In a Silicon Valley lab backed by the UAE, a handful of researchers have achieved what seemed unthinkable: building an artificial intelligence (AI) system that rivals tech giants’ billion-dollar models with just a fraction of the resources.
The system, called K2 Think, was developed by the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI). With 32 billion parameters, it matches the reasoning power of OpenAI and DeepSeek’s models that operate at more than 120 billion parameters, making it the leading open-source reasoning AI of its size.
Unlike other systems, K2 Think is fully free and publicly available on its website, positioning the UAE alongside the US and China as one of the few nations offering open-source reasoning models to the world.
“We started by fine-tuning K2 Think on carefully curated examples of long, step-by-step reasoning,” Hector Liu, Director of MBZUAI’s Silicon Valley lab, told Khaleej Times, in an exclusive interview. “This training helped the model learn not just to give answers, but to show its thought process in a clear and structured way.”
The breakthrough combines three distinct techniques that major technology companies have struggled to integrate effectively. The team taught K2 Think to break down complex problems step-by-step, building what Liu describes as “a strong foundation for tackling complex questions across areas like math, coding, and science.”
Beyond most AI systems that learn from human feedback, K2 Think is trained on tasks with clear right and wrong answers. “The model could directly optimize for correctness in areas such as logic, simulations, and tabular reasoning, giving it sharper accuracy and reliability,” Liu explained.
The third innovation involves real-time problem-solving. Before providing answers, K2 Think creates a plan, generates multiple potential responses, and then selects the best option. This approach delivers higher-quality results without requiring massive increases in computing power.
According to the team, the efficiency focus emerged from practical constraints. While technology giants invest hundreds of millions in developing frontier AI models, the MBZUAI team operated with limited resources.
“The entire project for K2 Think has centered around efficiency,” Liu said. “Many great things can be achieved with unlimited compute, but we wanted to explore how far we can push performance in a relatively constrained environment.”
The team exceeded their initial expectations despite technical challenges throughout development. “Our team set an ambitious goal to match the qualities of top models, and thanks to top talent on our team and a lot of hard work, we’ve achieved, and in some cases exceeded the goals and expectations of the K2 Think project,” Liu noted.
Liu acknowledges the competitive reality. OpenAI’s latest reasoning models remain “orders of magnitude larger than K2 Think,” with even their open-source releases using 120 billion parameters compared to K2 Think’s 32 billion.
However, K2 Think’s performance, despite its size, was able to compete with massive budgets. The system is completely free and available at k2think.ai.
“Not only does K2 Think put the UAE on the global stage with the USA and China with open-source reasoning LLMs (Large Language Models) but proves performance need not be sacrificed when striving for efficiency,” Liu said.
The team continues investigating efficiency improvements across their entire model development process, expanding capabilities while ensuring practical utility beyond benchmark performance.
