock. AWS highlights OpenCode, an open-source and model-agnostic coding agent with a native amazon-bedrock provider that uses the Converse API, and Hermes Agent, an open-source productivity assistant that natively supports models on Amazon Bedrock. AWS has also published a Moonshot AI on AWS samples repository on GitHub with additional examples.

Kimi K3 is available through the US Geo and Global cross-Region inference profiles, with the full list of supported Regions in the Bedrock documentation.