Overview
- Moonshot says Kimi K2 Thinking outperforms GPT-5 and Claude Sonnet 4.5 on Humanity’s Last Exam, BrowseComp, and Seal‑0, though these vendor-published results await independent verification.
- Kimi K2 Thinking’s weights are available for immediate download on Hugging Face, enabling community testing of its long‑horizon, multi‑step tool use and web‑browsing agent capabilities.
- Moonshot details a Mixture‑of‑Experts architecture with roughly one trillion parameters and claims coherent reasoning across hundreds of steps, including 200–300 sequential tool calls.
- The company and media reports cite a $4.6 million training bill, a figure that, if validated, challenges assumptions about the compute costs behind frontier‑level systems.
- Analysts say a high‑performing free model could pressure proprietary offerings and invite regulatory or security scrutiny, while reports highlighted Nvidia CEO Jensen Huang’s China‑lead remark later clarified to say China is “nanoseconds behind” the U.S.