The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), in partnership with G42, has launched K2 Think, an open-source AI reasoning system redefining efficiency and intelligence in advanced models.
With only 32 billion parameters, K2 Think outperforms reasoning models nearly 20 times larger, setting a new benchmark for parameter efficiency and logical depth. Built on six pillars of innovation, the system leverages chain-of-thought supervised fine-tuning, reinforcement learning with verifiable rewards, agentic planning, and test-time scaling. Soon, it will run on Cerebras’ wafer-scale inference platform, reaching speeds of 2000 tokens per second—making it one of the fastest reasoning models available.
Benchmarking results already place K2 Think among the best-performing reasoning systems globally, tying for top math scores on AIME ’24/’25, HMMT ’25, and OMNI-Math-HARD. Unlike most models labeled “open,” K2 Think is entirely transparent, releasing not just weights but also training data, deployment software, and optimization code—ensuring reproducibility and global collaboration.
Leaders including Khaldoon Khalifa Al Mubarak, Peng Xiao, and Professor Eric Xing hailed K2 Think as a milestone for AI innovation in the UAE, shifting the paradigm from “bigger is better” to “smarter is better.” The system builds on a legacy of UAE-developed open-source models such as Jais, NANDA, and SHERKALA, as well as the K2-65B model released in 2024.
K2 Think is now accessible worldwide at www.k2think.ai and via Hugging Face, marking a transformative step in advancing open, efficient, and globally impactful AI.
