The UAE entered the global AI arena with the launch of K2 Think, a reasoning model developed jointly by the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and tech group G42. At just 32 billion parameters, the model delivers reasoning performance matching much larger systems from OpenAI and DeepSeek while operating at a fraction of their size.
K2 Think is designed not just for raw power but for structured, efficient thinking. It integrates techniques like long chain-of-thought fine-tuning, agentic planning, reinforcement learning with verifiable rewards, and test-time scaling. Deployed on wafer-scale Cerebras hardware, it achieves a throughput of 2,000 tokens per second, a speed unmatched by open-source counterparts.
On math benchmarks such as AIME ’24/’25, HMMT ’25, and OMNI-Math-HARD, K2 Think ranks among the top open-source reasoning systems. It also delivers strong results in code and science tasks, proving its versatility across technical domains.
The model is fully open-source, users can access not only its weights but also training data, deployment code, and inference optimisations.
MBZUAI itself emphasises this transparency as foundational to advancing reproducible, accountable AI innovation.
MBZUAI’s president, Eric Xing, described K2 Think as a “disruption,” showcasing how a smaller model, designed effectively, can rival larger alternatives. G42’s CEO Peng Xiao echoed that vision, calling it proof that “smarter is better, not bigger.”
