AWS brings Moonshot's Kimi K3 to Amazon Bedrock

A new distribution channel for Moonshot

Amazon Web Services announced Kimi K3's availability on Amazon Bedrock on September 18. The launch makes Moonshot AI's model accessible through AWS's managed model platform for coding and knowledge-work applications. It gives organizations already using AWS another way to evaluate a Chinese-developed model within their existing cloud infrastructure.

AWS describes native visual understanding and a one-million-token context window, intended for work spanning repositories, documents and images. These capabilities concern the amount and type of input a model can process; they do not guarantee that it will correctly retrieve every detail or complete a complex assignment without checking.

A July model reaches a new cloud channel

Moonshot introduced Kimi K3 on July 17 and announced the publication of its weights and technical report on July 27. The September Bedrock launch is therefore a distribution milestone for an existing model. Moonshot already offered access through its own application, coding tools and API before AWS added this managed route.

Moonshot describes K3 as a 2.8-trillion-parameter mixture-of-experts model. Its technical release says each token activates 16 of 896 routed experts, combining a large total model with sparse computation. Those figures describe different things: the total parameter count is not the number of parameters activated for every token. The company also released infrastructure components for expert parallelism, attention and agent environments alongside the weights.

For a customer using Bedrock, operating that infrastructure is AWS's responsibility. The customer buys inference through the cloud service instead of downloading the checkpoint and arranging its own serving cluster. Self-hosting remains a separate deployment decision, with different infrastructure costs and operational work.

What is available through Bedrock

The Bedrock model card lists text and image input with text output, excluding audio and video input. Streaming, structured outputs and client-side tool calling are supported. AWS recommends its OpenAI-compatible APIs.

Customers choose US or global cross-region inference. The US profile processes requests within supported US-geography regions; the global profile can route them more widely. A regional endpoint therefore does not guarantee single-region processing.

The price of long context depends on reuse

At the time of this update, AWS's model card lists Standard global pricing of $3 per million input tokens and $15 per million output tokens. Global cache reads are listed at $0.30 and cache writes at $3.75 per million tokens. These are different billing categories, so a low cache-read rate should not be mistaken for the price of all input.

Prompt caching reuses previously processed context, such as a stable document or instructions. AWS documents explicit checkpoints of at least 1,024 tokens and retention of at least 30 minutes for Kimi K3. Repeated requests must actually hit the cache to realize its benefit; rewriting the reusable prefix or sending unrelated documents changes the economics.

Service tiers do not work through every API

AWS now lists three service tiers for K3: Standard, Priority and Flex. Priority is billed at 1.75 times the applicable Standard token rate, while Flex is billed at half that rate for less time-sensitive work. The distinction is available through the Responses and Chat Completions APIs; Converse and Invoke remain Standard-only for this model. A customer changing only the requested tier while keeping an unsupported API would not obtain the intended configuration.

The model identifiers also make the routing choice explicit: us.moonshotai.kimi-k3 selects the US cross-region profile, while global.moonshotai.kimi-k3 selects the global profile. At the checked Standard rates, US uncached input is $3.30 and output is $16.50 per million tokens, compared with $3 and $15 globally. The choice therefore changes both the permitted processing geography and the price used as the basis for tier calculations.

What the announcement establishes

The concrete development is availability on another major cloud platform. It does not establish an independently verified performance lead or disclose the commercial revenue split between AWS and Moonshot. Teams comparing this option with other hosted models still need to measure answer quality, completed-task cost and processing geography on representative work.