Kimi-k2 Thinking - A thinking model launched by the Dark Side of the Moon

What is Kimi-k2 Thinking?

Kimi-k2 Thinking is an AI model developed by Dark Side of the Moon, possessing general agentic capabilities and deep reasoning abilities. The model boasts powerful multi-turn thinking and tool invocation capabilities, enabling it to autonomously complete complex tasks without human intervention, making it suitable for step-by-step reasoning and planning in complex tasks. In multiple benchmark tests, including "Humanity's Last Exam," "Autonomous Web Browsing" (BrowseComp), and "Complex Information Gathering Reasoning" (SEAL-0), Kimi-k2 Thinking has achieved state-of-the-art (SOTA) performance, while also receiving comprehensive upgrades in agentic search, agentic programming, writing, and integrated reasoning capabilities. Kimi-k2 Thinking includes a high-speed version, Kimi-k2 Thinking-turbo, with a reasoning speed of up to 100 tokens/s, suitable for scenarios with high efficiency requirements.

The Kimi K2 Thinking model is now officially available on kimi.com and in the regular conversation mode of the latest version of the Kimi app. Kimi's Agent model will soon be upgraded to the Kimi K2 Thinking model, providing users with more powerful multi-turn thinking and tool invocation capabilities. The Kimi K2 Thinking API is now available on the Kimi Open Platform, which developers can access.

Main functions of Kimi-k2 Thinking

• Deep reasoningIt can perform complex logical reasoning and multi-step thinking to solve problems step by step, making it suitable for tasks that require in-depth analysis.
• Autonomous tool callIt can solve complex tasks autonomously by calling upon tools (such as search, programming, and web browsing) without human intervention.
• Long-term planning and multi-round interactionIt supports up to 300 rounds of tool calls and continuous, stable multi-round thinking, making it suitable for solving complex problems.
• Long context processingIt supports context lengths of up to 256k and can handle complex long text tasks, such as long text analysis and multi-step task planning.
• Visualizing the reasoning process:pass reasoning_content The fields display the reasoning process, helping users understand the model's logic and enhancing interpretability.
• Efficient ReasoningIt offers a high-speed version (Kimi-k2 Thinking-turbo) with an inference speed of up to 100 tokens/s, suitable for scenarios with high efficiency requirements.
• Cost optimizationIt strikes a balance between inference efficiency and cost, making it suitable for handling complex tasks that require high cost-effectiveness.

Performance of Kimi-k2 Thinking

• reasoning abilityIn "Humanity's Last Exam," which covers more than 100 professional fields, Kimi K2 Thinking achieved a 44.9% State-of-the-Art (SOTA) score, demonstrating strong reasoning and problem-solving abilities.
• Independent search and browsing capabilitiesIn the BrowseComp benchmark test released by OpenAI, Kimi K2 Thinking became the new state-of-the-art model with a score of 60.2%, far exceeding the human average score of 29.2%, demonstrating extremely strong information retrieval and research capabilities.
• Complex information gathering and reasoningIn the SEAL-0 benchmark test, Kimi K2 Thinking demonstrated outstanding ability to gather and reason about complex information, and was able to efficiently process and analyze large amounts of information.
• Agentic programming skillsKimi K2 Thinking has further improved its performance in benchmark tests such as SWE-Multilingual, SWE-bench validation set, and Terminal usage, and has performed well in handling front-end tasks such as HTML and React.

API Usage Guidelines for Kimi-k2 Thinking

• Enter the full contextWhen calling the model, all thought processes must be included.reasoning_content (Fields), which facilitate model analysis based on complete reasoning logic.
• Set large enough max_tokensRecommended settings max_tokens≥16000This ensures that the model can fully output the reasoning process and results.
• The temperature parameter is set to 1.0.:Will temperature Setting it to 1.0 will yield the best performance and inference stability.
• Enable streaming outputUse streaming output (stream=TrueThis improves user experience and avoids network timeouts caused by excessive output content.

The cost of using Kimi-k2 Thinking

• Standard API:
  • enterA fee of 4 yuan is charged for every million tokens entered.
  • OutputThe fee is 16 yuan per million tokens output.
  • Input that hits the cacheThe fee is 1 yuan.
• Turbo API(Speed up to 100 Tokens/s):
  • enterThe fee is 8 yuan per million tokens entered.
  • OutputThe fee is 58 yuan per million tokens output.
  • Input that hits the cacheThe fee is 1 yuan.

Kimi-k2 Thinking project address

• HuggingFace model libraryhttps://huggingface.co/moonshotai/Kimi-K2-Thinking
• Technical Papers: https://mp.weixinbridge.com/mp/wapredirect?url=https%3A%2F%2Fmoonshotai.github.io%2FKimi-K2%2Fthinking.html&action=appmsg_redirect&uin=MTE0MzM2NDMwMA==&biz=Mzk0NDU1MDkyNg==&mid=2247487508&idx=1&type=1&scene=0

Application Scenarios of Kimi-k2 Thinking

• Complex Problem SolvingIt is used for complex problems that require multi-step reasoning and logical analysis, such as scientific experimental design and engineering optimization.
• Automated task planningIn tasks that require dynamic adjustments and multiple rounds of decision-making, such as automated process design and resource allocation.
• Data Analysis and ReportingIt can handle analytical tasks involving large amounts of data and complex logic, and generate in-depth reports, such as market trend analysis and financial forecasts.
• Intelligent search and information integrationBy invoking multiple tools and integrating information from different sources, we provide users with comprehensive answers.
• Education and learning supportIt helps students solve complex academic problems step by step, providing problem-solving strategies and logical reasoning processes.