Startup Moonshot, backed by Alibaba, has introduced a new artificial intelligence model called Kimi K2 Thinking. This model builds on the previous K2 version from July and brings improvements in agentic capabilities, meaning it can better understand user requests without requiring detailed step-by-step instructions. Moonshot, founded in 2023, focuses on open AI models that are available to the general public and developers alike.
The Kimi K2 Thinking model is built as a Mixture-of-Experts (MoE) with a total of one trillion parameters, 32 billion of which are activated during each computation. This enables efficient processing of complex tasks. Developers can use it through platform.moonshot.ai or kimi.com, while its weights and code are available on Hugging Face. The model is released under a modified MIT license that permits commercial use but requires the name "Kimi K2" to be displayed in the user interface if the product exceeds 100 million monthly active users or generates more than USD 20 million per month.
Benchmark performance
Kimi K2 Thinking achieved record results in several key tests. In the Humanity’s Last Exam (HLE) benchmark, it scored 44.9% using tools, the best score for expert-level questions across subjects. In agentic web search and browsing, it achieved 60.2% in BrowseComp and 56.3% in Seal-0, which tests the collection of up-to-date information from the real world.
In coding, the model excels with 71.3% in SWE-Bench Verified, 83.1% in LiveCodeBench v6, and 44.9% in SWE-Multilingual. These results surpass proprietary models such as OpenAI’s GPT-5, Anthropic’s Claude Sonnet 4.5, and xAI’s Grok-4. For example, in BrowseComp, it outperformed GPT-5 with 60.2% versus 54.9% and Claude 4.5 with 24.1%. In GPQA Diamond, it achieved 85.7% compared with GPT-5’s 84.5%.
The model also surpassed the previous leading open model, MiniMax-M2, scoring 60.2% versus 44.0% in BrowseComp, for example, and 71.3% versus 69.4% in SWE-Bench Verified. Kimi K2 Thinking supports native INT4 inference and a context window of up to 256,000 tokens, ensuring speed and accuracy even with long sequences.
Agentic capabilities and tools
One of Kimi K2 Thinking’s main strengths is its ability to automatically select and use 200 to 300 tools in sequence without human intervention. This makes it possible to solve complex problems involving planning, search, analysis, and data synthesis across hundreds of steps. The model produces continuous reasoning traces in the reasoning_content field, ensuring transparency and consistency during lengthy tasks.
One example is a daily news report workflow, in which the model calls tools for the date, web searches, content analysis, and structured output. This agentic structure allows models to operate autonomously, which is essential for coding applications such as compiling, testing, and fixing code, or for browsing the web to gather information.
Efficiency and costs
Training the Kimi K2 Thinking model cost USD 4.6 million. This is significantly less than the billions spent by OpenAI, yet the model achieves comparable or better performance. Usage costs are USD 0.15 per million tokens for a cache hit, USD 0.60 per million tokens for a cache miss, and USD 2.50 per million output tokens. These rates are competitive with MiniMax-M2 at USD 0.30 for input and USD 1.20 for output, and significantly lower than GPT-5 at USD 1.25 for input and USD 10 for output.
The model is optimized for speed, with twice the inference speed thanks to INT4 QAT quantization, making it ideal for long reasoning sequences. This efficiency allows businesses to deploy open models without relying on proprietary APIs, while retaining full control over data and weights.
Availability and applications
Kimi K2 Thinking is available on kimi.com in chat mode, with full agentic mode coming soon. The API is accessible through platform.moonshot.ai. The model can also be tested on Hugging Face Spaces. Its open nature allows researchers and companies to customize it for specific tasks, such as agentic coding or search.
This model is another addition to the growing competition among open Chinese AI systems, such as DeepSeek, which trained its V3 model for USD 5.6 million. Companies such as Airbnb are already publicly praising Chinese models for their cost and performance compared with OpenAI. Moonshot is thus contributing to a trend in which open models are reaching the level of closed systems.
Sources: moonshotai.github.io, cnbc.com and venturebeat.com
