Kimi K2’s thinking model was trained with just $4.6 million??? While OpenAI raised trillions, K2 Thinking was built for 0.1% the cost of its US counterparts, yet still reached SOTA performance! It even beats GPT-5 and Claude 4.5 Sonnet in various benchmarks.

It generated a complete novel from a single prompt, an entire book with a collection of 15 short sci-fi stories, reaching up to 300 tool requests per session.

Kimi K2 Thinking has built this 'Live coding music with Strudel.cc' from a single prompt:

Kimi K2 Thinking sets new records across benchmarks that assess reasoning, coding, and agent capabilities. - 44.9% on HLE with tools - 60.2% on BrowseComp - 71.3% on SWE-Bench Verified

It solved a PhD-level math problem using ​23 interleaved reasoning and tool calls, showcasing its deep, structured reasoning and long-term problem-solving abilities.

K2 Thinking exhibits substantial gains in coding and software development tasks. Check out this Math Explainer example: Visualization of gradient descent

K2 Thinking excels in reasoning and problem-solving. On Humanity’s Last Exam (HLE), a challenging test with thousands of expert-level questions in over 100 subjects, K2 Thinking scored a state-of-the-art 44.9%. Using search, Python, and web-browsing tools, it set new records in multi-domain expert reasoning.

K2 Thinking is now live on kimi.com in chat mode, with full agentic mode coming soon. It is also accessible via API. 🔌 API is live: platform.moonshot.ai 🔗 Tech blog: moonshotai.github.io/Kimi-K2… 🔗 Weights & code: huggingface.co/moonshotai