Moonshot's Kimi K3.1 Pro (Highspeed) arrives on Amazon Bedrock with 1M-token context

Kimi K3.1 Pro from Beijing-based Moonshot AI is now a Bedrock option. The Highspeed variant is tuned for output tokens per second, which matters for agent loops that make many sequential model calls. It offers a 1-million-token context window and belongs to a roughly 3-trillion-parameter model class. It runs under Bedrock's standard security and governance controls.

AWS highlights explicit prompt caching. Developers can cache large, repeated prefixes such as codebases or document corpora, then pay less and wait less on subsequent calls. For long-context agentic coding, where the same repository is re-read many times, caching often dominates total cost.

The addition reinforces AWS's strategy of building a broad model marketplace. Jassy called Bedrock AWS's fastest-growing multibillion-dollar business this week. Hosting Chinese open-weight models alongside Anthropic's Claude, Meta's Llama and Mistral gives enterprises price-performance options while keeping data inside AWS. Chinese open models, including DeepSeek, Qwen and Kimi, currently lead many open-weight leaderboards.

Caveats: some enterprises restrict the use of Chinese-origin models for compliance reasons, even when they are self-hosted on US infrastructure. Pricing per token was not detailed in the sources. Watch whether Bedrock adds Qwen 4 quickly if Alibaba ships it.