Kimi K3 Lands on Amazon Bedrock, Two Months After Moonshot AI’s Own Launch

AWS announced on September 18, 2026, that it's introducing Kimi K3 on Amazon Bedrock, giving developers managed access to Moonshot AI's largest open-weight model without hosting it themselves. The post, credited to AWS's Alex Thewsey, Sofian Hamiti, Tanvi Girinath, Saurabh Trikande, and William Yap, frames it as part of a broader Bedrock strategy: since 2025, the platform has added dozens of open-weight models from providers including DeepSeek, Google, MiniMax, Mistral AI, Moonshot AI, NVIDIA, OpenAI, and Qwen.

One clarification worth making up front: Kimi K3 itself is not new. Moonshot AI launched it through its own Kimi products and API on July 16, 2026, with the full open weights following on July 27, 2026, according to contemporaneous coverage from outlets including The Next Web and DataNorth. What's new today is that AWS customers can now call it through Bedrock's managed infrastructure, with the security posture and API compatibility that implies, rather than through Moonshot's own endpoints or a self-hosted deployment.

What AWS Actually Announced

Amazon Bedrock now offers Kimi K3 through two cross-Region inference profiles: global.moonshotai.kimi-k3, which AWS says routes requests to any supported commercial Region and costs approximately 10% less than a geographic profile, and us.moonshotai.kimi-k3, which keeps processing inside the US for data residency needs. AWS says Kimi K3 is the first open-weight model on Bedrock to support explicit prompt caching, a feature previously available for Bedrock's own first-party models.

AWS's stated security posture for the model matches what it offers other open-weight models on Bedrock: your data stays inside the AWS data boundary, isn't shared with Moonshot AI, and isn't used to train the underlying model. AWS says zero data retention is always enabled for inference requests, and zero operator access is meant to prevent AWS's own staff from viewing prompts and completions during inference.

What Kimi K3 Actually Is

According to Moonshot AI, Kimi K3 is its most capable model and the first open model to reach 2.8 trillion total parameters. It's a sparse mixture of experts design that activates 16 of 896 experts per token, built around what Moonshot calls Kimi Delta Attention, a hybrid linear attention architecture, plus a technique it calls Attention Residuals. Moonshot attributes an approximate 2.5x improvement in scaling efficiency over Kimi K2 to these changes, a company claim rather than an independently reproduced figure.

The model carries a 1,048,576 token (roughly 1 million) context window and native vision, meaning it can process text and images in the same request. That combination, per AWS's post, is aimed at long-running coding and knowledge work that needs to hold large repositories, documents, or image sets in context across a session.

Independent coverage from its July launch adds detail AWS's Bedrock post doesn't include. The Next Web reported that Moonshot positioned K3, on the company's own benchmark results, as ranking behind only Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol. DataNorth's reporting, citing the Artificial Analysis Intelligence Index, an independent third party benchmark aggregator, put K3 fourth of 189 tracked models with a score of 57, level with Claude Opus 4.8 and GPT-5.5. Those two rankings aren't necessarily contradictory since they likely reflect different benchmark methodologies, but neither is the kind of controlled, single-harness comparison that would let you treat either ranking as settled. Industry outlet implicator.ai noted at launch that Moonshot's own benchmark table mixed agent harnesses (KimiCode, Claude Code, Codex) depending on which score was being reported, which limits direct model to model comparison.

Kimi K2 vs. Kimi K3 on Bedrock

|

|Kimi K2 Thinking | Kimi K3 |
|—|—|—|
|Context window | 256K tokens | 1M tokens |
|Native vision | Not specified in Bedrock documentation | Yes |
|Explicit prompt caching on Bedrock | Not available | Available |
|Bedrock launch date | November 6, 2025 | September 18, 2026 |

Getting Started on Bedrock

AWS lists three prerequisites: an active AWS account with Bedrock access, Python 3.10 or later, and IAM permissions covering bedrock:InvokeModel, bedrock:InvokeModelWithResponseStream, and bedrock:CreateInference. From the Bedrock console, you can reach Kimi K3 under Test, then Playground. Programmatically, AWS's example uses the OpenAI Python SDK together with the aws-bedrock-token-generator library to authenticate against the bedrock-runtime endpoint, which supports both the OpenAI-compatible Responses and Chat Completions APIs alongside Bedrock's own Invoke and Converse APIs.

For explicit prompt caching specifically, AWS says you mark the end of a reusable prompt prefix, after a minimum of 1,024 tokens, with a prompt_cache_breakpoint. Tokens written to cache are billed at a higher rate but stay cached for at least 30 minutes; requests that hit a matching cached prefix get input tokens billed at a discount and those tokens don't count against input-tokens-per-minute quotas.

Where Developers Will Actually Use This

AWS's post highlights two agent tools as integration examples rather than as AWS products. OpenCode, an open source, model-agnostic coding agent, has a native amazon-bedrock provider built on the Converse API, configurable through a project or user-level opencode.json file. Hermes Agent, an open source productivity assistant from Nous Research, natively supports Bedrock models through its terminal setup flow, though AWS notes Hermes currently labels the provider "AWS Bedrock" rather than "Amazon Bedrock" in its interface, and that named AWS credential profiles require an environment variable workaround at the time of writing.

Why It Matters

The more interesting story here isn't Kimi K3's specs, which have been public since July. It's what AWS keeps doing with its open-weight model shelf. Bedrock has spent the past year adding tool calling, structured output, reasoning support, streaming, and OpenAI-compatible APIs as platform-level capabilities rather than per-model integrations, which is exactly why a two-month-old Chinese open-weight model can show up on Bedrock with a full feature set on day one instead of a stripped-down version.

That pattern sits inside a bigger, more contested question BriefFlash has covered repeatedly this year: enterprises are already in a shaky spot as AI slowdown warnings collide with cheap, capable open-weight models out of China, and Moonshot AI specifically has come up in Anthropic's own reporting on distillation activity tied to Chinese labs. AWS putting Kimi K3 behind its own security boundary, with data residency options and zero operator access, is a direct answer to the trust gap that keeps coming up around these models: it's a way to get the capability and price point of a frontier-scale Chinese open-weight model without routing your prompts through Moonshot's own infrastructure. It's the same basic move Vercel made putting GLM 5.3 FlashX on its AI Gateway: Western infrastructure providers increasingly compete on being the trusted front door to Chinese open-weight models, not on building rival models themselves.

What to Watch

Watch for independent, single-harness benchmark results now that K3's full weights have been public since late July, since the rankings circulating today still mix vendor-run tests with early third-party index scores that don't fully agree with each other. Also watch Bedrock's per-token pricing for Kimi K3 specifically once AWS publishes it on its pricing page, since AWS's post doesn't include a number and Moonshot's own API pricing has shifted with nearly every point release this year.

Key Takeaways

• AWS made Kimi K3 available on Amazon Bedrock on September 18, 2026, through global and US-only cross-Region inference profiles
• Kimi K3 itself launched through Moonshot AI's own products on July 16, 2026, with full open weights following July 27, 2026; the Bedrock listing is a platform integration, not the model's debut
• Moonshot AI describes Kimi K3 as a 2.8 trillion parameter mixture of experts model (16 of 896 experts active per token) with a 1 million token context window and native vision
• Kimi K3 is the first open-weight model on Bedrock with explicit prompt caching, which discounts input tokens on repeated prompt prefixes cached for at least 30 minutes

FAQ

Is Kimi K3 a new model, or is Amazon just now adding it to Bedrock?

Kimi K3 is not new. Moonshot AI launched it through its own Kimi products and API on July 16, 2026, with full open weights following on July 27, 2026, according to reporting at the time from outlets including The Next Web and DataNorth. AWS's September 18, 2026 post covers Kimi K3 becoming available through Amazon Bedrock specifically, roughly two months after the model's original release.

How much does Kimi K3 cost on Amazon Bedrock?

AWS's announcement doesn't list per-token pricing directly and points readers to the Amazon Bedrock pricing page instead. AWS does confirm that using the global cross-Region inference profile costs approximately 10% less than using a geographic (US-only) profile, and that explicit prompt caching bills cached tokens at a higher rate upfront but discounts input tokens on later requests that hit the cache.

What makes Kimi K3 different from Kimi K2?

According to Moonshot AI, Kimi K3 scales up to 2.8 trillion total parameters using a sparse mixture of experts design (16 of 896 experts active per token), versus Kimi K2's roughly 1 trillion parameters. It adds native vision and a 1 million token context window, up from Kimi K2 Thinking's 256,000 token context window as listed in AWS's Bedrock documentation, and Moonshot claims roughly 2.5 times better scaling efficiency, a company figure not yet independently verified.

How do you access Kimi K3 on Amazon Bedrock?

You need an AWS account with Bedrock access, Python 3.10 or later, and IAM permissions for bedrock:InvokeModel, bedrock:InvokeModelWithResponseStream, and bedrock:CreateInference. From there you can test it in the Bedrock console's Playground, or call it programmatically through the bedrock-runtime endpoint using either Bedrock's native Invoke and Converse APIs or the OpenAI-compatible Responses and Chat Completions APIs.