Overview
- Beijing-based Moonshot says Kimi K2 Thinking outscored OpenAI’s GPT-5 and Anthropic’s Claude Sonnet 4.5 on benchmarks such as Humanity’s Last Exam, BrowseComp, and Seal-0.
- The Mixture-of-Experts model uses roughly 1 trillion parameters with open weights available on Hugging Face, extending July’s Kimi K2 release.
- Moonshot reports training spend at about $4.6 million, a fraction of typical frontier budgets, with external validation of results still pending.
- Demonstrations highlight multi-step agentic behavior, including hundreds of sequential tool calls for tasks like complex math, coding projects, and data visualizations.
- Coverage notes potential pressure on paid proprietary offerings, as debate over U.S.–China AI leadership intensifies following Jensen Huang’s remark later clarified to say China is “nanoseconds behind America.”