Kimi K2: New AI model from China outperforms GPT-4.1 in key benchmarks
Chinese company Moonshot AI has introduced a new language model, Kimi K2, which is specifically aimed at developers and professional users. The approach is reminiscent of OpenAI's GPT models, but is intended to deliver better results in some areas. Kimi K2 is now available via an API.
As Reuters reports, Kimi K2 was officially introduced in July 2025. The model is based on a mixture-of-experts architecture with around one trillion parameters, although only 32 billion of these are used simultaneously. This saves computing power and improves efficiency.
Kimi K2 is available in two versions: Instruct is aimed at users who want to use the model directly. Base is intended for research and individual fine-tuning. Both versions can be integrated via an OpenAI-compatible application programming interface (API). However, according to Moonshot AI, commercial use is limited in cases of high user numbers or large sales volumes.
A key difference from many other language models lies in its training strategy: Kimi K2 was specifically designed to structure tasks independently, use tools and create simple program code. The model is designed to provide reliable answers even without specific chain-of-thought instructions.
According to Shinkai, Kimi K2 performed well against GPT-4.1 in various tests, including practical programming and knowledge benchmarks. The specific results vary depending on the application, but generally favor the new model, particularly in mathematics, science and multilingualism. Reuters points out, however, that there are also weaknesses. For example, ambiguous questions or unclear tasks can result in longer or incomplete output.
Moonshot AI is one of a growing number of Chinese providers developing their own AI models and making them publicly available. Kimi K2 is a powerful and flexible model that could gain a foothold in international applications in the long term.
