The global AI race is taking a new turn: Kimi K2 Thinking, a new model developed by Beijing-based Moonshot AI Lab is now available. The company claims that its latest embodiment outperforms OpenAI’s GPT-5 and Anthropic’s Claude Sonnet 4.5 on some of the most important benchmarks going. What’s more, it’s totally free – because Kimi K2 Thinking is open-source.
Moonshot describes Kimi K2 Thinking as a Mixture-of-Experts (MoE) model focusing on reasoning capabilities. The system combines long-range planning, adaptive reasoning and integrated online tools like browsing software into a powerful package. It can handle complex academic and analytical tasks.On tests like Humanity’s Last Exam, BrowseComp, and Seal-0, the model outperformed GPT-5 and Claude Sonnet 4.5 in reasoning, planning, and web browsing tasks. Its coding performance was similar to that of main competitors, but did not markedly outstrip them.
The model’s architecture enables it to split ambiguous problems into clear steps, each of them actionable. Kimi K2 was trained with some 1 trillion parameters and is publicly available from Hugging Face, in effect the GitHub for Large Language Models (LLMs).
Kimi K2 Thinking is open source. Developers can freely access its code and weights, in a direct challenge to the prevailing trend of closed subscription based platforms in artificial intelligence. According to Moonshot AI Lab, the model cost only $4.6 million to develop. This is only a small fraction of what major US AI labs spend. If these results stand up under independent scrutiny, it could well be another DeepSeek situation for both Silicon Valley and Wall Street later this year.