In a Silicon Valley lab supported by the UAE, a team of researchers has achieved the extraordinary: crafting an AI system that rivals the models of tech giants, using far fewer resources.
Named K2 Think, this system was developed by the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI). With 32 billion parameters, it competes with OpenAI and DeepSeek’s models, which use over 120 billion parameters, making it the most advanced open-source reasoning AI of its size.
K2 Think is unique in its full availability for free on its website, positioning the UAE alongside the US and China as one of the few countries providing open-source reasoning models globally.
Keep up with the latest news. Follow KT on WhatsApp Channels.
“We began by fine-tuning K2 Think with carefully selected examples of detailed, step-by-step reasoning,” said Hector Liu, Director of MBZUAI’s Silicon Valley lab, in an exclusive interview with Khaleej Times. “This training enabled the model not only to provide answers but to clearly articulate its thought process.”
The breakthrough merges three unique techniques that major tech firms have struggled to integrate effectively. The team instructed K2 Think to deconstruct complex problems systematically, creating what Liu calls “a robust foundation for addressing intricate questions in fields like math, coding, and science.”
Hector Liu
Unlike most AI systems that rely on human feedback, K2 Think is trained on tasks with definitive right and wrong answers. “The model can optimize for accuracy in areas such as logic, simulations, and tabular reasoning, enhancing its precision and reliability,” Liu explained.
The third innovation involves real-time problem-solving. Before answering, K2 Think devises a plan, generates multiple possible responses, and selects the best one. This method results in higher-quality outcomes without the need for extensive computing power.
The team’s focus on efficiency stemmed from practical limitations. While tech giants pour hundreds of millions into cutting-edge AI models, the MBZUAI team worked with limited resources.
“The K2 Think project has always prioritized efficiency,” Liu said. “While unlimited computing power can achieve many great things, we aimed to see how far we could push performance in a resource-constrained setting.”
The team surpassed their expectations despite technical hurdles throughout development. “We set out to match the capabilities of top models, and thanks to our team’s exceptional talent and hard work, we’ve met and sometimes exceeded the goals and expectations for K2 Think,” Liu remarked.
Liu acknowledges the competitive landscape. OpenAI’s latest reasoning models are significantly larger than K2 Think, with even their open-source versions utilizing 120 billion parameters compared to K2 Think’s 32 billion.
Nonetheless, K2 Think’s performance, despite its smaller size, rivals those with vast budgets. The system is entirely free and accessible at k2think.ai
“K2 Think not only positions the UAE on a global platform with the US and China in open-source reasoning LLMs (Large Language Models) but also demonstrates that efficiency doesn’t equate to compromised performance,” Liu stated.
The team is continually exploring efficiency improvements in their model development process, expanding capabilities while ensuring practical relevance beyond mere benchmark performance.
