Kimi K3 on Amazon Bedrock Brings a 2.8T Open Model to AWS

Kimi K3 is now generally available on Amazon Bedrock, placing Moonshot AI's 2.8-trillion-parameter open-weight model inside AWS's managed inference, guardrails and data-residency controls. Amazon confirmed general availability for September 18, 2026. Moonshot describes K3 as its most capable system and the first open model to reach 2.8 trillion parameters, a claim that puts the Beijing lab at the top of the open-weight tier on raw scale.

The model is not new to the market. Moonshot launched K3 as a hosted service on July 16, 2026, then published downloadable weights in late July, close to the schedule the company had communicated. What changes with the Bedrock listing is who can realistically use it. Running a Mixture-of-Experts system of this size demands cluster capacity that few corporate buyers can justify for a single workload, which makes managed cloud distribution the practical route to enterprise adoption.

Open weights mean the trained parameters can be downloaded and modified, but they do not make a model cheap to run. Moonshot's own guidance steers users toward hosted endpoints for anything beyond experimentation, and the July weight release was aimed at research teams with their own clusters. For everyone else, the choice was Moonshot's API or nothing.

Why a 2.8-Trillion-Parameter Model Needs a Cloud

K3 carries 2.8 trillion total parameters but activates roughly 104 billion of them per token through its Mixture-of-Experts design. Only a slice of the network runs for any given request, so inference cost tracks active parameters rather than total ones. Training and hosting the full weight set remains an infrastructure problem measured in data-center floors rather than servers.

Bedrock absorbs that problem. Customers reach K3 through the same managed endpoints, identity controls and regional data-residency guarantees they already apply to other models on the service. AWS has built this catalogue deliberately: Moonshot's Kimi K2.5 arrived on Bedrock in January 2026, followed by a reasoning variant, so K3 extends an existing relationship rather than opening a new one.

That continuity shapes procurement. Enterprises that already cleared K2.5 through security and compliance review face a shorter path to approving K3, because the vendor, the data-handling posture and the contracting vehicle are unchanged. A model from an unfamiliar lab would restart that review from the beginning.

The gap between the mid-July hosted launch and the Bedrock listing is typical for this kind of integration. Weight publication, serving optimisation and a security review of the model and its supplier all sit between a launch event and a managed endpoint. K3 cleared that sequence faster than many open releases do.

What Kimi K3 on Amazon Bedrock Changes for Buyers

Two specifications drive the enterprise pitch. The first is a context window of 1,048,576 tokens, which lets a single session hold codebases, contracts or research archives that would otherwise need chunking and retrieval scaffolding. The second is native vision, so the model reads screenshots, diagrams and scanned documents without a separate pipeline bolted on.

A window that size is roughly several hundred thousand words of prose, or a mid-sized codebase with its documentation attached. Retrieval pipelines add their own failure modes: chunk boundaries split relevant passages, embedding search misses exact identifiers, and the assembly step consumes engineering time. Holding the corpus in the window removes a moving part, at the cost of paying for those tokens on every call.

Moonshot and AWS both aim K3 at coding and knowledge work. Those are the two workloads where long context converts directly into fewer engineering hours: a repository-wide refactor or a multi-document compliance review becomes one request instead of dozens of orchestrated calls.

Explicit prompt caching addresses the cost side. Caching reuses the processed prefix of a prompt across calls, cutting both latency and input-token spend on repeated queries against the same document set. For agentic loops that re-read a large context on every step, that is the difference between a viable budget and an unworkable one.

Budgeting is the reason caching receives top billing in AWS's description of the launch. A model that reprocesses a million-token context on every call burns input tokens at a rate few finance teams will sign off on. Cache hits on a stable prefix change the arithmetic, and in document-heavy pipelines the prefix is almost always stable.

Data-residency controls matter most to regulated buyers. Banks, healthcare providers and public-sector agencies that cannot send documents to an unvetted endpoint can run K3 inside the same regional boundary as their other Bedrock workloads, which is often the deciding factor in whether a model is approved at all. Guardrails apply the same way: administrators can filter inputs and outputs against policy, and content controls plus personally identifiable information detection cover K3 as they cover every other hosted model.

Cost Per Token Is Now the Selection Criterion

AWS frames the addition around routing rather than loyalty. With frontier-class open weights sitting beside proprietary models on one service, buyers can send a task to whichever model clears the quality bar at the lowest cost and latency, then switch without renegotiating a contract or rebuilding an integration. On Bedrock that switch is a model identifier in an API call. The low switching cost makes the routing argument credible, and it also means the launch of Kimi K3 on Amazon Bedrock has to win on measured quality and price rather than on lock-in.

Routing in practice starts with an evaluation harness rather than a procurement decision. Teams run the same task set across several Bedrock endpoints, compare accuracy against token cost and latency, and assign each workload to the model that clears the bar. Because K3 sits behind the same API surface as the rest of the catalogue, adding it to that comparison costs an afternoon rather than a project.

Parameter count is a marketing number as much as a technical one, and Moonshot's framing leans on it. The company bills K3 as the largest open model yet, says it outperforms some leading US systems, and places it just behind the top Anthropic models.

Benchmark comparisons between open and closed models are noisy. Reported scores depend on prompt formats, sampling settings and which version of a competitor was tested, and vendors rarely publish the full harness. The useful signal for a buyer is a task-level test on their own data, run through the same endpoint and guardrails they intend to use in production.

Scale also invites scrutiny. Moonshot is a Beijing-based lab, one of China's so-called AI Tigers, and it has been negotiating with Microsoft and Google as well as Amazon to place K3 on US clouds. A multi-cloud strategy serves two purposes: it reaches Western enterprise buyers who will not procure directly from a Chinese vendor, and it spreads the infrastructure burden of serving a model this large.

The company's corporate trajectory adds urgency to the distribution push. Moonshot has filed for a Hong Kong listing that would value it near $50 billion, a figure that depends on converting research prestige into recurring enterprise revenue. Cloud marketplaces are the fastest route to that revenue because they attach K3 to budgets that are already approved.

Why This Matters

For enterprise buyers, the practical effect of Kimi K3 on Amazon Bedrock is that a frontier-scale open model becomes a line item rather than a capital project. Teams can test K3 against their own coding and document workloads without provisioning hardware, and prompt caching makes repeated long-context calls predictable enough to budget.

For the wider market, the listing confirms that open-weight distribution is now a cloud competition rather than a licensing question. AWS is betting that hosting the largest open models alongside proprietary ones keeps enterprise inference spend on its platform, and Moonshot is betting that managed access is the only way a 2.8-trillion-parameter model reaches corporate customers at all.

Sources

Kimi K3 by Moonshot AI is now generally available on Amazon Bedrock

Introducing Kimi K3 on Amazon Bedrock

Related Articles

• Moonshot AI's Kimi K3 Open Weights Reshape the Frontier Model Economics
• Google DeepMind Brings Gemma 4 Open Models to Amazon Bedrock for Enterprise AI
• Amazon Bedrock Integrates OpenAI GPT OSS and NVIDIA Nemotron to Diversify Enterprise AI Options

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.