The Pulse
Amazon Bedrock Adds Moonshot AI’s Kimi K3
Amazon Bedrock now offers Moonshot AI’s Kimi K3, an open-weight multimodal model with a one-million-token context window. AWS lists Chat Completions and Responses access, US Geo and Global cross-Region inference, and token-based pricing.
AI.info Team ·
A one-million-token model enters Bedrock
Amazon Web Services lists Moonshot AI’s Kimi K3 as available through Amazon Bedrock, giving developers managed access to an open-weight model built for long coding and knowledge workflows. AWS’s model card identifies September 18, 2026, as Kimi K3’s launch date and describes it as Moonshot AI’s most capable open-weight model.
Kimi K3 combines native vision with a one-million-token context window. That capacity allows a single workflow to retain large repositories, long documents and images without repeatedly dividing the material into smaller prompts. AWS presents the model as a fit for applications that need to work across substantial text and visual inputs.
The Bedrock listing supports the bedrock-runtime endpoint and recommends the Chat Completions API for Kimi K3. The model also supports Responses, Invoke and Converse, although AWS documents restrictions that make the OpenAI-compatible APIs the preferred route for new integrations.
US Geo and Global routing define access
Customers can call Kimi K3 with the model identifier moonshotai.kimi-k3 through Bedrock’s runtime service. AWS provides two cross-Region inference profiles: us.moonshotai.kimi-k3 for US Geo routing and global.moonshotai.kimi-k3 for requests that may be routed across supported commercial AWS Regions worldwide.
The US Geo option is designed for workloads that require requests to remain within the US geography. The Global profile provides broader routing across AWS’s supported commercial regions. Kimi K3 does not offer in-Region inference in the regions listed on the model card, so customers must choose between those cross-Region options.
AWS lists support across North American, European, Asian-Pacific, Middle Eastern, African and South American regions through the Global profile. The precise availability varies by region, making the model card the relevant reference for deployment planning rather than a general Bedrock model catalogue.
Pricing starts at $3 per million input tokens
Standard-tier pricing is $3 per million input tokens and $15 per million output tokens through Global Cross-Region Inference. The US Geo option costs $3.30 per million input tokens and $16.50 per million output tokens.
Bedrock also lists prompt-cache pricing for Kimi K3. Global cache reads cost $0.30 per million tokens and cache writes with a 30-minute retention period cost $3.75. US Geo cache reads cost $0.33 per million tokens, while cache writes cost $4.125 per million tokens.
Priority service is billed at 1.75 times the Standard rate, while Flex service costs half the Standard rate. AWS says those service tiers are available through the Responses and Chat Completions APIs; Converse and Invoke remain limited to Standard on-demand inference.
API limits matter for multi-turn applications
The Bedrock documentation warns that Converse has known limitations with Kimi K3. A multi-turn request can fail with an InternalServerException when reasoning content from earlier turns is included, an issue AWS says can affect frameworks such as LangChain and Strands Agents in their default configurations.
Converse also rejects attached PDF and HTML documents for this model. AWS says developers who need multi-turn Converse requests should remove reasoning blocks from earlier turns, or use the Responses or Chat Completions APIs instead.
Video inputs are not supported. For combined image-and-text prompts, AWS recommends testing image blocks before text blocks because that ordering can produce higher-quality answers, while the image-detail control is honored only through Chat Completions.
Kimi K3 joins an expanding open-weight catalogue
The addition follows Bedrock’s February expansion with six open-weight models, including Moonshot AI’s Kimi K2.5. AWS has also published deployment guidance for running Kimi K3 on SageMaker HyperPod and Amazon EKS, describing the model as a 2.8-trillion-parameter mixture-of-experts system with 896 specialist experts and 104 billion active parameters per token.
Those self-hosting requirements differ sharply from the Bedrock path. AWS’s deployment guide calls for an instance with eight NVIDIA B300 GPUs, while the Bedrock model card presents Kimi K3 as a managed API with usage-based pricing and AWS-controlled inference capacity.
For developers, the immediate distinction is operational: Bedrock supplies a model ID, regional routing profiles, APIs and service tiers without requiring them to assemble the infrastructure described in AWS’s self-hosting guide. The tradeoff is that application teams must account for the documented API restrictions, cross-Region routing choices and higher US Geo rates before moving production workloads to Kimi K3.