The Chinese company Moonshot introduced a new AI model Kimi K2 Thinking, which, according to the developers, surpasses GPT-5 from OpenAI and Claude Sonnet 4.5 from Anthropic in key tests of logical reasoning and planning. These tests include Humanity’s Last Exam, BrowseComp for testing web browser work, and Seal-0, which assesses analytical abilities. The model also demonstrates a programming level comparable to leading closed systems.
Kimi K2 Thinking has about 1 trillion parameters and uses the “Mixture-of-Experts” approach, combining long-term planning, adaptive thinking, and working with online tools. It can break down complex tasks into simpler steps, verify evidence, generate and refine hypotheses, and create coherent responses.
The main feature of Kimi K2 Thinking is its openness: the model is available for free use on the Hugging Face platform, and Moonshot has published its code and weights. The cost of training the model was 4.6 million dollars, which is significantly less than the expenses for creating similar products in the USA.
The novelty may attract businesses by offering a free alternative to paid corporate solutions from Western companies. Some enterprises have already started preferring Chinese open AI models due to their efficiency and lower cost, although security issues remain relevant.
