Moonshot AI launches Kimi K3.1 Pro Highspeed on Amazon Bedrock

Moonshot AI has released Kimi K3.1 Pro Highspeed on Amazon Bedrock, expanding the availability of its latest open-weight large language model. This new variant, part of the K3.1 series, is designed to deliver significantly higher output token generation rates while maintaining the architecture’s existing capabilities for long-context reasoning and complex professional tasks.

The release targets developers building agentic systems and coding assistants who need faster iteration cycles without sacrificing the model’s ability to handle million-token context windows or specialized domains like computer-aided design scripts.

What happened

Kimi K3.1 Pro Highspeed is a post-trained version of the base Kimi K3 model, utilizing the same 3T-class open-weight architecture and supporting a context window of up to one million tokens. Moonshot AI applied substantially more compute during the post-training phase to reduce factual errors and improve performance in long-horizon agentic work. The company reports that this specific variant generates output tokens at a rate two to three times faster than the standard Kimi K3.1 Pro.

The model is accessible through Amazon Bedrock using cross-Region inference profiles. Developers can route requests via global.moonshotai.kimi-k3.1 for lower costs across supported commercial regions, or use us.moonshotai.kimi-k3.1 and in.moonshotai.kimi-k3.1 to keep data processing within the United States or India respectively. These geographic profiles address specific data residency requirements while providing access to the same underlying model capabilities.

Integration requires minimal changes for existing users. The model shares endpoints, APIs, and features with previous Kimi versions on Bedrock. Developers already using Kimi K3 can switch to the Highspeed variant simply by updating the model ID in their code. The model supports tool calling, structured output, image input, streaming, and configurable reasoning efforts, allowing it to fit into diverse application architectures without requiring significant refactoring.

How it works

The model operates with a built-in reasoning step before generating final answers, a feature controlled by the reasoning_effort parameter. Developers can set this to low, high, or max, with max being the default. Lowering the effort level reduces latency and output token count for simpler queries, while higher levels engage deeper logical processing for complex problems. This mechanism allows applications to balance accuracy and speed dynamically based on the task at hand.

Cost optimization is achieved through two primary mechanisms: the Flex tier and explicit prompt caching. The Flex tier offers inference at half the Standard price, though it may result in longer response times. This makes it suitable for background tasks, model evaluations, or summarization where immediate latency is less critical. Additionally, explicit prompt caching allows developers to mark stable prefixes in their prompts, such as system instructions or repository guidelines. The system caches these segments for at least 30 minutes, charging a discounted rate for cache reads and excluding them from input token quotas, which significantly lowers costs for repetitive agentic workflows.

Key details

• Kimi K3.1 Pro Highspeed reports 2–3 times the output speed of Kimi K3.1 Pro.
• The model supports a 1-million-token context window and uses a 3T-class open-weight architecture.
• Data processed on Amazon Bedrock remains within the AWS data boundary and is not shared with Moonshot AI or used for training.
• The Flex service tier provides inference at 50% of the Standard price, ideal for non-latency-critical workloads.
• Explicit prompt caching requires at least 256 tokens of stable content and retains cached data for a minimum of 30 minutes.
• Integration is available via OpenAI-compatible Chat Completions, Responses APIs, and Amazon Bedrock Converse APIs.

Why it matters

For engineers building autonomous agents, output speed is a critical bottleneck. Agentic workflows often involve sequential calls where the model reads files, plans changes, edits code, and runs tests. Each step must complete before the next begins, meaning faster token generation directly shortens the entire edit-test-fix loop. By doubling or tripling output speed, Kimi K3.1 Pro Highspeed can significantly reduce the wait time for developers interacting with coding assistants or automated debugging tools.

The introduction of the Flex tier and explicit caching also addresses the economic challenges of deploying large context models. Agentic systems frequently resend large blocks of static context, such as documentation or codebase structures, with every request. Without caching, this leads to redundant processing costs. By enabling developers to explicitly mark these stable sections, AWS allows for more predictable and lower inference bills, making it feasible to run complex, long-running agents in production environments without prohibitive expenses.

What you can do

• Update your existing Bedrock integration by changing the model ID to global.moonshotai.kimi-k3.1 or a regional equivalent.
• Experiment with the reasoning_effort parameter, setting it to low for simple tasks to reduce latency and token usage.
• Implement the Flex tier for background jobs by setting service_tier to flex in your API requests.
• Use explicit prompt caching for system prompts and reference documents by adding prompt_cache_breakpoint markers after at least 256 tokens of stable content.
• Monitor your usage patterns using CloudWatch dimensions ServiceTier and ResolvedServiceTier to optimize the mix of Standard and Flex traffic.
• Connect open-source coding agents like OpenCode to Bedrock using short-term bearer tokens for secure, local development assistance.