Kimi K3.1 Pro Highspeed lands on Bedrock, claiming 2-3x faster output

Amazon has added Kimi K3.1 Pro Highspeed from Moonshot AI to Amazon Bedrock. According to AWS’s announcement, Moonshot reports 2 to 3 times the output speed of Kimi K3.1 Pro, with the same endpoint and APIs as Kimi K3.

A faster Kimi that swaps in by model ID

Kimi K3.1 Pro Highspeed is part of Moonshot AI’s K3.1 series. AWS describes it as a post-trained version of Kimi K3 that keeps the same open-weight architecture, in the 3-trillion-parameter class, and the 1-million-token context window. Moonshot says the added post-training compute targets long-running agent work and fewer factual errors. It also claims gains in financial modeling, office documents, and coding for game engines, 3D scenes, and CAD scripts.

The model supports tool calling, structured output, image input, streaming, and configurable reasoning. It always reasons before answering, and the reasoning_effort setting accepts low, high, or max. If a developer sets nothing, the model uses max, so AWS suggests a lower value to cut latency and output tokens on simple tasks.

Teams already building on Kimi K3 do not need new plumbing. AWS says they can switch by changing the model ID in existing code. The model is reachable through the OpenAI-compatible Chat Completions and Responses APIs, which AWS recommends, as well as the Bedrock Converse and Invoke APIs.

Three inference profiles, a half-price tier, and caching limits

Access runs through cross-Region inference profiles:

• global.moonshotai.kimi-k3.1 can send requests to whichever commercial AWS Region is supported, and AWS prices it per token below the regional profiles. AWS recommends it for most workloads.
• us.moonshotai.kimi-k3.1 keeps processing in US Regions.
• in.moonshotai.kimi-k3.1 keeps processing in India Regions.

The post does not print per-token prices; it points readers to the Bedrock pricing page. It does say the Flex tier costs half the Standard price in exchange for potentially slower responses, and suggests it for evaluations, summarization, and background agent tasks. Tier is chosen per request with service_tier set to flex, so one application can mix Priority, Standard, and Flex traffic. All three tiers share one on-demand quota.

Explicit prompt caching is available on Chat Completions, Responses, and Invoke, while Converse does not support it. Each prompt_cache_breakpoint needs a stable prefix of at least 256 tokens before it, and cached content is kept for at least 30 minutes. Cache writes cost more than standard input, while reads are discounted and do not count toward the input tokens-per-minute quota.

AWS also walks through connecting the model to OpenCode, an open source terminal coding agent. Its argument is that agent loops make many sequential calls, so a faster output rate shortens each step of the edit-test-fix cycle.

Data residency is the real selling point for regulated teams

The analysis here turns on privacy and data handling, where the announcement makes its strongest claims. AWS says prompts are processed inside the AWS data boundary, are not shared with Moonshot AI, and are not used for training. It also says zero data retention is always on for inference, and that zero operator access means AWS staff cannot see prompts or completions during inference.

The residency options are concrete. In a separate November 3 post on India-based inference for Kimi K3, AWS said the India profile keeps requests within the Mumbai (ap-south-1) and Hyderabad (ap-south-2) Regions, and that cross-Region traffic stays on the AWS network rather than the public internet. That post also cited Ultrahuman, which reported an 80% LLM cost reduction from prompt caching; that is the customer’s own figure, not a published price.

Geography is wider than the announcement suggests. A separate AWS What’s New notice dated November 6 says the model is available in all Regions where Bedrock is offered through cross-Region inference. That fits the global profile, but the engineering post names only three profiles, so teams with other residency rules should check the model card and Regional availability page. We have also explained how Bedrock attaches controls to many models under one account.

What the announcement does not settle

The performance claims are Moonshot’s. In our weekly model roundup, the K3.1 Pro model card listed 94.2 on GPQA-Diamond, 92.1 on BrowseComp, and 65.7 on Terminal-Bench 4.0, all self-reported and not independently verified. The 2 to 3 times speed figure is likewise Moonshot’s, and AWS gives no tokens-per-second numbers or latency tests.

The AWS notices also blur the line between the two variants. The What’s New page says it does not separately specify how K3.1 Pro and the Highspeed version differ, and the engineering post covers only Highspeed. The roundup said the weights ship under the Kimi K3.1 Pro License and advised reading its terms before commercial use. Pricing for Highspeed, the Priority tier, and the global versus geographic discount remain unstated in the announcement.

For developers, the practical step is to test it on a small share of traffic. AWS itself suggests starting Flex with a small share and comparing latency and cost before moving more.