Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- AWS announced availability on 6 November; the model card separately gives 4 November as the model launch date and marks the lifecycle Active. These are distinct source records and do not establish a GA or Preview label.
- The card supports the bedrock-runtime endpoint and recommends Chat Completions for new Kimi applications. The base model ID and the US, India and Global inference profile IDs are different.
- Published Standard rates are per million tokens, with different input, output and cache rates for Global and US/India. The stated Priority and Flex multipliers do not specify an India pricing basis.
- None of the listed source Regions supports In-Region inference. Converse has known multi-turn reasoning and PDF/HTML attachment limitations, while the 10x output-token quota burn is not a 10x bill.
Keep the announcement and model-card dates separate
On 6 November 2026, AWS announced that Moonshot AI's Kimi K3.1 Pro - Highspeed was available through Amazon Bedrock. The model card lists 4 November as its launch date and marks its lifecycle Active; these dates describe separate records and do not establish a GA or Preview label or the precise first moment the service became available.
AWS's model card describes it as an open-weight model from Moonshot AI and relays claims about native image capability and a 1-million-token context window. Those are vendor descriptions, not benchmarks conducted for this article or a guarantee of quality or speed on a particular task.
Choose a Runtime API, then confirm the model ID
The model card supports bedrock-runtime and recommends Chat Completions for new Kimi applications; Responses, InvokeModel and Converse are also listed as supported APIs. bedrock-mantle is not supported, and the card lists no supported In-Region option.
The base model ID is moonshotai.kimi-k3.1. The cross-Region profile IDs are us.moonshotai.kimi-k3.1, in.moonshotai.kimi-k3.1 and global.moonshotai.kimi-k3.1. The capability icons mark image and text input and text output as supported; audio input, embedding output, speech input or output, image output and video input are not supported.
Cross-Region availability depends on the request's source Region
All 32 bedrock-runtime source Regions listed in the model card show In-Region as unsupported and Global as supported. US Geo can be requested from us-east-1, us-east-2, us-west-1, us-west-2, ca-central-1 and ca-west-1; India Geo can be requested from ap-south-1 and ap-south-2.
The other listed source Regions show Global only: eu-central-1, eu-central-2, eu-north-1, eu-south-1, eu-south-2, eu-west-1, eu-west-2, eu-west-3, ap-east-2, ap-northeast-1, ap-northeast-2, ap-northeast-3, ap-southeast-1, ap-southeast-2, ap-southeast-3, ap-southeast-4, ap-southeast-5, ap-southeast-6, ap-southeast-7, il-central-1, me-central-1, me-south-1, af-south-1 and sa-east-1.
US Geo destinations vary by source Region. Most US sources route to us-east-1, us-east-2 and us-west-2; a us-west-1 source also includes itself. A ca-central-1 or ca-west-1 source can route to that Canadian Region and those three US Regions, while India Geo routes between Mumbai and Hyderabad. The card's narrative says the US profile routes only to US Regions, but the destination table also includes a Canadian Region for Canadian sources. It therefore does not establish US-only routing for every source.
Global can route to supported commercial AWS Regions and does not promise data residency; the AWS Region used for a request is not the reader's residence or proof of account eligibility. If residency affects the choice, call GetInferenceProfile from the actual source Region to check current destinations; this article did not make that call.
Estimate token costs from the published Standard rates
The model card lists Standard rates per million tokens. Global is $4.00 for input, $20.00 for output, $0.20 for cache reads and $5.00 for a 30-minute cache write; US and India are each $4.40, $22.00, $0.22 and $5.50, respectively.
Priority is listed at 1.75 times the applicable Standard base rate and Flex at 0.5 times; the card says these multipliers apply to the Global or US CRIS Standard base rate. It does not state an India multiplier basis, so do not apply either multiplier to India by assumption. Standard, Priority and Flex are marked supported and Reserved is not; the tier parameter differs by API, with a serviceTier object for Converse and service_tier for the other listed APIs.
As a hypothetical calculation, 1,000 uncached input tokens plus 100 output tokens at the Global Standard rates come to $0.006, excluding caching, other charges and taxes. This is arithmetic based on the published rates, not a measured workload, an account bill or a final quote.
Account separately for caching, reasoning and quota
Implicit prompt caching is enabled by default. Explicit caching is supported with Responses, Chat Completions and InvokeModel, but not with Converse; a cache checkpoint needs at least 256 tokens and the listed retention time (TTL) is at least 30 minutes. These conditions do not guarantee a cache hit, and this article has not measured hit rates or savings.
Reasoning defaults to max. Supported values are none, low, medium, high, xhigh and max; minimal is unsupported. Responses uses reasoning.effort, while Chat Completions and InvokeModel use reasoning_effort. Converse accepts the field but does not validate its values for this model.
On bedrock-runtime, each output token consumes 10 quota tokens. This is quota accounting, not a billing multiplier, and it does not imply a generic 700-million-token daily allowance. Actual account quotas can vary; check that account's Service Quotas.
Understand Converse limits before choosing an integration
AWS recommends using the OpenAI-compatible Responses or Chat Completions APIs where possible. A known Converse issue can produce an InternalServerException when earlier reasoning content is included in a multi-turn request, which may affect default LangChain and Strands Agents configurations. For Converse multi-turn requests, remove reasoning blocks from prior turns.
Converse rejects attached PDF and HTML documents, and the model does not support video input. For mixed image-and-text input, placing image blocks before text may produce better answers, but the effect is prompt-dependent, so test both orderings for the workload. The image detail setting low or high is honored only by Chat Completions; Responses always processes images at high detail.
Before integrating, confirm that the account identity has the required authorization, check the profile and current routing from the request's source Region, then estimate cost with the intended API, tier and caching strategy and review the account quota. This article draws on the public announcement, model card and saved tables; it did not access an AWS account, call an inference API or verify account-level access.
Sources
Frequently Asked Questions
Does a 1-million-token context window mean the model can generate 1 million tokens in one response?
That does not follow from the context-window figure. The model card describes context size, but the material used here does not give a maximum output-token limit; check the request limits for the API you plan to use.
Can I keep the Global profile ID but budget directly at the US rates?
Base the estimate on the inference profile and service tier you will actually use. The model card lists separate Global, US and India rates. If you plan to use a different profile, also check that the actual request source Region supports it, rather than simply substituting a price in the budget.