Photo by Alex Andrews on Pexels

Chinese AI startup Moonshot AI is making a bold move in the artificial intelligence arena with its Kimi K2 Thinking model, an open-source offering that reportedly rivals, and in some cases surpasses, the performance of leading models like OpenAI’s GPT-5 and Anthropic’s Claude Sonnet 4.5 on specific benchmarks. This development is fueling discussions about a potential power shift in AI dominance, as Chinese companies increasingly challenge the established players in the United States.

Backed by tech giants Alibaba and Tencent, Moonshot AI’s Kimi K2 Thinking model has garnered attention from industry experts, drawing comparisons to the impact of DeepSeek. Its strengths lie in areas such as reasoning, agentic search capabilities, and coding proficiency. A key feature is its extensive 256K context window, allowing it to process significantly larger amounts of information.

Benchmark tests have shown Kimi K2 Thinking achieving a 44.9% score on Humanity’s Last Exam, outperforming GPT-5’s 41.7%. The model also achieved a score of 60.2% on the BrowseComp benchmark. Notably, the model’s training is estimated to have cost approximately US$4.6 million, with significantly lower API costs than comparable models from OpenAI and Anthropic.

Built using a Mixture-of-Experts architecture with one trillion parameters, Kimi K2 Thinking is generating discussion within the AI community. While figures like Thomas Wolf from Hugging Face have recognized its potential, others, such as Nathan Lambert, suggest a performance gap still exists when compared to closed-source models. The model provides a glimpse into China’s growing competitiveness in AI through innovative, open-source models.

Moonshot AI’s emergence underscores the intensifying competition within the AI sector, with Chinese firms such as DeepSeek, Qwen, and Baichuan employing innovation and open-source strategies to challenge American AI supremacy.