What is Kimi K3 on Amazon Bedrock and when did it launch?

Kimi K3 from Moonshot AI became generally available on Amazon Bedrock on September 18, 2026. It is a 2.8-trillion-parameter open-weight model with a 1-million-token context window, native vision, and explicit prompt caching. It runs in all AWS Regions where Bedrock is available via cross-region inference profiles.

How does enterprise data security work with Kimi K3 on Bedrock?

AWS enforces a full data boundary: your prompts and completions stay within the AWS infrastructure, are not shared with Moonshot AI, and are not used to train the underlying model. Zero data retention is always enabled, and zero operator access prevents even AWS operators from reading your prompts during inference.

What is explicit prompt caching on Amazon Bedrock and why does it matter?

Explicit prompt caching lets developers mark stable content at the top of a prompt (after at least 1,024 tokens) so it does not need to be reprocessed on subsequent calls. Cached tokens are billed at a higher write rate but then cost less on every cache hit and do not count against token-per-minute quotas. The cache holds for at least 30 minutes. For agentic workflows that repeatedly resend system instructions or repository context, this can meaningfully reduce both latency and inference cost.

How does Kimi K3 relate to the earlier Kimi K3 open-weight release in July 2026?

The July 2026 release made Kimi K3 weights downloadable under an open license. The September 18, 2026 Bedrock launch adds a managed-inference path: enterprises can now run the same model through the Amazon Bedrock console and APIs without provisioning GPU clusters, and with the enterprise security controls (encryption, audit logging, private network access) already in place for proprietary models on the platform.

Kimi K3 on Amazon Bedrock: Enterprise Open-Weight AI

On September 18, 2026, Amazon Web Services announced that Moonshot AI’s Kimi K3 is generally available on Amazon Bedrock. The model is notable for two reasons. First, according to Moonshot AI, it is the first open model to reach 2.8 trillion parameters. Second, Kimi K3 is the first open-weight model on Amazon Bedrock to support explicit prompt caching, a feature that reduces both latency and inference cost for agentic and long-context workloads.

For enterprise leaders, this launch closes a gap that has defined the open-weight AI market for the past two years: the tension between model openness and enterprise governance. Hosting on Bedrock puts Kimi K3 inside the same security boundary as proprietary models, which changes the calculation for procurement and compliance teams.

What Kimi K3 Brings to Amazon Bedrock

The AWS Machine Learning Blog launch post details the model’s core specifications. Kimi K3 combines a 1-million-token context window with native vision and an approximately 2.5x improvement in scaling efficiency over Kimi K2. At 2.8 trillion parameters, it is the largest open-weight model available on any major managed inference platform as of its Bedrock launch date.

The model targets two primary use cases:

Long-running coding workflows. A 1-million-token context window lets the model hold an entire large repository in memory across a session. Engineers running iterative coding agents no longer need to reload context between tool calls or manage external memory systems for repository-scale projects.

Multi-document knowledge work. Native vision means the model can process scanned pages, screenshots, and diagrams without a separate OCR step. For knowledge workers in legal, financial services, or research, this enables multi-document analysis within a single model call.

Kimi K3 is accessible via the Amazon Bedrock console and programmatically through the bedrock-runtime endpoint. AWS supports both the OpenAI-compatible Responses and Chat Completions APIs and the native Bedrock Invoke and Converse APIs, which means teams that have already built against OpenAI-compatible interfaces can adopt the model with minimal code changes.

Kimi K3 Inference Profiles: Specifications at a Glance

Profile	Model ID	Use Case	Cost Note
Global	`global.moonshotai.kimi-k3`	No regional restrictions	~10% lower than US geo
US Geo	`us.moonshotai.kimi-k3`	US data residency required	Standard rate
Parameters	2.8T	Largest open-weight model on Bedrock	First at this scale
Context window	1M tokens	Code repos, multi-doc analysis	Native vision included
Prompt caching	Explicit mode	Agentic loops, repeated system prompts	Cache TTL: 30 min min
Scaling efficiency	2.5x vs K2	Long-horizon tasks	Moonshot AI reported

Source: AWS What’s New, September 18, 2026

Explicit Prompt Caching: Why It Matters for Enterprise Agentic Workflows

The prompt caching feature deserves specific attention because it directly addresses one of the dominant cost and latency problems in production agentic AI.

Most agentic workflows resend the same system instructions, tool definitions, or reference documents with every request. Without caching, the model reprocesses that context on each call, which inflates both cost and time-to-first-token. With explicit prompt caching, developers designate a reusable prefix by adding a prompt_cache_breakpoint at or after the 1,024-token mark. AWS then holds that prefix in cache for at least 30 minutes.

The billing mechanics create a useful trade-off: cached token writes are billed at a premium, but reads are discounted and exempt from input-token-per-minute quotas. For workflows that fire many calls against the same stable context, the math typically favors caching significantly. For one-off or low-frequency workloads, the premium on writes may not pay off.

Because this is the first open-weight model on Amazon Bedrock to support explicit caching, it also marks a shift in what enterprises can reasonably expect from open-weight inference infrastructure. Previously, prompt caching was effectively a proprietary-model feature on Bedrock. The Kimi K3 launch begins extending that capability to the open-weight tier.

The Enterprise Security Story: Open-Weight Without the Governance Risk

The defining structural advantage of Amazon Bedrock for enterprise teams is its unified security posture. As the AWS announcement states, Kimi K3 runs with the same controls as proprietary models on the platform:

Zero data retention: Inference request data is not persisted beyond the response.
Zero operator access: AWS operators cannot access prompts or completions during inference.
AWS data boundary: Prompts are processed inside AWS infrastructure and not transmitted to Moonshot AI.
Existing audit and encryption: The same access controls, private network access, and audit logging that enterprises use for closed models apply directly to Kimi K3.

This matters most for regulated industries. A healthcare organization using Bedrock for other workloads can evaluate Kimi K3 without obtaining separate data processing agreements with Moonshot AI. A financial services firm with a US data residency policy can use the us.moonshotai.kimi-k3 inference profile and maintain the same compliance posture as with proprietary models. Government contractors operating in GovCloud-adjacent architectures can assess the model within their existing cloud governance frameworks.

Open-weight models downloaded and self-hosted have always offered flexibility on weights. What they have rarely offered is a pre-existing security-certified deployment environment. Bedrock provides exactly that, and the Kimi K3 launch is the clearest example yet of the managed-inference channel for open models reaching frontier-model scale.

What This Means for Enterprise AI Strategy in Late 2026

The Kimi K3 Bedrock launch fits a broader pattern that enterprise AI leaders should track. As Enera previously covered when Kimi K3 first appeared in July 2026, the model’s architecture was designed for scaling efficiency from the start. Moonshot AI’s 2.5x efficiency gain over K2 at 2.8T parameters suggests the model was engineered to run at lower cost per token than a naive parameter-count comparison would imply.

On Amazon Bedrock, that efficiency improvement translates to more favorable unit economics for enterprise teams. Inference cost pressure is now a central concern for organizations that have moved from proof-of-concept to production AI deployment. The AI pricing dynamics this market has seen through mid-2026 put cost-efficient frontier-scale models in strong demand.

Three practical implications for enterprise architects:

For coding agent infrastructure: Kimi K3’s 1-million-token context plus explicit caching make it a plausible backend for coding agents that maintain session memory across large codebases. Teams already using Bedrock can test it through the Bedrock console without changing their cloud vendor relationship.

For multimodal document processing: Native vision at 2.8T parameters provides a credible alternative to proprietary multimodal models for document-intensive workflows where open-weight flexibility and model inspection matter.

For procurement and compliance teams: The managed-inference path on Bedrock removes the most common enterprise objection to open-weight models: that using them requires either exposing data to a third-party API or building and maintaining custom inference infrastructure. Both objections no longer apply.

What to Watch Next

AWS noted that Kimi K3 is one of three Moonshot AI models now on Bedrock, alongside Kimi K2.5 (a multimodal reasoning model) and Kimi K2 Thinking (a chain-of-thought reasoning model). The existence of a reasoning model variant in the same model family suggests Moonshot AI is building toward a model suite for Bedrock rather than a single offering.

The broader trend is the normalization of frontier-scale open-weight models in enterprise infrastructure. In 2025, open-weight models on managed inference platforms were exceptions. In September 2026, a 2.8-trillion-parameter open model is available with a few console clicks, inside an enterprise security boundary, with prompt caching that matches proprietary-model features. The gap between open and closed is closing faster than most enterprise roadmaps anticipated.

If your team is evaluating open-weight models for production agentic workloads, Kimi K3 on Bedrock is worth a direct comparison against your current stack. Enera’s team can help you design that evaluation.