Abu Dhabi Unveils AI Reasoning Model Challenging OpenAI and DeepSeek

The global artificial intelligence landscape is witnessing the emergence of a new contender, signaling a potential shift in the balance of power.

The Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), a research institution established in the United Arab Emirates, has unveiled K2 Think, a novel reasoning model positioned to challenge established players like OpenAI and DeepSeek. The announcement highlights the UAE’s escalating ambitions in the AI domain.

This development follows DeepSeek’s earlier introduction of its R1 model, a powerful offering from the Chinese AI lab that purportedly surpasses OpenAI’s performance metrics while demanding significantly lower training costs. These advancements highlight the increasing competition and innovation in the AI sector.

K2 Think stands out due to its compact size, operating with just 32 billion parameters. This is in stark contrast to models such as DeepSeek’s R1, which utilizes 671 billion parameters, representing the variables influencing the model’s language understanding and generation capabilities. Although OpenAI doesn’t disclose the size of their models, it is understood that are significantly larger. K2 Think leverages Alibaba’s open-source Qwen 2.5 model as its foundation and relies on Cerebras’ AI accelerator hardware for operations and testing. This strategy not only offers a cost-effective advantage but also promotes collaboration and open-source development within the AI community.

Developed in collaboration with G42, a prominent UAE-based AI firm backed by Microsoft, K2 Think aims to deliver comparable performance to the leading reasoning models despite its reduced size. This partnership reflects the UAE’s broader strategy of forging alliances with established tech giants to bolster its AI capabilities and global influence.

MBZUAI cites benchmark results from AIME24, AIME25, HMMT25, OMNI-Math-HARD (all math-related assessments), LiveCodeBenchv5 (coding), and GPQA-Diamond (science) to support its claims of competitive performance. These benchmarks serve as crucial metrics for evaluating the model’s efficacy across various cognitive tasks, underscoring its potential across diverse fields.

Hector Liu, Director of MBZUAI’s Institute of Foundation Models, explained that K2 Think achieves its performance through a combination of sophisticated techniques.

These methods include long chain-of-thought (CoT) supervised fine-tuning, enabling the model to emulate step-by-step reasoning, and test-time scaling, a technique that dynamically allocates extra computing resources during “inferencing” to optimize performance on unseen data. These optimizations highlight a strategic focus on enhancing the model’s reasoning prowess with a smaller footprint.

“Our approach treats the model as part of a larger system,” Liu stated. “Unlike typical open-source releases, we actively deploy K2 Think and continuously seek ways to enhance its performance over time.” This iterative approach, prioritizing real-world implementation and ongoing refinement, differentiates MBZUAI’s strategy from traditional model development paradigms.

“It’s difficult to isolate a single most important element. The final performance is the culmination of a system-level approach where the combined methods contribute to the overall result,” he added.

The US and China currently dominate the AI landscape, but the UAE is actively working to carve out a prominent position.

While companies like OpenAI have spurred the US to seize early momentum with foundation models, China’s DeepSeek R1 has demonstrated a formidable competitive edge. The UAE is strategically investing in AI to diversify its economy beyond oil dependency and expand its influence on the global stage.

G42 exemplifies the UAE’s progress, but the region faces increasing competition from Saudi Arabia, which aims to build full-stack AI capabilities through its Humain initiative. This rivalry contributes to a dynamic regional landscape with numerous stakeholders vying for AI leadership.

Geopolitical complexities surrounding the UAE’s AI ambitions also warrant attention. Microsoft’s partnership with G42 drew scrutiny in the US due to concerns about the company’s ties to China. These concerns underscore the intricate interplay between technological innovation and geopolitical considerations in the AI sphere.

Despite rapidly growing, the UAE’s AI sector remains smaller that those in the US and China. These established players hold a head start, and AI remains a major point of strategic interest for Beijing.

Although K2 Think exhibits performance comparable to OpenAI’s models, its developers emphasize that the objective is not to replicate a chatbot like ChatGPT. Instead, MBZUAI aims to focus the model on applications in fields such as mathematics and science.

“Mathematical reasoning is the underlying principle for all thought processes,” Richard Morton, Managing Director for MBZUAI’s Institute of Foundation Models, told CNBC.

“With this specific application, instead of requiring 1,000 human scientists 5 years to explore a question, this radically condenses the period involved, which could have important clinical and research implications.”

This strategy could allow advanced AI technologies to reach regions with limited resources. “We are finding that you can achieve more with less resources,” said Morton.