The September 9 2025 press release and paper announce and detail K2 Think, an advanced open-source AI reasoning system developed by the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and G42 in the UAE. K2 Think stands out for its parameter efficiency, achieving performance comparable to much larger models, particularly in mathematical reasoning, with only 32 billion parameters. This breakthrough is attributed to a six-pillar approach, including supervised fine-tuning, reinforcement learning with verifiable rewards, agentic planning, test-time scaling, and optimization for Cerebras Wafer-Scale Engine hardware for rapid inference. The system also undergoes red-teaming for safety evaluation, demonstrating a solid baseline in content refusal and conversational robustness, while identifying areas for improvement in cybersecurity and jailbreak resistance.

Sources:

https://k2think-about.pages.dev/assets/tech-report/K2-Think_Tech-Report.pdf

https://k2think-about.pages.dev/assets/announcement/K2_Think_-_Sep_9_English_Release.pdf

https://www.k2think.ai/k2think