
Kimi K3.1 Pro Highspeed arrives on Amazon Bedrock
Kimi K3.1 Pro Highspeed from Moonshot AI is available on Amazon Bedrock. It is a post-trained version of Kimi K3 and retains the same 3T-class open-weight architecture and one-million-token context window. Moonshot AI reports that it generates output tokens two to three times as fast as Kimi K3.1 Pro. That speed comparison is a vendor-reported claim and is not independently verified by the supplied evidence.
A compatible path from Kimi K3
For teams already using Kimi K3 on Amazon Bedrock, the main integration detail is continuity. The Highspeed model uses the same Bedrock endpoint, APIs, and features as Kimi K3. Existing Kimi K3 code can therefore try the new variant by changing the model ID, without replacing the entire application-facing integration.
The model is accessible through the OpenAI-compatible Chat Completions and Responses APIs, as well as the Amazon Bedrock Converse and Invoke APIs. Teams can select an interface that fits their current client and workflow. Supported capabilities include tool calling, structured output, configurable reasoning, image input, and streaming. Together, these features cover interactive responses, machine-readable results, tool-oriented flows, and requests that include images.
Reasoning behavior and inference profiles
The model always reasons before answering. Applications can set reasoning effort to low, high, or max. If no value is supplied, max is used by default. The setting therefore adjusts the level of reasoning effort; it does not switch reasoning off.
Three cross-Region inference profiles are available: global.moonshotai.kimi-k3.1, us.moonshotai.kimi-k3.1, and in.moonshotai.kimi-k3.1. The US profile keeps processing within its respective US Regions, while the India profile keeps processing within its respective India Regions. The global profile is a separate available option, but the supplied facts do not assign it the same regional-processing restriction.
Data handling and service tiers
AWS states that inference data remains inside the AWS data boundary, is not shared with Moonshot AI, and is not used to train the model. AWS also states that zero data retention and zero operator access apply to inference requests. These properties describe the handling of inference data, while the choice between global, US, and India profiles remains an application-level deployment decision.
Applications can select Priority, Standard, or Flex for each request. Flex serves Kimi K3.1 Pro Highspeed at half the Standard price in exchange for potentially longer response times. All three tiers share one on-demand quota. This creates a request-level choice: an application can use a lower-priced tier where waiting is acceptable and choose another tier when its processing preference differs, without separating the tiers into different on-demand quotas.
Explicit prompt caching
Explicit prompt caching is supported through Chat Completions, Responses, and Invoke, but it is not supported through Converse. A cache breakpoint requires at least 256 preceding tokens of stable content, and cached content is retained for at least 30 minutes. The API choice therefore affects whether caching is available. The feature is relevant when requests repeatedly reuse a sufficiently long, stable prefix rather than when every prompt is entirely different.
Using the model with OpenCode
OpenCode can connect to Kimi K3.1 Pro Highspeed through Amazon Bedrock’s Chat Completions API. The setup uses OpenCode’s OpenAI-compatible provider together with an Amazon Bedrock API key. Short-term keys last for up to 12 hours and operate only in the Region where they were generated, so the key, client configuration, and selected Region need to remain aligned.
The release brings the model’s long context, API compatibility, image input, streaming, tool calling, and structured output into one Bedrock offering. A practical evaluation can begin with a model-ID change, then account for the desired inference profile, reasoning effort, service tier, and API-specific caching support. Any assessment of the reported speed advantage should continue to identify it as Moonshot AI’s claim rather than an independently established measurement.